bigcode/jupyter-parsed · Datasets at Hugging Face

4ad526bf716f40599bab35d7f2cd7512f9ba249b

63,092

ipynb

Jupyter Notebook

Dual_Pixel_Net.ipynb

RugvedKatole/Learning-Single-Camera-Depth-Estimation-using-Dual-Pixels

4f05cd80cea3d25474f14a0c928bce1d700b49d2

[ "MIT" ]

null

Dual_Pixel_Net.ipynb

RugvedKatole/Learning-Single-Camera-Depth-Estimation-using-Dual-Pixels

4f05cd80cea3d25474f14a0c928bce1d700b49d2

[ "MIT" ]

null

Dual_Pixel_Net.ipynb

RugvedKatole/Learning-Single-Camera-Depth-Estimation-using-Dual-Pixels

4f05cd80cea3d25474f14a0c928bce1d700b49d2

[ "MIT" ]

null

81.304124

281

0.254533

[ [ [ "<a href=\"https://colab.research.google.com/github/RugvedKatole/Learning-Single-Camera-Depth-Estimation-using-Dual-Pixels/blob/main/Dual_Pixel_Net.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>", "_____no_output_____" ], [ "# Dual Pixel Net implementation\nLink to Paper: [Learning Single Camera Depth Estimation using Dual Pixels](https://arxiv.org/abs/1904.05822)\n", "_____no_output_____" ], [ "Import libraries ", "_____no_output_____" ] ], [ [ "import keras\nimport os\nimport copy\nimport json\nimport numpy as np\nfrom matplotlib import pyplot as plt\nimport tensorflow as tf\nfrom scipy.interpolate import interp2d\nimport numpy.random as random\nfrom tensorflow.keras.layers import Input, Conv2D ,Conv2DTranspose, MaxPooling2D, concatenate, Add, Dense, Dropout, Activation, Flatten, BatchNormalization, SeparableConv2D, LeakyReLU\nfrom tensorflow.keras.optimizers import Adam", "_____no_output_____" ] ], [ [ "Paper uses a Unet Architecture with Residual Blocks.\nUnet Architecture consists of a Encoder Decoder Network. Encoder Downsamples given images while decoder upsamples the downsampled images.k", "_____no_output_____" ] ], [ [ "# Encoder block A\ndef EncoderA(inputs=None, i_filters=32, o=32, s=2, max_pooling=True):\n \"\"\"\n Convolutional downsampling block\n \n Arguments:\n inputs -- Input tensor\n n_filters -- Number of filters for the convolutional layers \n dropout_prob -- Dropout probability\n max_pooling -- Use MaxPooling2D to reduce the spatial dimensions of the output volume\n Returns: \n next_layer, skip_connection -- Next layer and skip connection outputs\n \"\"\"\n # first Layer of Encoder Block\n #Note E_a(i,o,s) == E(i,o,s)\n conv = BatchNormalization()(inputs)\n\n conv = Conv2D(i_filters, # Number of filters i.e i in paper (E(i,o,s))\n (3,3), # 3x3 Kernel size \n padding='same',\n strides=(s,s))(conv) # s from E(i,o,s)\n \n conv = LeakyReLU(alpha=0.05)(conv)\n \n # Second Layer of Encoder Block Is a Depthwise Separable Convolution layer with 3x3 kernel\n conv = BatchNormalization()(conv)\n conv = SeparableConv2D(i_filters,(3,3),\n padding = 'same')(conv)\n conv = LeakyReLU(alpha=0.05)(conv)\n\n # Third layer of Encoder Block is 1x1 convolution Layer with o filters from E(i,o,s)\n conv = BatchNormalization()(conv)\n conv = Conv2D(o,(1,1), padding = 'same')(conv)\n conv = LeakyReLU(alpha=0.05)(conv)\n\n next_layer = BatchNormalization()(inputs)\n next_layer = SeparableConv2D(o,(3,3),\n padding = 'same')(next_layer)\n next_layer = LeakyReLU(alpha=0.05)(next_layer)\n next_layer = MaxPooling2D(pool_size=(s,s), strides=(s,s),padding='same')(next_layer)\n next_layer = Add()([conv,next_layer])\n \n skip_connection = conv\n \n return next_layer, skip_connection", "_____no_output_____" ], [ "# Encoder Block B\ndef EncoderB(inputs=None, o=32, s=2, max_pooling=True):\n \"\"\"\n Convolutional downsampling block\n \n Arguments:\n inputs -- Input tensor\n n_filters -- Number of filters for the convolutional layers \n dropout_prob -- Dropout probability\n max_pooling -- Use MaxPooling2D to reduce the spatial dimensions of the output volume\n Returns: \n next_layer, skip_connection -- Next layer and skip connection outputs\n \"\"\"\n # first Layer of Encoder Block\n conv = BatchNormalization()(inputs)\n conv = Conv2D(o, # Number of filters i.e o in paper (E_b(o,s))\n (7,7), # 3x3 Kernel size \n padding='same',\n kernel_initializer='he_normal',\n strides=(s,s))(conv) # s from E(o,s)\n conv = LeakyReLU(alpha=0.05)(conv)\n\n # the output of conv is added to max pooled input images\n Pooled_input = MaxPooling2D(pool_size=(s,s), strides=(s,s))(inputs)\n next_layer = concatenate([conv,Pooled_input],axis = 3)\n skip_connection = conv\n \n return next_layer, skip_connection", "_____no_output_____" ] ], [ [ "Now we create a Decoder block for our Network", "_____no_output_____" ] ], [ [ "# Decoder Block\ndef Decoder(expansive_input, contractive_input, i_filters = 32, o = 32):\n \"\"\"\n Convolutional upsampling block\n \n Arguments:\n expansive_input -- Input tensor from previous layer\n contractive_input -- Input tensor from previous skip layer\n i_filters -- Number of filters for the convolutional layers (o from (D(i,o)))\n Returns: \n conv -- Tensor output\n \"\"\"\n # first layer of decoder block i.e transpose conv to previous layer\n up = BatchNormalization()(expansive_input)\n up = Conv2DTranspose(\n i_filters, # number of filters\n (4,4), # Kernel size\n strides=(2,2),\n padding='same')(up)\n up = LeakyReLU(alpha=0.05)(up)\n \n \n # second layer of decoder block i.e 3x3 depth seperable conv \n up = BatchNormalization()(up)\n up = SeparableConv2D(i_filters,(3,3),\n padding = 'same')(up)\n up = LeakyReLU(alpha=0.05)(up)\n\n # Third layer of Decoder Block i.e 1x1 conv with i filters\n up = BatchNormalization()(up)\n up = Conv2D(i_filters,(1,1), padding = 'same')(up)\n up = LeakyReLU(alpha=0.05)(up)\n\n #fourth layer of Decoder block i.e 3x3 \n up = BatchNormalization()(up)\n up = SeparableConv2D(i_filters,(3,3),strides=(2,2),padding = 'same')(up)\n up = LeakyReLU(alpha=0.05)(up)\n\n # fifth layer \n up = BatchNormalization()(up)\n contractive_input = SeparableConv2D(i_filters,(3,3),\n padding = 'same')(contractive_input)\n\n # BC kitne layers hai\n next_layer = Add()([up,contractive_input])\n next_layer = LeakyReLU(alpha=0.05)(next_layer)\n #Finally the final layer\n next_layer = BatchNormalization()(next_layer)\n next_layer = Conv2D(o,(1,1), padding = 'same')(next_layer)\n next_layer = LeakyReLU(alpha=0.05)(next_layer)\n\n return next_layer", "_____no_output_____" ] ], [ [ "Now we have completed the require Encoder Decoder blocks with now create our model architecture", "_____no_output_____" ] ], [ [ "def Unet_model(input_size=(1024,1024,1)):\n \"\"\"\n Unet model\n \n Arguments:\n input_size -- Input shape\n Returns: \n model -- tf.keras.Model\n \"\"\"\n #Encoding\n inputs = Input(input_size)\n Block1E_b = EncoderB(inputs,8,2)\n Block1E_a = EncoderA(Block1E_b[0],11,11,1) # E^1_a\n\n Block2E_a = EncoderA(Block1E_b[0],16,32,2) \n Block2E_a = EncoderA(Block1E_b[0],16,32,1)\n Block2E_a = EncoderA(Block1E_b[0],16,32,1) # E^2_a\n\n Block3E_a = EncoderA(Block2E_a[0],16,64,2) \n Block3E_a = EncoderA(Block2E_a[0],16,64,1) \n Block3E_a = EncoderA(Block2E_a[0],16,64,1) #E^3_a\n \n Block4E_a = EncoderA(Block3E_a[0],32,128,2)\n Block4E_a = EncoderA(Block3E_a[0],32,128,1)\n Block4E_a = EncoderA(Block3E_a[0],32,128,1) #E^4_a\n\n Block5E_a = EncoderA(Block4E_a[0],32,128,2)\n Block5E_a = EncoderA(Block4E_a[0],32.128,1)\n Block5E_a = EncoderA(Block4E_a[0],32,128,1) \n\n #Decoding\n\n Block4D = Decoder(Block5E_a[0],Block4E_a[1],32,128) #D^4\n \n Block3D = Decoder(Block4D,Block3E_a[1],16,64) #D^4\n\n Block2D = Decoder(Block3D,Block2E_a[1],16,32) #D^4\n\n Block1D = Decoder(Block2D,Block1E_a[1],8,8) #D^4\n\n #Creating model\n model = tf.keras.Model(inputs=inputs, outputs=Block1D)\n\n return model\n\n\n\n", "_____no_output_____" ], [ "model=Unet_model((256,256,1))\nmodel.compile(optimizer= Adam(beta_2 = 0.9),loss='mean_squared_error',metrics=['mse'])\nmodel.summary()", "Model: \"model_6\"\n__________________________________________________________________________________________________\n Layer (type) Output Shape Param # Connected to \n==================================================================================================\n input_18 (InputLayer) [(None, 256, 256, 1 0 [] \n )] \n \n batch_normalization_421 (Batch (None, 256, 256, 1) 4 ['input_18[0][0]'] \n Normalization) \n \n conv2d_238 (Conv2D) (None, 128, 128, 8) 400 ['batch_normalization_421[0][0]']\n \n leaky_re_lu_317 (LeakyReLU) (None, 128, 128, 8) 0 ['conv2d_238[0][0]'] \n \n max_pooling2d_101 (MaxPooling2 (None, 128, 128, 1) 0 ['input_18[0][0]'] \n D) \n \n concatenate_30 (Concatenate) (None, 128, 128, 9) 0 ['leaky_re_lu_317[0][0]', \n 'max_pooling2d_101[0][0]'] \n \n batch_normalization_434 (Batch (None, 128, 128, 9) 36 ['concatenate_30[0][0]'] \n Normalization) \n \n conv2d_245 (Conv2D) (None, 128, 128, 16 1312 ['batch_normalization_434[0][0]']\n ) \n \n leaky_re_lu_330 (LeakyReLU) (None, 128, 128, 16 0 ['conv2d_245[0][0]'] \n ) \n \n batch_normalization_435 (Batch (None, 128, 128, 16 64 ['leaky_re_lu_330[0][0]'] \n Normalization) ) \n \n separable_conv2d_191 (Separabl (None, 128, 128, 16 416 ['batch_normalization_435[0][0]']\n eConv2D) ) \n \n leaky_re_lu_331 (LeakyReLU) (None, 128, 128, 16 0 ['separable_conv2d_191[0][0]'] \n ) \n \n batch_normalization_437 (Batch (None, 128, 128, 9) 36 ['concatenate_30[0][0]'] \n Normalization) \n \n batch_normalization_436 (Batch (None, 128, 128, 16 64 ['leaky_re_lu_331[0][0]'] \n Normalization) ) \n \n separable_conv2d_192 (Separabl (None, 128, 128, 32 401 ['batch_normalization_437[0][0]']\n eConv2D) ) \n \n conv2d_246 (Conv2D) (None, 128, 128, 32 544 ['batch_normalization_436[0][0]']\n ) \n \n leaky_re_lu_333 (LeakyReLU) (None, 128, 128, 32 0 ['separable_conv2d_192[0][0]'] \n ) \n \n leaky_re_lu_332 (LeakyReLU) (None, 128, 128, 32 0 ['conv2d_246[0][0]'] \n ) \n \n max_pooling2d_105 (MaxPooling2 (None, 128, 128, 32 0 ['leaky_re_lu_333[0][0]'] \n D) ) \n \n add_45 (Add) (None, 128, 128, 32 0 ['leaky_re_lu_332[0][0]', \n ) 'max_pooling2d_105[0][0]'] \n \n batch_normalization_446 (Batch (None, 128, 128, 32 128 ['add_45[0][0]'] \n Normalization) ) \n \n conv2d_251 (Conv2D) (None, 128, 128, 16 4624 ['batch_normalization_446[0][0]']\n ) \n \n leaky_re_lu_342 (LeakyReLU) (None, 128, 128, 16 0 ['conv2d_251[0][0]'] \n ) \n \n batch_normalization_447 (Batch (None, 128, 128, 16 64 ['leaky_re_lu_342[0][0]'] \n Normalization) ) \n \n separable_conv2d_197 (Separabl (None, 128, 128, 16 416 ['batch_normalization_447[0][0]']\n eConv2D) ) \n \n leaky_re_lu_343 (LeakyReLU) (None, 128, 128, 16 0 ['separable_conv2d_197[0][0]'] \n ) \n \n batch_normalization_449 (Batch (None, 128, 128, 32 128 ['add_45[0][0]'] \n Normalization) ) \n \n batch_normalization_448 (Batch (None, 128, 128, 16 64 ['leaky_re_lu_343[0][0]'] \n Normalization) ) \n \n separable_conv2d_198 (Separabl (None, 128, 128, 64 2400 ['batch_normalization_449[0][0]']\n eConv2D) ) \n \n conv2d_252 (Conv2D) (None, 128, 128, 64 1088 ['batch_normalization_448[0][0]']\n ) \n \n leaky_re_lu_345 (LeakyReLU) (None, 128, 128, 64 0 ['separable_conv2d_198[0][0]'] \n ) \n \n leaky_re_lu_344 (LeakyReLU) (None, 128, 128, 64 0 ['conv2d_252[0][0]'] \n ) \n \n max_pooling2d_108 (MaxPooling2 (None, 128, 128, 64 0 ['leaky_re_lu_345[0][0]'] \n D) ) \n \n add_48 (Add) (None, 128, 128, 64 0 ['leaky_re_lu_344[0][0]', \n ) 'max_pooling2d_108[0][0]'] \n \n batch_normalization_458 (Batch (None, 128, 128, 64 256 ['add_48[0][0]'] \n Normalization) ) \n \n conv2d_257 (Conv2D) (None, 128, 128, 32 18464 ['batch_normalization_458[0][0]']\n ) \n \n leaky_re_lu_354 (LeakyReLU) (None, 128, 128, 32 0 ['conv2d_257[0][0]'] \n ) \n \n batch_normalization_459 (Batch (None, 128, 128, 32 128 ['leaky_re_lu_354[0][0]'] \n Normalization) ) \n \n separable_conv2d_203 (Separabl (None, 128, 128, 32 1344 ['batch_normalization_459[0][0]']\n eConv2D) ) \n \n leaky_re_lu_355 (LeakyReLU) (None, 128, 128, 32 0 ['separable_conv2d_203[0][0]'] \n ) \n \n batch_normalization_461 (Batch (None, 128, 128, 64 256 ['add_48[0][0]'] \n Normalization) ) \n \n batch_normalization_460 (Batch (None, 128, 128, 32 128 ['leaky_re_lu_355[0][0]'] \n Normalization) ) \n \n separable_conv2d_204 (Separabl (None, 128, 128, 12 8896 ['batch_normalization_461[0][0]']\n eConv2D) 8) \n \n conv2d_258 (Conv2D) (None, 128, 128, 12 4224 ['batch_normalization_460[0][0]']\n 8) \n \n leaky_re_lu_357 (LeakyReLU) (None, 128, 128, 12 0 ['separable_conv2d_204[0][0]'] \n 8) \n \n leaky_re_lu_356 (LeakyReLU) (None, 128, 128, 12 0 ['conv2d_258[0][0]'] \n 8) \n \n max_pooling2d_111 (MaxPooling2 (None, 128, 128, 12 0 ['leaky_re_lu_357[0][0]'] \n D) 8) \n \n add_51 (Add) (None, 128, 128, 12 0 ['leaky_re_lu_356[0][0]', \n 8) 'max_pooling2d_111[0][0]'] \n \n batch_normalization_470 (Batch (None, 128, 128, 12 512 ['add_51[0][0]'] \n Normalization) 8) \n \n conv2d_263 (Conv2D) (None, 128, 128, 32 36896 ['batch_normalization_470[0][0]']\n ) \n \n leaky_re_lu_366 (LeakyReLU) (None, 128, 128, 32 0 ['conv2d_263[0][0]'] \n ) \n \n batch_normalization_471 (Batch (None, 128, 128, 32 128 ['leaky_re_lu_366[0][0]'] \n Normalization) ) \n \n separable_conv2d_209 (Separabl (None, 128, 128, 32 1344 ['batch_normalization_471[0][0]']\n eConv2D) ) \n \n leaky_re_lu_367 (LeakyReLU) (None, 128, 128, 32 0 ['separable_conv2d_209[0][0]'] \n ) \n \n batch_normalization_473 (Batch (None, 128, 128, 12 512 ['add_51[0][0]'] \n Normalization) 8) \n \n batch_normalization_472 (Batch (None, 128, 128, 32 128 ['leaky_re_lu_367[0][0]'] \n Normalization) ) \n \n separable_conv2d_210 (Separabl (None, 128, 128, 12 17664 ['batch_normalization_473[0][0]']\n eConv2D) 8) \n \n conv2d_264 (Conv2D) (None, 128, 128, 12 4224 ['batch_normalization_472[0][0]']\n 8) \n \n leaky_re_lu_369 (LeakyReLU) (None, 128, 128, 12 0 ['separable_conv2d_210[0][0]'] \n 8) \n \n leaky_re_lu_368 (LeakyReLU) (None, 128, 128, 12 0 ['conv2d_264[0][0]'] \n 8) \n \n max_pooling2d_114 (MaxPooling2 (None, 128, 128, 12 0 ['leaky_re_lu_369[0][0]'] \n D) 8) \n \n add_54 (Add) (None, 128, 128, 12 0 ['leaky_re_lu_368[0][0]', \n 8) 'max_pooling2d_114[0][0]'] \n \n batch_normalization_474 (Batch (None, 128, 128, 12 512 ['add_54[0][0]'] \n Normalization) 8) \n \n conv2d_transpose_24 (Conv2DTra (None, 256, 256, 32 65568 ['batch_normalization_474[0][0]']\n nspose) ) \n \n leaky_re_lu_370 (LeakyReLU) (None, 256, 256, 32 0 ['conv2d_transpose_24[0][0]'] \n ) \n \n batch_normalization_475 (Batch (None, 256, 256, 32 128 ['leaky_re_lu_370[0][0]'] \n Normalization) ) \n \n separable_conv2d_211 (Separabl (None, 256, 256, 32 1344 ['batch_normalization_475[0][0]']\n eConv2D) ) \n \n leaky_re_lu_371 (LeakyReLU) (None, 256, 256, 32 0 ['separable_conv2d_211[0][0]'] \n ) \n \n batch_normalization_476 (Batch (None, 256, 256, 32 128 ['leaky_re_lu_371[0][0]'] \n Normalization) ) \n \n conv2d_265 (Conv2D) (None, 256, 256, 32 1056 ['batch_normalization_476[0][0]']\n ) \n \n leaky_re_lu_372 (LeakyReLU) (None, 256, 256, 32 0 ['conv2d_265[0][0]'] \n ) \n \n batch_normalization_477 (Batch (None, 256, 256, 32 128 ['leaky_re_lu_372[0][0]'] \n Normalization) ) \n \n separable_conv2d_212 (Separabl (None, 128, 128, 32 1344 ['batch_normalization_477[0][0]']\n eConv2D) ) \n \n leaky_re_lu_373 (LeakyReLU) (None, 128, 128, 32 0 ['separable_conv2d_212[0][0]'] \n ) \n \n batch_normalization_478 (Batch (None, 128, 128, 32 128 ['leaky_re_lu_373[0][0]'] \n Normalization) ) \n \n separable_conv2d_213 (Separabl (None, 128, 128, 32 5280 ['leaky_re_lu_356[0][0]'] \n eConv2D) ) \n \n add_55 (Add) (None, 128, 128, 32 0 ['batch_normalization_478[0][0]',\n ) 'separable_conv2d_213[0][0]'] \n \n leaky_re_lu_374 (LeakyReLU) (None, 128, 128, 32 0 ['add_55[0][0]'] \n ) \n \n batch_normalization_479 (Batch (None, 128, 128, 32 128 ['leaky_re_lu_374[0][0]'] \n Normalization) ) \n \n conv2d_266 (Conv2D) (None, 128, 128, 12 4224 ['batch_normalization_479[0][0]']\n 8) \n \n leaky_re_lu_375 (LeakyReLU) (None, 128, 128, 12 0 ['conv2d_266[0][0]'] \n 8) \n \n batch_normalization_480 (Batch (None, 128, 128, 12 512 ['leaky_re_lu_375[0][0]'] \n Normalization) 8) \n \n conv2d_transpose_25 (Conv2DTra (None, 256, 256, 16 32784 ['batch_normalization_480[0][0]']\n nspose) ) \n \n leaky_re_lu_376 (LeakyReLU) (None, 256, 256, 16 0 ['conv2d_transpose_25[0][0]'] \n ) \n \n batch_normalization_481 (Batch (None, 256, 256, 16 64 ['leaky_re_lu_376[0][0]'] \n Normalization) ) \n \n separable_conv2d_214 (Separabl (None, 256, 256, 16 416 ['batch_normalization_481[0][0]']\n eConv2D) ) \n \n leaky_re_lu_377 (LeakyReLU) (None, 256, 256, 16 0 ['separable_conv2d_214[0][0]'] \n ) \n \n batch_normalization_482 (Batch (None, 256, 256, 16 64 ['leaky_re_lu_377[0][0]'] \n Normalization) ) \n \n conv2d_267 (Conv2D) (None, 256, 256, 16 272 ['batch_normalization_482[0][0]']\n ) \n \n leaky_re_lu_378 (LeakyReLU) (None, 256, 256, 16 0 ['conv2d_267[0][0]'] \n ) \n \n batch_normalization_483 (Batch (None, 256, 256, 16 64 ['leaky_re_lu_378[0][0]'] \n Normalization) ) \n \n separable_conv2d_215 (Separabl (None, 128, 128, 16 416 ['batch_normalization_483[0][0]']\n eConv2D) ) \n \n leaky_re_lu_379 (LeakyReLU) (None, 128, 128, 16 0 ['separable_conv2d_215[0][0]'] \n ) \n \n batch_normalization_484 (Batch (None, 128, 128, 16 64 ['leaky_re_lu_379[0][0]'] \n Normalization) ) \n \n separable_conv2d_216 (Separabl (None, 128, 128, 16 1616 ['leaky_re_lu_344[0][0]'] \n eConv2D) ) \n \n add_56 (Add) (None, 128, 128, 16 0 ['batch_normalization_484[0][0]',\n ) 'separable_conv2d_216[0][0]'] \n \n leaky_re_lu_380 (LeakyReLU) (None, 128, 128, 16 0 ['add_56[0][0]'] \n ) \n \n batch_normalization_485 (Batch (None, 128, 128, 16 64 ['leaky_re_lu_380[0][0]'] \n Normalization) ) \n \n conv2d_268 (Conv2D) (None, 128, 128, 64 1088 ['batch_normalization_485[0][0]']\n ) \n \n leaky_re_lu_381 (LeakyReLU) (None, 128, 128, 64 0 ['conv2d_268[0][0]'] \n ) \n \n batch_normalization_486 (Batch (None, 128, 128, 64 256 ['leaky_re_lu_381[0][0]'] \n Normalization) ) \n \n conv2d_transpose_26 (Conv2DTra (None, 256, 256, 16 16400 ['batch_normalization_486[0][0]']\n nspose) ) \n \n leaky_re_lu_382 (LeakyReLU) (None, 256, 256, 16 0 ['conv2d_transpose_26[0][0]'] \n ) \n \n batch_normalization_487 (Batch (None, 256, 256, 16 64 ['leaky_re_lu_382[0][0]'] \n Normalization) ) \n \n separable_conv2d_217 (Separabl (None, 256, 256, 16 416 ['batch_normalization_487[0][0]']\n eConv2D) ) \n \n leaky_re_lu_383 (LeakyReLU) (None, 256, 256, 16 0 ['separable_conv2d_217[0][0]'] \n ) \n \n batch_normalization_488 (Batch (None, 256, 256, 16 64 ['leaky_re_lu_383[0][0]'] \n Normalization) ) \n \n conv2d_269 (Conv2D) (None, 256, 256, 16 272 ['batch_normalization_488[0][0]']\n ) \n \n leaky_re_lu_384 (LeakyReLU) (None, 256, 256, 16 0 ['conv2d_269[0][0]'] \n ) \n \n batch_normalization_489 (Batch (None, 256, 256, 16 64 ['leaky_re_lu_384[0][0]'] \n Normalization) ) \n \n separable_conv2d_218 (Separabl (None, 128, 128, 16 416 ['batch_normalization_489[0][0]']\n eConv2D) ) \n \n leaky_re_lu_385 (LeakyReLU) (None, 128, 128, 16 0 ['separable_conv2d_218[0][0]'] \n ) \n \n batch_normalization_490 (Batch (None, 128, 128, 16 64 ['leaky_re_lu_385[0][0]'] \n Normalization) ) \n \n separable_conv2d_219 (Separabl (None, 128, 128, 16 816 ['leaky_re_lu_332[0][0]'] \n eConv2D) ) \n \n add_57 (Add) (None, 128, 128, 16 0 ['batch_normalization_490[0][0]',\n ) 'separable_conv2d_219[0][0]'] \n \n leaky_re_lu_386 (LeakyReLU) (None, 128, 128, 16 0 ['add_57[0][0]'] \n ) \n \n batch_normalization_491 (Batch (None, 128, 128, 16 64 ['leaky_re_lu_386[0][0]'] \n Normalization) ) \n \n conv2d_270 (Conv2D) (None, 128, 128, 32 544 ['batch_normalization_491[0][0]']\n ) \n \n leaky_re_lu_387 (LeakyReLU) (None, 128, 128, 32 0 ['conv2d_270[0][0]'] \n ) \n \n batch_normalization_492 (Batch (None, 128, 128, 32 128 ['leaky_re_lu_387[0][0]'] \n Normalization) ) \n \n conv2d_transpose_27 (Conv2DTra (None, 256, 256, 8) 4104 ['batch_normalization_492[0][0]']\n nspose) \n \n leaky_re_lu_388 (LeakyReLU) (None, 256, 256, 8) 0 ['conv2d_transpose_27[0][0]'] \n \n batch_normalization_493 (Batch (None, 256, 256, 8) 32 ['leaky_re_lu_388[0][0]'] \n Normalization) \n \n batch_normalization_422 (Batch (None, 128, 128, 9) 36 ['concatenate_30[0][0]'] \n Normalization) \n \n separable_conv2d_220 (Separabl (None, 256, 256, 8) 144 ['batch_normalization_493[0][0]']\n eConv2D) \n \n conv2d_239 (Conv2D) (None, 128, 128, 11 902 ['batch_normalization_422[0][0]']\n ) \n \n leaky_re_lu_389 (LeakyReLU) (None, 256, 256, 8) 0 ['separable_conv2d_220[0][0]'] \n \n leaky_re_lu_318 (LeakyReLU) (None, 128, 128, 11 0 ['conv2d_239[0][0]'] \n ) \n \n batch_normalization_494 (Batch (None, 256, 256, 8) 32 ['leaky_re_lu_389[0][0]'] \n Normalization) \n \n batch_normalization_423 (Batch (None, 128, 128, 11 44 ['leaky_re_lu_318[0][0]'] \n Normalization) ) \n \n conv2d_271 (Conv2D) (None, 256, 256, 8) 72 ['batch_normalization_494[0][0]']\n \n separable_conv2d_185 (Separabl (None, 128, 128, 11 231 ['batch_normalization_423[0][0]']\n eConv2D) ) \n \n leaky_re_lu_390 (LeakyReLU) (None, 256, 256, 8) 0 ['conv2d_271[0][0]'] \n \n leaky_re_lu_319 (LeakyReLU) (None, 128, 128, 11 0 ['separable_conv2d_185[0][0]'] \n ) \n \n batch_normalization_495 (Batch (None, 256, 256, 8) 32 ['leaky_re_lu_390[0][0]'] \n Normalization) \n \n batch_normalization_424 (Batch (None, 128, 128, 11 44 ['leaky_re_lu_319[0][0]'] \n Normalization) ) \n \n separable_conv2d_221 (Separabl (None, 128, 128, 8) 144 ['batch_normalization_495[0][0]']\n eConv2D) \n \n conv2d_240 (Conv2D) (None, 128, 128, 11 132 ['batch_normalization_424[0][0]']\n ) \n \n leaky_re_lu_391 (LeakyReLU) (None, 128, 128, 8) 0 ['separable_conv2d_221[0][0]'] \n \n leaky_re_lu_320 (LeakyReLU) (None, 128, 128, 11 0 ['conv2d_240[0][0]'] \n ) \n \n batch_normalization_496 (Batch (None, 128, 128, 8) 32 ['leaky_re_lu_391[0][0]'] \n Normalization) \n \n separable_conv2d_222 (Separabl (None, 128, 128, 8) 195 ['leaky_re_lu_320[0][0]'] \n eConv2D) \n \n add_58 (Add) (None, 128, 128, 8) 0 ['batch_normalization_496[0][0]',\n 'separable_conv2d_222[0][0]'] \n \n leaky_re_lu_392 (LeakyReLU) (None, 128, 128, 8) 0 ['add_58[0][0]'] \n \n batch_normalization_497 (Batch (None, 128, 128, 8) 32 ['leaky_re_lu_392[0][0]'] \n Normalization) \n \n conv2d_272 (Conv2D) (None, 128, 128, 8) 72 ['batch_normalization_497[0][0]']\n \n leaky_re_lu_393 (LeakyReLU) (None, 128, 128, 8) 0 ['conv2d_272[0][0]'] \n \n==================================================================================================\nTotal params: 250,533\nTrainable params: 247,729\nNon-trainable params: 2,804\n__________________________________________________________________________________________________\n" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ] ]

4ad531c7f53551360c342dc34900b3429f61f5d3

25,150

ipynb

Jupyter Notebook

One-Shot Classification/One_Shot_Classification_V1.ipynb

Naren-Jegan/Deep-Learning-Keras

8a962226c3527b6b382a46d0981f78a20c4f8708

[ "MIT" ]

null

One-Shot Classification/One_Shot_Classification_V1.ipynb

Naren-Jegan/Deep-Learning-Keras

8a962226c3527b6b382a46d0981f78a20c4f8708

[ "MIT" ]

null

One-Shot Classification/One_Shot_Classification_V1.ipynb

Naren-Jegan/Deep-Learning-Keras

8a962226c3527b6b382a46d0981f78a20c4f8708

[ "MIT" ]

null

40.304487

226

0.513042

[ [ [ "[View in Colaboratory](https://colab.research.google.com/github/Naren-Jegan/Deep-Learning-Keras/blob/master/One_Shot_Classification_V1.ipynb)", "_____no_output_____" ], [ "# One Shot Learning on Omniglot Dataset\n\nThe [Omniglot](https://github.com/brendenlake/omniglot) dataset contains 1623 different handwritten characters from 50 different alphabets.\nEach of the 1623 characters was drawn online via Amazon's Mechanical Turk by 20 different people.\nThis dataset has been the baseline for any one-shot learning algorithm.\n\n\nSome of the machine learning algorithms used for learning this dataset over the years are listed below in order of accuracy:\n* Hierarchical Bayesian Program Learning - 95.2%\n* Convolutional Siamese Net - 92.0%\n* Affine model - 81.8%\n* Hierarchical Deep - 65.2%\n* Deep Boltzmann Machine - 62.0%\n* Siamese Neural Net - 58.3%\n* Simple Stroke - 35.2%\n* 1-Nearest Neighbor - 21.7%\n\n\nThis notebook implements a [Convolutional Siamese Neural Network](https://https://www.cs.cmu.edu/~rsalakhu/papers/oneshot1.pdf) using a background set of 30 alphabets for training and evaluate on set of 20 alphabets.", "_____no_output_____" ] ], [ [ "from google.colab import auth, drive\nauth.authenticate_user()\ndrive.mount('/content/drive')", "_____no_output_____" ], [ "%matplotlib inline\nimport matplotlib.pyplot as plt\nimport tensorflow as tf\nimport numpy as np\nimport math\nimport os\nfrom PIL import Image, ImageFilter, ImageOps, ImageMath\nimport numpy.random as rnd\nimport pickle\nfrom time import sleep\nfrom copy import deepcopy", "_____no_output_____" ], [ "# from tf.keras.models import Sequential # This does not work!\nfrom tensorflow.python.keras.models import Model, Sequential\nfrom tensorflow.python.keras.layers import InputLayer, Input, Lambda\nfrom tensorflow.python.keras.layers import Reshape, MaxPooling2D, Dropout, BatchNormalization\nfrom tensorflow.python.keras.layers import Conv2D, Dense, Flatten\nfrom tensorflow.python.keras.preprocessing.image import ImageDataGenerator\nfrom tensorflow.python.keras.models import load_model\nfrom tensorflow.python.keras import backend as K\nfrom tensorflow.python.keras.regularizers import l2\nfrom tensorflow.python.keras.initializers import RandomNormal\nfrom tensorflow import test, logging\nfrom keras.wrappers.scikit_learn import KerasClassifier\nfrom keras.wrappers.scikit_learn import GridSearchCV\nlogging.set_verbosity(tf.logging.ERROR)\ntest.gpu_device_name()", "_____no_output_____" ], [ "tf.__version__", "_____no_output_____" ], [ "one_shot_path = os.path.join(\"drive\", \"My Drive\", \"Colab Notebooks\", \"One-Shot Classification\")\nbackground_path = os.path.join(one_shot_path, \"background\")\nevaluation_path = os.path.join(one_shot_path, \"evaluation\")\nrecognition_model_path = os.path.join(one_shot_path, \"recognition_model.h5\")", "_____no_output_____" ], [ "##creating training set\ntrain_data = np.ndarray(shape=(964, 20, 105, 105))\ntrain_alphabets = dict()\n\n#for alphabet in os.listdir(background_path):\n# alphabet_path = os.path.join(background_path, alphabet)\n# for character in os.listdir(alphabet_path):\n# character_path = os.path.join(alphabet_path, character)\n# for image in os.listdir(character_path):\n# index = int(image[0:4]) - 1\n# writer = int(image[5:7]) - 1\n# train_data[index][writer] = np.array(Image.open(os.path.join(character_path, image)))\n# train_alphabets[alphabet] = index if alphabet not in train_alphabets or train_alphabets[alphabet] > index else train_alphabets[alphabet]\n\n#with open(os.path.join(\"train.pickle\"), 'wb') as f:\n# pickle.dump([train_data, train_alphabets], f, protocol=2)", "_____no_output_____" ], [ "with open(os.path.join(one_shot_path, \"train.pickle\"), 'rb') as f:\n train_data, train_alphabets = pickle.load(f, encoding='latin1')", "_____no_output_____" ], [ "#@title Inputs\n\nconv_activation = 'relu' #@param ['relu', 'softplus', 'tanh', 'sigmoid'] {type:\"string\"}\ndense_activation = 'sigmoid' #@param ['relu', 'softplus', 'tanh', 'sigmoid'] {type:\"string\"}\nlearning_rate = 1e-2 #@param {type:\"number\"}\nconv_regularization_parameter = 1e-2 #@param {type:\"number\"}\ndense_regularization_parameter = 1e-4 #@param {type:\"number\"}\nbatch_size = 128 #@param {type:\"slider\", min:0, max:1024, step:16}\nbatches_per_epoch = 75 #@param {type:\"slider\", min:0, max:100, step:5}\nn_epochs = 200 #@param {type:\"slider\", min:25, max:500, step:25}\n\n\nbatch_size = 1 if batch_size == 0 else batch_size\nbatches_per_epoch = 1 if batches_per_epoch == 0 else batches_per_epoch\n", "_____no_output_____" ], [ "#@title Data Augmentation\nimage_size = 105 #@param {type:\"slider\", min:32, max:512, step:1}\n\nrotation_range = 10 #@param {type:\"slider\", min:0, max:90, step:1}\nwidth_shift_range = 2 #@param {type:\"slider\", min:0, max:10, step:0.1}\nheight_shift_range = 2 #@param {type:\"slider\", min:0, max:10, step:0.1}\nshear_range = 0.3 #@param {type:\"slider\", min:0, max:1, step:0.1}\nzoom_range = 0.2 #@param {type:\"slider\", min:0, max:1, step:0.01}", "_____no_output_____" ], [ "# this is the augmentation configuration we will use for training\ndatagen = ImageDataGenerator()\n\ndef transform_image(image):\n return datagen.apply_transform(image.reshape((image_size, image_size, 1)), \n transform_parameters = \n {'theta': rnd.uniform(-rotation_range, rotation_range),\n 'tx' : rnd.uniform(-width_shift_range, width_shift_range),\n 'ty' : rnd.uniform(-height_shift_range, height_shift_range),\n 'shear': rnd.uniform(-shear_range, shear_range),\n 'zx' : rnd.uniform(-zoom_range, zoom_range),\n 'zy' : rnd.uniform(-zoom_range, zoom_range)\n })\n\n#generate image pairs [x1, x2] with target y = 1/0 representing same/different\ndef datagen_flow(datagen, val = False):\n while True:\n X1 = np.ndarray(shape=(batch_size, image_size, image_size, 1))\n X2 = np.ndarray(shape=(batch_size, image_size, image_size, 1))\n Y = np.ndarray(shape=(batch_size,))\n \n s_alphabets = sorted(train_alphabets.values())\n a_indices = list(range(len(s_alphabets)))\n times = batch_size//(2*len(a_indices))\n remainder = (batch_size//2)%len(a_indices)\n \n aindices = a_indices*times + list(rnd.choice(a_indices, remainder))\n rnd.shuffle(aindices)\n \n w_range = list(range(12, 20) if val else range(12))\n \n i = 0 \n for a in aindices:\n end_index = (len(train_data) if a+1 == len(s_alphabets) else s_alphabets[a+1])\n c_range = list(range(s_alphabets[a], end_index))\n \n writers = rnd.choice(w_range, 2)\n same = rnd.choice(c_range)\n X1[2*i] = transform_image(train_data[same, writers[0]])\n X2[2*i] = transform_image(train_data[same, writers[1]])\n Y[2*i] = 1.0\n \n writers = rnd.choice(w_range, 2)\n diff = rnd.choice(c_range, 2)\n X1[2*i + 1] = transform_image(train_data[diff[0], writers[0]])\n X2[2*i + 1] = transform_image(train_data[diff[1], writers[1]])\n Y[2*i + 1] = 0.0\n \n i += 1\n \n yield [X1, X2], Y\n\ntrain_generator = datagen_flow(datagen) \n\n# this is a similar generator, for validation data that takes only the remaining 8 writers\ntrain_dev_generator = datagen_flow(datagen, val=True)", "_____no_output_____" ], [ "\nw_init = RandomNormal(mean=0.0, stddev=1e-2)\nb_init = RandomNormal(mean=0.5, stddev=1e-2)", "_____no_output_____" ], [ "input_shape=(image_size, image_size, 1)\nleft_input = Input(input_shape)\nright_input = Input(input_shape)\n\n# Start construction of the Keras Sequential model.\nconvnet = Sequential()\n\n# First convolutional layer with activation, batchnorm and max-pooling.\nconvnet.add(Conv2D(kernel_size=10, strides=1, filters=64, padding='valid',\n input_shape=input_shape, bias_initializer=b_init,\n activation=conv_activation,\n name='layer_conv1', kernel_regularizer=l2(conv_regularization_parameter)))\nconvnet.add(BatchNormalization(axis = 3, momentum=0.5, name = 'bn1'))\nconvnet.add(MaxPooling2D(pool_size=2, strides=2, name=\"max_pooling1\"))\n\n# Second convolutional layer with activation, batchnorm and max-pooling.\nconvnet.add(Conv2D(kernel_size=7, strides=1, filters=128, padding='valid',\n kernel_initializer=w_init, bias_initializer=b_init,\n activation=conv_activation, name='layer_conv2', kernel_regularizer=l2(conv_regularization_parameter)))\nconvnet.add(BatchNormalization(axis = 3, name = 'bn2'))\nconvnet.add(MaxPooling2D(pool_size=2, strides=2, name=\"max_pooling2\"))\n\n# Third convolutional layer with activation, batchnorm and max-pooling.\nconvnet.add(Conv2D(kernel_size=4, strides=1, filters=128, padding='valid',\n kernel_initializer=w_init, bias_initializer=b_init,\n activation=conv_activation, name='layer_conv3', kernel_regularizer=l2(conv_regularization_parameter)))\nconvnet.add(BatchNormalization(axis = 3, name = 'bn3'))\nconvnet.add(MaxPooling2D(pool_size=2, strides=2, name=\"max_pooling3\"))\n\n# Fourth convolutional layer with activation, batchnorm and max-pooling.\nconvnet.add(Conv2D(kernel_size=4, strides=1, filters=256, padding='valid',\n kernel_initializer=w_init, bias_initializer=b_init,\n activation=conv_activation, name='layer_conv4', kernel_regularizer=l2(conv_regularization_parameter)))\nconvnet.add(BatchNormalization(axis = 3, name = 'bn4'))\nconvnet.add(MaxPooling2D(pool_size=2, strides=2, name=\"max_pooling4\"))\n\n# Flatten the 4-rank output of the convolutional layers\n# to 2-rank that can be input to a fully-connected / dense layer.\nconvnet.add(Flatten())\n\n\n# First fully-connected / dense layer with activation.\nconvnet.add(Dense(4096, activation=dense_activation,\n kernel_initializer=w_init, bias_initializer=b_init,\n name = \"dense_1\", kernel_regularizer=l2(dense_regularization_parameter)))\nconvnet.add(BatchNormalization(axis = 1, name = 'bn5'))\n\n#call the convnet Sequential model on each of the input tensors so params will be shared\nencoded_l = convnet(left_input)\nencoded_r = convnet(right_input)\n\n#layer to merge two encoded inputs with the l1 distance between them\nL1_layer = Lambda(lambda tensors:K.abs(tensors[0] - tensors[1]))\n\n#call this layer on list of two input tensors.\nL1_distance = L1_layer([encoded_l, encoded_r])\n\nprediction = Dense(1,activation='sigmoid',bias_initializer=b_init)(L1_distance)\n\nmodel = Model(inputs=[left_input,right_input],outputs=prediction)", "_____no_output_____" ], [ "from tensorflow.python.keras.optimizers import SGD, Adam\n\n#optimizer = SGD(lr=learning_rate, momentum=0.5)\noptimizer = Adam(lr=learning_rate)\n\nmodel.compile(optimizer=optimizer,\n loss='binary_crossentropy',\n metrics=['accuracy'])\n\nsteps_train = batches_per_epoch\nsteps_validation = batches_per_epoch ", "_____no_output_____" ], [ "from tensorflow.python.keras.callbacks import ModelCheckpoint, Callback, LearningRateScheduler, ReduceLROnPlateau\n\nmodel_checkpoint = ModelCheckpoint(recognition_model_path, monitor='val_loss',\n save_best_only=True, period=10)\n\nlr_scheduler = LearningRateScheduler(lambda epoch, lr: 0.99*lr)\n\nreduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.5,\n patience=5, min_lr=1e-4)\n\nclass LearningRateFinder(Callback):\n def __init__(self, steps=100, period=10):\n super(LearningRateFinder, self).__init__()\n self.steps = steps\n self.batch_size=batch_size\n self.period = period\n self.best_lr = 1e-4\n self.best_loss = 1000\n self.find_lr = True\n self.current_lr = None\n self.training_path = os.path.join(one_shot_path, \"training_model.h5\")\n self.model_weights = None\n \n def reset_values(self):\n K.set_value(self.model.optimizer.lr, self.best_lr)\n self.best_lr = 1e-4\n self.best_loss = 1000\n self.model = load_model(self.training_path)\n \n def on_train_begin(self, logs={}):\n return\n\n def on_train_end(self, logs={}):\n return\n\n def on_epoch_begin(self, epoch, logs={}):\n self.find_lr = epoch % self.period == 0\n if epoch % self.period == 1:\n print(\"Learning Rate: \" + \"{0:.2g}\".format(K.get_value(self.model.optimizer.lr)))\n if(self.find_lr):\n self.current_lr = K.get_value(self.model.optimizer.lr)\n self.model.save(self.training_path)\n self.model_weights = self.model.get_weights()\n \n def on_epoch_end(self, epoch, logs={}):\n if(self.find_lr):\n self.reset_values()\n return \n\n def on_batch_begin(self, batch, logs={}):\n if(self.find_lr):\n K.set_value(self.model.optimizer.lr, 10**(2*batch/self.steps + np.log10(self.current_lr) - 1))\n return\n\n def on_batch_end(self, batch, logs={}):\n if(self.find_lr):\n loss = logs.get('loss')\n if loss < self.best_loss:\n self.best_loss = loss\n self.best_lr = K.get_value(self.model.optimizer.lr)\n elif loss >= 1.25*self.best_loss:\n self.find_lr = False\n self.reset_values()\n self.model.set_weights(self.model_weights)\n \n return\n \nlr_finder = LearningRateFinder(steps=steps_train, period=n_epochs//4)", "_____no_output_____" ], [ "model.fit_generator(train_generator, \n steps_per_epoch = steps_train,\n epochs=n_epochs,\n validation_data = train_dev_generator,\n validation_steps = steps_validation,\n callbacks = [model_checkpoint, lr_scheduler, reduce_lr]\n )", "_____no_output_____" ], [ "model = load_model(recognition_model_path)", "_____no_output_____" ], [ "##creating test set\ntest_data = np.ndarray(shape=(659, 20, 105, 105))\ntest_alphabets = dict()\n\n#for alphabet in os.listdir(evaluation_path):\n# alphabet_path = os.path.join(evaluation_path, alphabet)\n# for character in os.listdir(alphabet_path):\n# character_path = os.path.join(alphabet_path, character)\n# for image in os.listdir(character_path):\n# index = int(image[0:4]) - 965\n# writer = int(image[5:7]) - 1\n# test_data[index][writer] = np.array(Image.open(os.path.join(character_path, image)))\n# test_alphabets[alphabet] = index if alphabet not in test_alphabets or test_alphabets[alphabet] > index else test_alphabets[alphabet]\n\n#with open(os.path.join(\"test.pickle\"), 'wb') as f:\n# pickle.dump([test_data, test_alphabets], f, protocol=2)", "_____no_output_____" ], [ "with open(os.path.join(one_shot_path, \"test.pickle\"), 'rb') as f:\n test_data, test_alphabets = pickle.load(f, encoding='latin1')", "_____no_output_____" ], [ "N = 20\nst_alphabets = sorted(test_alphabets.values())\ncorrect = 0\nshow = True\nfor i in range(len(st_alphabets)):\n end_index = len(test_data) if i+1 == len(st_alphabets) else st_alphabets[i+1] \n c_range = list(range(st_alphabets[i],end_index))\n \n for j in range(2):\n c_list = rnd.choice(c_range, N)\n w_list = rnd.choice(range(20), 2)\n \n for c_i in range(N):\n image = test_data[c_list[c_i]][w_list[0]]\n \n X1 = np.array([image]*N).reshape((N, image_size, image_size, 1))\n X2 = np.array(test_data[c_list][w_list[1]]).reshape((N, image_size, image_size, 1))\n if show and c_i == 2 and i == 3:\n plt.imshow(image)\n plt.show()\n for m in range(N):\n plt.imshow(test_data[c_list[m]][w_list[1]])\n plt.show()\n \n \n targets = np.zeros((N,))\n targets[c_i] = 1\n predictions = model.predict([X1, X2])\n \n if show and c_i == 2 and i == 3:\n print(targets)\n print(predictions)\n show = False\n \n if(np.argmax(predictions) == np.argmax(targets)):\n correct += 1\n\nprint(str(N) + \"-Way Classification Accuracy: \" + \"{0:.2f}\".format(correct/(N*20*2))) ", "_____no_output_____" ] ] ]

[ "markdown", "code" ]

[ [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad53254b3472915021185584ecd4e257468771d

93,327

ipynb

Jupyter Notebook

squash trainer/notebooks/multitask_transformer/Train.ipynb

kilshaw/predictions-training

6c0225d15661717e3c46b9c45ded8a1a129e0b14

[ "MIT" ]

402

2019-07-31T00:37:10.000Z

2022-03-27T22:21:29.000Z

squash trainer/notebooks/multitask_transformer/Train.ipynb

kilshaw/predictions-training

6c0225d15661717e3c46b9c45ded8a1a129e0b14

[ "MIT" ]

26

2019-08-20T13:44:30.000Z

2022-01-27T10:42:28.000Z

squash trainer/notebooks/multitask_transformer/Train.ipynb

kilshaw/predictions-training

6c0225d15661717e3c46b9c45ded8a1a129e0b14

[ "MIT" ]

81

2019-08-14T06:55:55.000Z

2022-03-19T09:49:15.000Z

78.624263

20,436

0.713781

[ [ [ "%reload_ext autoreload\n%autoreload 2\n%matplotlib inline", "_____no_output_____" ], [ "import os\nos.chdir('../../')", "_____no_output_____" ], [ "from musicautobot.numpy_encode import *\nfrom musicautobot.utils.file_processing import process_all, process_file\nfrom musicautobot.config import *\nfrom musicautobot.music_transformer import *\nfrom musicautobot.multitask_transformer import *\nfrom musicautobot.utils.stacked_dataloader import StackedDataBunch", "_____no_output_____" ], [ "from fastai.text import *", "_____no_output_____" ] ], [ [ "## MultitaskTransformer Training\n\nMultitask Training is an extension of [MusicTransformer](../music_transformer/Train.ipynb).\n\nInstead a basic language model that predicts the next word...\n\nWe train on multiple tasks\n* [Next Word](../music_transformer/Train.ipynb)\n* [Bert Mask](https://arxiv.org/abs/1810.04805)\n* [Sequence to Sequence Translation](http://jalammar.github.io/illustrated-transformer/)\n\nThis gives a more generalized model and also let's you do some really cool [predictions](Generate.ipynb)", "_____no_output_____" ], [ "## End to end training pipeline \n\n1. Create and encode dataset\n2. Initialize Transformer MOdel\n3. Train\n4. Predict", "_____no_output_____" ] ], [ [ "# Location of your midi files\nmidi_path = Path('data/midi/examples')\nmidi_path.mkdir(parents=True, exist_ok=True)\n\n# Location to save dataset\ndata_path = Path('data/numpy')\ndata_path.mkdir(parents=True, exist_ok=True)\n\ndata_save_name = 'musicitem_data_save.pkl'\ns2s_data_save_name = 'multiitem_data_save.pkl'", "_____no_output_____" ] ], [ [ "## 1. Gather midi dataset", "_____no_output_____" ], [ "Make sure all your midi data is in `musicautobot/data/midi` directory", "_____no_output_____" ], [ "Here's a pretty good dataset with lots of midi data: \nhttps://www.reddit.com/r/datasets/comments/3akhxy/the_largest_midi_collection_on_the_internet/\n\nDownload the folder and unzip it to `data/midi`", "_____no_output_____" ], [ "## 2. Create dataset from MIDI files", "_____no_output_____" ] ], [ [ "midi_files = get_files(midi_path, '.mid', recurse=True); len(midi_files)", "_____no_output_____" ] ], [ [ "### 2a. Create NextWord/Mask Dataset", "_____no_output_____" ] ], [ [ "processors = [Midi2ItemProcessor()]\ndata = MusicDataBunch.from_files(midi_files, data_path, processors=processors, \n encode_position=True, dl_tfms=mask_lm_tfm_pitchdur, \n bptt=5, bs=2)\ndata.save(data_save_name)", "_____no_output_____" ], [ "xb, yb = data.one_batch(); xb", "_____no_output_____" ] ], [ [ "Key:\n* 'msk' = masked input\n* 'lm' = next word input\n* 'pos' = timestepped postional encoding. This is in addition to relative positional encoding\n\nNote: MultitaskTransformer trains on both the masked input ('msk') and next word input ('lm') at the same time.\n\nThe encoder is trained on the 'msk' data, while the decoder is trained on 'lm' data.\n\n", "_____no_output_____" ], [ "### 2b. Create sequence to sequence dataset", "_____no_output_____" ] ], [ [ "processors = [Midi2MultitrackProcessor()]\ns2s_data = MusicDataBunch.from_files(midi_files, data_path, processors=processors, \n preloader_cls=S2SPreloader, list_cls=S2SItemList,\n dl_tfms=melody_chord_tfm,\n bptt=5, bs=2)\ns2s_data.save(s2s_data_save_name)", "_____no_output_____" ] ], [ [ "Structure", "_____no_output_____" ] ], [ [ "xb, yb = s2s_data.one_batch(); xb", "_____no_output_____" ] ], [ [ "Key:\n* 'c2m' = chord2melody translation\n * enc = chord\n * dec = melody\n* 'm2c' = next word input\n * enc = melody\n * dec = chord\n* 'pos' = timestepped postional encoding. Gives the model a better reference when translating\n\nNote: MultitaskTransformer trains both translations ('m2c' and 'c2m') at the same time. ", "_____no_output_____" ], [ "## 3. Initialize Model", "_____no_output_____" ] ], [ [ "# Load Data\nbatch_size = 2\nbptt = 128\n\nlm_data = load_data(data_path, data_save_name, \n bs=batch_size, bptt=bptt, encode_position=True,\n dl_tfms=mask_lm_tfm_pitchdur)\n\ns2s_data = load_data(data_path, s2s_data_save_name, \n bs=batch_size//2, bptt=bptt,\n preloader_cls=S2SPreloader, dl_tfms=melody_chord_tfm)\n\n# Combine both dataloaders so we can train multiple tasks at the same time\ndata = StackedDataBunch([lm_data, s2s_data])", "_____no_output_____" ], [ "# Create Model\nconfig = multitask_config(); config\n\nlearn = multitask_model_learner(data, config.copy())\n# learn.to_fp16(dynamic=True) # Enable for mixed precision", "_____no_output_____" ], [ "learn.model", "_____no_output_____" ] ], [ [ "# 4. Train", "_____no_output_____" ] ], [ [ "learn.fit_one_cycle(4)", "_____no_output_____" ], [ "learn.save('example')", "_____no_output_____" ] ], [ [ "## Predict\n\n---\nSee [Generate.ipynb](Generate.ipynb) to use a pretrained model and generate better predictions\n\n---", "_____no_output_____" ] ], [ [ "# midi_files = get_files(midi_path, '.mid', recurse=True)\nmidi_file = Path('data/midi/notebook_examples/single_bar_example.mid'); midi_file", "_____no_output_____" ], [ "next_word = nw_predict_from_midi(learn, midi_file, n_words=20, seed_len=8); next_word.show()", "_____no_output_____" ], [ "pred_melody = s2s_predict_from_midi(learn, midi_file, n_words=20, seed_len=4, pred_melody=True); pred_melody.show()", "_____no_output_____" ], [ "pred_notes = mask_predict_from_midi(learn, midi_file, predict_notes=True); pred_notes.show()", "_____no_output_____" ] ] ]

[ "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code" ] ]

4ad5443bd1ce83ade9ff4ddafbc5640007c48488

347,867

ipynb

Jupyter Notebook

.ipynb_checkpoints/take-home-exam-problem6-checkpoint.ipynb

wilsonify/AppliedRegression

9daf878140ae6e14ff94b59bde6a60f44b966819

[ "MIT" ]

null

.ipynb_checkpoints/take-home-exam-problem6-checkpoint.ipynb

wilsonify/AppliedRegression

9daf878140ae6e14ff94b59bde6a60f44b966819

[ "MIT" ]

null

.ipynb_checkpoints/take-home-exam-problem6-checkpoint.ipynb

wilsonify/AppliedRegression

9daf878140ae6e14ff94b59bde6a60f44b966819

[ "MIT" ]

null

354.243381

51,312

0.514102

[ [ [ "library(tidyverse)\nlibrary(broom) #glance\nlibrary(knitr) #kable\nlibrary(MASS) #stepAIC\nlibrary(data.table) #fread\nlibrary(car) #vif", "_____no_output_____" ] ], [ [ "Recommend a model to predict the y variable -- “Opening Weekend Gross” with the possible predictors (no interactions) -- Runtime, Production Budget, Critic Rating, Audience Rating, and/or Month of Release. \n\nExplain how you determined your model and why you recommended it over other models.\n\n", "_____no_output_____" ] ], [ [ "data6 <- fread('https://raw.githubusercontent.com/wilsonify/AppliedRegression/master/data/wide_release_movies.csv')", "_____no_output_____" ], [ "colnames(data6) <- colnames(data6) %>%\n tolower() %>% \n str_replace_all(' ','_') %>% \n str_replace_all('[()]','')", "_____no_output_____" ], [ "data6 %>% str()", "Classes ‘data.table’ and 'data.frame':\t680 obs. of 17 variables:\n $ title : chr \"Toy Story 3\" \"Alice in Wonderland (2010)\" \"Iron Man 2\" \"The Twilight Saga: Eclipse\" ...\n $ majorstudio : int 1 1 1 0 1 1 1 1 1 1 ...\n $ runtime : int 103 109 125 124 150 148 95 93 98 100 ...\n $ rating : chr \"G\" \"PG\" \"PG-13\" \"PG-13\" ...\n $ production_budget_in_millions: num 200 200 200 68 125 160 69 165 165 260 ...\n $ in_release : int 168 126 104 114 139 175 196 112 119 191 ...\n $ widest_release : int 4028 3739 4390 4468 4125 3792 3602 4386 4060 3603 ...\n $ opening_weekend_gross : int 110307189 116101023 128122480 64832191 125017372 62785337 56397125 70838207 43732319 48767052 ...\n $ total_gross : int 415004880 334191110 312433331 300531751 295983305 292576195 251513985 238736787 217581231 200821936 ...\n $ critic_rating : int 99 51 73 49 78 86 81 58 98 89 ...\n $ audience_rating : int 89 55 73 60 84 91 82 54 91 87 ...\n $ year : int 2010 2010 2010 2010 2010 2010 2010 2010 2010 2010 ...\n $ monthofrelease : int 6 3 5 6 11 7 7 5 3 11 ...\n $ logopeningweekendgross : num 18.5 18.6 18.7 18 18.6 ...\n $ logproductionbudget : num 19.1 19.1 19.1 18 18.6 ...\n $ logwidestrelease : num 8.3 8.23 8.39 8.4 8.32 ...\n $ logtotalgross : num 19.8 19.6 19.6 19.5 19.5 ...\n - attr(*, \".internal.selfref\")=<externalptr> \n" ], [ "subdata6 <- data6[,c('opening_weekend_gross','runtime','production_budget_in_millions','critic_rating','audience_rating','monthofrelease')]", "_____no_output_____" ], [ "subdata6 <- subdata6 %>% drop_na()", "_____no_output_____" ], [ "subdata6 %>% str()", "Classes ‘data.table’ and 'data.frame':\t642 obs. of 6 variables:\n $ opening_weekend_gross : int 110307189 116101023 128122480 64832191 125017372 62785337 56397125 70838207 43732319 48767052 ...\n $ runtime : int 103 109 125 124 150 148 95 93 98 100 ...\n $ production_budget_in_millions: num 200 200 200 68 125 160 69 165 165 260 ...\n $ critic_rating : int 99 51 73 49 78 86 81 58 98 89 ...\n $ audience_rating : int 89 55 73 60 84 91 82 54 91 87 ...\n $ monthofrelease : int 6 3 5 6 11 7 7 5 3 11 ...\n - attr(*, \".internal.selfref\")=<externalptr> \n" ], [ "full_fit <- lm(formula = opening_weekend_gross~\n runtime+\n production_budget_in_millions+\n critic_rating+\n audience_rating+\n monthofrelease\n ,data = subdata6)", "_____no_output_____" ], [ "summary(full_fit)", "_____no_output_____" ], [ "models <- interaction( c('runtime','production_budget_in_millions','critic_rating','audience_rating','monthofrelease')\n ,c('runtime','production_budget_in_millions','critic_rating','audience_rating','monthofrelease')\n ,c('runtime','production_budget_in_millions','critic_rating','audience_rating','monthofrelease')\n ,c('runtime','production_budget_in_millions','critic_rating','audience_rating','monthofrelease')\n ,c('runtime','production_budget_in_millions','critic_rating','audience_rating','monthofrelease')\n ,c('runtime','production_budget_in_millions','critic_rating','audience_rating','monthofrelease')\n , sep='+') %>% \n levels() %>% \n paste(\"opening_weekend_gross ~\",.)", "_____no_output_____" ], [ "scaled <- scale(subdata6) %>% as.data.frame()", "_____no_output_____" ], [ "scaled %>% cor() %>% round(2)", "_____no_output_____" ], [ "AICc_from_AIC <- function(AIC,fit) { \n n <- length(fit$residuals) \n k <- length(fit$coefficients) - 1 \n correction <- (2*k^2 + 2*k) / (n - k - 1)\n return ( AIC + correction )\n }\n \nresult <- data_frame()\nfor (form in models) {\n #print(form)\n fit <- lm(data=scaled, formula = as.formula(form)) \n glance_of_fit <- glance(fit) %>%\n mutate( model=form\n ,k = length(fit$coefficients) - 1 \n ,AICc = AICc_from_AIC(AIC,fit)) %>% \n dplyr::select(c( 'model' \n ,'k'\n ,'adj.r.squared'\n ,'AIC'\n ,'AICc'\n ,'BIC')\n )\n result <- rbind(result,glance_of_fit)\n}", "_____no_output_____" ], [ "result", "_____no_output_____" ], [ "rbind(result[which.max(result$adj.r.squared),]\n ,result[which.min(result$AIC),]\n ,result[which.min(result$AICc),]\n ,result[which.min(result$BIC),]\n)", "_____no_output_____" ], [ "fit <- lm(opening_weekend_gross ~ production_budget_in_millions+audience_rating, scaled)", "_____no_output_____" ], [ "summary(fit)", "_____no_output_____" ], [ "plot(y=studres(fit),x=fit$fitted.values)", "_____no_output_____" ], [ "plot(y=abs(studres(fit)),x= fit$fitted.values)", "_____no_output_____" ], [ "rfit <- lm(abs(studres(fit)) ~ fit$fitted.values)", "_____no_output_____" ], [ "rfit$fitted.values %>% str()", " Named num [1:642] 1.812 1.607 1.715 0.737 1.27 ...\n - attr(*, \"names\")= chr [1:642] \"1\" \"2\" \"3\" \"4\" ...\n" ], [ "scaled %>% str()", "'data.frame':\t642 obs. of 6 variables:\n $ opening_weekend_gross : num 3.54 3.77 4.23 1.78 4.11 ...\n $ runtime : num -0.204 0.179 1.202 1.138 2.8 ...\n $ production_budget_in_millions: num 2.847 2.847 2.847 0.364 1.436 ...\n $ critic_rating : num 1.8513 0.0246 0.8619 -0.0515 1.0521 ...\n $ audience_rating : num 1.902 -0.177 0.924 0.129 1.596 ...\n $ monthofrelease : num -0.176 -1.075 -0.476 -0.176 1.322 ...\n" ], [ "wfit <- lm(opening_weekend_gross ~ production_budget_in_millions+audience_rating, scaled,weights=(1/(rfit$fitted.values^2)))", "_____no_output_____" ], [ "summary(wfit)", "_____no_output_____" ], [ "plot(y=studres(wfit),x=wfit$fitted.values)", "_____no_output_____" ], [ "AICc_from_AIC <- function(AIC,fit) { \n n <- length(fit$residuals) \n k <- length(fit$coefficients) - 1 \n correction <- (2*k^2 + 2*k) / (n - k - 1)\n return ( AIC + correction )\n }\n \nresult <- data_frame()\nfor (form in models) {\n #print(form)\n fit <- lm(data=scaled, formula = as.formula(form)) \n rfit <- lm(abs(studres(fit)) ~ fit$fitted.values)\n wfit <- lm(data = scaled, as.formula(form), weights=(1/(rfit$fitted.values^2)))\n glance_of_fit <- glance(fit) %>%\n mutate( model=form\n ,k = length(fit$coefficients) - 1 \n ,AICc = AICc_from_AIC(AIC,fit)) %>% \n dplyr::select(c( 'model' \n ,'k'\n ,'adj.r.squared'\n ,'AIC'\n ,'AICc'\n ,'BIC')\n )\n result <- rbind(result,glance_of_fit)\n}", "_____no_output_____" ] ] ]

[ "code", "markdown", "code" ]

[ [ "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad544c0eb6101a46daaffcb90507b77cfa7cf39

179,678

ipynb

Jupyter Notebook

ElasticProperties.ipynb

richardotis/pycalphad-sandbox

43d8786eee8f279266497e9c5f4630d19c893092

[ "MIT" ]

1

2017-03-08T18:21:30.000Z

ElasticProperties.ipynb

richardotis/pycalphad-sandbox

43d8786eee8f279266497e9c5f4630d19c893092

[ "MIT" ]

null

ElasticProperties.ipynb

richardotis/pycalphad-sandbox

43d8786eee8f279266497e9c5f4630d19c893092

[ "MIT" ]

1

2018-11-03T01:31:57.000Z

492.268493

164,508

0.915972

[ [ [ "import matplotlib\nfrom matplotlib.axes import Axes\nfrom matplotlib.patches import Polygon\nfrom matplotlib.path import Path\nfrom matplotlib.ticker import NullLocator, Formatter, FixedLocator\nfrom matplotlib.transforms import Affine2D, BboxTransformTo, IdentityTransform\nfrom matplotlib.projections import register_projection\nimport matplotlib.spines as mspines\nimport matplotlib.axis as maxis\nimport matplotlib.pyplot as plt\n\nimport numpy as np\n\nclass TriangularAxes(Axes):\n \"\"\"\n A custom class for triangular projections.\n \"\"\"\n\n name = 'triangular'\n\n def __init__(self, *args, **kwargs):\n Axes.__init__(self, *args, **kwargs)\n self.set_aspect(1, adjustable='box', anchor='SW')\n self.cla()\n\n def _init_axis(self):\n self.xaxis = maxis.XAxis(self)\n self.yaxis = maxis.YAxis(self)\n self._update_transScale()\n\n def cla(self):\n \"\"\"\n Override to set up some reasonable defaults.\n \"\"\"\n # Don't forget to call the base class\n Axes.cla(self)\n \n x_min = 0\n y_min = 0\n x_max = 1\n y_max = 1\n x_spacing = 0.1\n y_spacing = 0.1\n self.xaxis.set_minor_locator(NullLocator())\n self.yaxis.set_minor_locator(NullLocator())\n self.xaxis.set_ticks_position('bottom')\n self.yaxis.set_ticks_position('left')\n Axes.set_xlim(self, x_min, x_max)\n Axes.set_ylim(self, y_min, y_max)\n self.xaxis.set_ticks(np.arange(x_min, x_max+x_spacing, x_spacing))\n self.yaxis.set_ticks(np.arange(y_min, y_max+y_spacing, y_spacing))\n\n def _set_lim_and_transforms(self):\n \"\"\"\n This is called once when the plot is created to set up all the\n transforms for the data, text and grids.\n \"\"\"\n # There are three important coordinate spaces going on here:\n #\n # 1. Data space: The space of the data itself\n #\n # 2. Axes space: The unit rectangle (0, 0) to (1, 1)\n # covering the entire plot area.\n #\n # 3. Display space: The coordinates of the resulting image,\n # often in pixels or dpi/inch.\n\n # This function makes heavy use of the Transform classes in\n # ``lib/matplotlib/transforms.py.`` For more information, see\n # the inline documentation there.\n\n # The goal of the first two transformations is to get from the\n # data space (in this case longitude and latitude) to axes\n # space. It is separated into a non-affine and affine part so\n # that the non-affine part does not have to be recomputed when\n # a simple affine change to the figure has been made (such as\n # resizing the window or changing the dpi).\n\n # 1) The core transformation from data space into\n # rectilinear space defined in the HammerTransform class.\n self.transProjection = IdentityTransform()\n # 2) The above has an output range that is not in the unit\n # rectangle, so scale and translate it so it fits correctly\n # within the axes. The peculiar calculations of xscale and\n # yscale are specific to a Aitoff-Hammer projection, so don't\n # worry about them too much.\n self.transAffine = Affine2D.from_values(\n 1., 0, 0.5, np.sqrt(3)/2., 0, 0)\n self.transAffinedep = Affine2D.from_values(\n 1., 0, -0.5, np.sqrt(3)/2., 0, 0)\n #self.transAffine = IdentityTransform()\n \n # 3) This is the transformation from axes space to display\n # space.\n self.transAxes = BboxTransformTo(self.bbox)\n\n # Now put these 3 transforms together -- from data all the way\n # to display coordinates. Using the '+' operator, these\n # transforms will be applied \"in order\". The transforms are\n # automatically simplified, if possible, by the underlying\n # transformation framework.\n self.transData = \\\n self.transProjection + \\\n self.transAffine + \\\n self.transAxes\n\n # The main data transformation is set up. Now deal with\n # gridlines and tick labels.\n\n # Longitude gridlines and ticklabels. The input to these\n # transforms are in display space in x and axes space in y.\n # Therefore, the input values will be in range (-xmin, 0),\n # (xmax, 1). The goal of these transforms is to go from that\n # space to display space. The tick labels will be offset 4\n # pixels from the equator.\n\n self._xaxis_pretransform = IdentityTransform()\n self._xaxis_transform = \\\n self._xaxis_pretransform + \\\n self.transData\n self._xaxis_text1_transform = \\\n Affine2D().scale(1.0, 0.0) + \\\n self.transData + \\\n Affine2D().translate(0.0, -20.0)\n self._xaxis_text2_transform = \\\n Affine2D().scale(1.0, 0.0) + \\\n self.transData + \\\n Affine2D().translate(0.0, -4.0)\n\n # Now set up the transforms for the latitude ticks. The input to\n # these transforms are in axes space in x and display space in\n # y. Therefore, the input values will be in range (0, -ymin),\n # (1, ymax). The goal of these transforms is to go from that\n # space to display space. The tick labels will be offset 4\n # pixels from the edge of the axes ellipse.\n\n self._yaxis_transform = self.transData\n yaxis_text_base = \\\n self.transProjection + \\\n (self.transAffine + \\\n self.transAxes)\n self._yaxis_text1_transform = \\\n yaxis_text_base + \\\n Affine2D().translate(-8.0, 0.0)\n self._yaxis_text2_transform = \\\n yaxis_text_base + \\\n Affine2D().translate(8.0, 0.0)\n\n def get_xaxis_transform(self,which='grid'):\n assert which in ['tick1','tick2','grid']\n return self._xaxis_transform\n\n def get_xaxis_text1_transform(self, pad):\n return self._xaxis_text1_transform, 'bottom', 'center'\n\n def get_xaxis_text2_transform(self, pad):\n return self._xaxis_text2_transform, 'top', 'center'\n\n def get_yaxis_transform(self,which='grid'):\n assert which in ['tick1','tick2','grid']\n return self._yaxis_transform\n\n def get_yaxis_text1_transform(self, pad):\n return self._yaxis_text1_transform, 'center', 'right'\n\n def get_yaxis_text2_transform(self, pad):\n return self._yaxis_text2_transform, 'center', 'left'\n\n def _gen_axes_spines(self):\n dep_spine = mspines.Spine.linear_spine(self,\n 'right')\n # Fix dependent axis to be transformed the correct way\n dep_spine.set_transform(self.transAffinedep + self.transAxes)\n return {'left':mspines.Spine.linear_spine(self,\n 'left'),\n 'bottom':mspines.Spine.linear_spine(self,\n 'bottom'),\n\t\t'right':dep_spine}\n\n def _gen_axes_patch(self):\n \"\"\"\n Override this method to define the shape that is used for the\n background of the plot. It should be a subclass of Patch.\n Any data and gridlines will be clipped to this shape.\n \"\"\"\n\n return Polygon([[0,0], [0.5,np.sqrt(3)/2], [1,0]], closed=True)\n\n # Interactive panning and zooming is not supported with this projection,\n # so we override all of the following methods to disable it.\n def can_zoom(self):\n \"\"\"\n Return True if this axes support the zoom box\n \"\"\"\n return False\n def start_pan(self, x, y, button):\n pass\n def end_pan(self):\n pass\n def drag_pan(self, button, key, x, y):\n pass\n\n\n# Now register the projection with matplotlib so the user can select\n# it.\nregister_projection(TriangularAxes)", "_____no_output_____" ], [ "import pycalphad.io.tdb_keywords\npycalphad.io.tdb_keywords.TDB_PARAM_TYPES.extend(['EM', 'BULK', 'SHEAR', 'C11', 'C12', 'C44'])\nfrom pycalphad import Database, Model, calculate, equilibrium\nimport numpy as np\nimport pycalphad.variables as v\nimport sympy\nfrom tinydb import where\n\nclass ElasticModel(Model):\n def build_phase(self, dbe, phase_name, symbols, param_search):\n phase = dbe.phases[phase_name]\n self.models['ref'] = self.reference_energy(phase, param_search)\n self.models['idmix'] = self.ideal_mixing_energy(phase, param_search)\n self.models['xsmix'] = self.excess_mixing_energy(phase, param_search)\n self.models['mag'] = self.magnetic_energy(phase, param_search)\n\n # Here is where we add our custom contribution\n # EM, BULK, SHEAR, C11, C12, C44\n for prop in ['EM', 'BULK', 'SHEAR', 'C11', 'C12', 'C44']:\n prop_param_query = (\n (where('phase_name') == phase.name) & \\\n (where('parameter_type') == prop) & \\\n (where('constituent_array').test(self._array_validity))\n )\n prop_val = self.redlich_kister_sum(phase, param_search, prop_param_query)\n setattr(self, prop, prop_val)\n\n # Extra code necessary for compatibility with order-disorder model\n ordered_phase_name = None\n disordered_phase_name = None\n try:\n ordered_phase_name = phase.model_hints['ordered_phase']\n disordered_phase_name = phase.model_hints['disordered_phase']\n except KeyError:\n pass\n if ordered_phase_name == phase_name:\n self.models['ord'] = self.atomic_ordering_energy(dbe,\n disordered_phase_name,\n ordered_phase_name)", "_____no_output_____" ], [ "dbf = Database('ElasticTi.tdb')\nmod = ElasticModel(dbf, ['TI', 'MO', 'NB', 'VA'], 'BCC_A2')\nsymbols = dict([(sympy.Symbol(s), val) for s, val in dbf.symbols.items()])\nmod.EM = mod.EM.xreplace(symbols)\nx1 = np.linspace(0,1, num=100)\nx2 = np.linspace(0,1, num=100)\nmesh = np.meshgrid(x1, x2)\nX = mesh[0]\nY = mesh[1]\nmesh_arr = np.array(mesh)\nmesh_arr = np.moveaxis(mesh_arr, 0, 2)\ndep_col = 1 - np.sum(mesh_arr, axis=-1, keepdims=True)\nmesh_arr = np.concatenate((mesh_arr, dep_col), axis=-1)\nmesh_arr = np.concatenate((mesh_arr, np.ones(mesh_arr.shape[:-1] + (1,))), axis=-1)\norig_shape = tuple(mesh_arr.shape[:-1])\nmesh_arr = mesh_arr.reshape(-1, mesh_arr.shape[-1])\nmesh_arr[np.any(mesh_arr < 0, axis=-1), :] = np.nan\nres = calculate(dbf, ['TI', 'MO', 'NB', 'VA'], 'BCC_A2', T=300, P=101325,\n model=mod, output='EM', points=mesh_arr)\nres_EM = res.EM.values.reshape(orig_shape)", "_____no_output_____" ], [ "%matplotlib inline\nimport matplotlib.pyplot as plt\n\nfig = plt.figure(figsize=(12,12))\nax = fig.gca(projection='triangular')\nCS = ax.contour(X, Y, res_EM, levels=list(range(-10, 310, 10)), linewidths=4, cmap='cool')\nax.clabel(CS, inline=1, fontsize=13, fmt='%1.0f')\n#PCM=ax.get_children()[0] #get the mappable, the 1st and the 2nd are the x and y axes\n#plt.colorbar(PCM, ax=ax)\nax.set_xlabel('Mole Fraction Mo', fontsize=18)\nax.set_ylabel('Mole Fraction Nb', fontsize=18, rotation=60, labelpad=-180)\nax.tick_params(axis='both', which='major', labelsize=18)\nax.tick_params(axis='both', which='minor', labelsize=18)\nfig.savefig('TiMoNb-EM.pdf')", "_____no_output_____" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code" ] ]

4ad54f073e7e71f751a8684cfb49e13859295fc4

317,747

ipynb

Jupyter Notebook

escape_mutations/ranks_analysis.ipynb

sberbank-ai-lab/viral-mutation

c0d53cfb88a9ae60c24cfc083b0a2610a31c27e3

[ "MIT" ]

null

escape_mutations/ranks_analysis.ipynb

sberbank-ai-lab/viral-mutation

c0d53cfb88a9ae60c24cfc083b0a2610a31c27e3

[ "MIT" ]

null

escape_mutations/ranks_analysis.ipynb

sberbank-ai-lab/viral-mutation

c0d53cfb88a9ae60c24cfc083b0a2610a31c27e3

[ "MIT" ]

null

39.693567

117,051

0.575036

[ [ [ "import numpy as np\nimport pandas as pd\nimport plotly.graph_objects as go\nfrom sklearn.metrics import roc_auc_score", "_____no_output_____" ], [ "ranks = pd.read_csv('ranks_with_escapes.csv')", "_____no_output_____" ], [ "escapes_df = ranks[ranks.is_escape == 'ESCAPE']\nnot_escapes_df = ranks[ranks.is_escape != 'ESCAPE']", "_____no_output_____" ], [ "fig = go.Figure()\n\nfig.add_trace(go.Scatter(\n x=not_escapes_df.predicted,\n y=not_escapes_df.sem_log10,\n marker=dict(\n color=not_escapes_df.rank_sum,\n colorscale='Viridis',\n showscale=True\n ),\n name='not escapes',\n mode='markers'\n))\n\nfig.add_trace(go.Scatter(\n x=escapes_df.predicted,\n y=escapes_df.sem_log10,\n name='escapes',\n mode='markers',\n marker_color='red'\n))\n\nfig.update_layout(\n title='Escape prediction (Sars-Cov2)',\n xaxis_title='Grammaticality, log10(p(xi | x[N] \\ i))',\n yaxis_title='Semantic change, log10(Δz)'\n)\n\nfig.show()", "_____no_output_____" ], [ "ranks = ranks.sort_values(by='rank_sum', ascending=True)\n\ntrue_labels = ranks.is_escape == 'ESCAPE'\npredicted = []\npredicted.extend([False] * len(not_escapes_df))\npredicted.extend([True] * len(escapes_df))\n\nprint(roc_auc_score(true_labels, predicted))", "0.5073858157271851\n" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code", "code" ] ]

4ad557e76472b159cab2b56ede6db4b5718bb150

6,086

ipynb

Jupyter Notebook

ArkoudaDistance.ipynb

bradcray/ArkoudaNotebooks

aae93045b9c03216160527c50799a862e0182a2d

[ "MIT" ]

null

ArkoudaDistance.ipynb

bradcray/ArkoudaNotebooks

aae93045b9c03216160527c50799a862e0182a2d

[ "MIT" ]

null

ArkoudaDistance.ipynb

bradcray/ArkoudaNotebooks

aae93045b9c03216160527c50799a862e0182a2d

[ "MIT" ]

null

22.540741

154

0.534177

[ [ [ "## Arkouda example of cosine distance and euclidean distance\n- Two random arkouda int64 pdarrays are created then the distance is measured between them...\n- The cosine and euclidean distance functions are compared against the scipy variants for correctness\n\nArkouda functions used:\n- `ak.connect`\n- `ak.randint`\n- `ak.sum`\n- `ak.pdarray.__mul__`\n- `ak.pdarray.to_ndarray`\n- `ak.disconnect`\n- `ak.shutdown`", "_____no_output_____" ] ], [ [ "import arkouda as ak", "_____no_output_____" ], [ "import math\nfrom scipy.spatial import distance\nimport numpy as np", "_____no_output_____" ] ], [ [ "### Connect to the Arkouda server", "_____no_output_____" ] ], [ [ "# connect to the arkopuda server using the connect_url which the server prints out\nak.connect(connect_url=\"tcp://localhost:5555\")", "_____no_output_____" ] ], [ [ "### Create two pdarrays and fill with random integers", "_____no_output_____" ] ], [ [ "# create two in64 pdarrays and add them\na = ak.randint(0,10,100)\nb = ak.randint(0,10,100)\n\nprint(a+b)", "_____no_output_____" ] ], [ [ "### Check for pdarray", "_____no_output_____" ] ], [ [ "def ak_check_pda(u):\n if not isinstance(u, ak.pdarray):\n raise TypeError(\"argument must be a pdarray\")", "_____no_output_____" ] ], [ [ "### Dot Product of two Arkouda pdarrays\n$u \\cdot v = \\sum_i{u_i v_i}$", "_____no_output_____" ] ], [ [ "# define the dot product or two arkouda pdarrays\ndef ak_dot(u, v):\n ak_check_pda(u)\n ak_check_pda(v)\n return ak.sum(u*v)", "_____no_output_____" ] ], [ [ "### Magnatude ($L_2$ norm) of an Arkouda pdarray\n$\\|u\\|_2 = \\sqrt{\\sum_i{u_i^2}}$", "_____no_output_____" ] ], [ [ "# define the magnitude/L_2-norm of arkouda pdarray\ndef ak_mag2(u):\n ak_check_pda(u)\n return math.sqrt(ak_dot(u,u))", "_____no_output_____" ] ], [ [ "### Cosine Distance of two Arkouda pdarrays\n$D_C = 1 - \\cos(\\theta) = 1 - \\frac{u \\cdot v}{\\|u\\|_2\\|v\\|_2} = 1 - \\frac{\\sum_i{u_i v_i}}{\\sqrt{\\sum_i{u_i^2}}\\sqrt{\\sum_i{v_i^2}}}$", "_____no_output_____" ] ], [ [ "# define the cosine distance of two arkouda pdarrays\n# should function similarly to scipy.spatial.distance.cosine\ndef ak_cos_dist(u, v):\n ak_check_pda(u)\n ak_check_pda(v)\n return (1.0 - ak_dot(u,v)/(ak_mag2(u)*ak_mag2(v)))", "_____no_output_____" ] ], [ [ "### Euclidean Distance of two Arkouda pdarrays\n$D_E = \\|u-v\\|_2$", "_____no_output_____" ] ], [ [ "# define the euclidean distance of two arkouda pdarrays\n# should function similarly to scipy.spatial.distance.euclidean\ndef ak_euc_dist(u, v):\n ak_check_pda(u)\n ak_check_pda(v)\n return (ak_mag2(u-v))", "_____no_output_____" ] ], [ [ "### Measure cosine distance and check against scipy", "_____no_output_____" ] ], [ [ "# check the arkouda version against the scipy version\nd1 = ak_cos_dist(a,b)\nd2 = distance.cosine(a.to_ndarray(), b.to_ndarray())\nprint(d1,d2)\nprint(np.allclose(d1,d2))", "_____no_output_____" ] ], [ [ "### Measure euclidean distance and check against scipy", "_____no_output_____" ] ], [ [ "# check the arkouda version against the scipy version\nd1 = ak_euc_dist(a,b)\nd2 = distance.euclidean(a.to_ndarray(), b.to_ndarray())\nprint(d1,d2)\nprint(np.allclose(d1,d2))", "_____no_output_____" ] ], [ [ "### Disconnect from Arkouda server or Shutdown Arkouda server", "_____no_output_____" ] ], [ [ "# disconnect from the the arkouda server\n#ak.disconnect()\n# shutdown the arkouda server\n#ak.shutdown()", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad576069331a3a4e0f3af5ff0161095b789cd90

107,807

ipynb

Jupyter Notebook

data_wrangling.ipynb

73Shivam/Machine_Learning

d338a9a8f0d0f20cbff6f11287886f30fcd640bf

[ "MIT" ]

null

data_wrangling.ipynb

73Shivam/Machine_Learning

d338a9a8f0d0f20cbff6f11287886f30fcd640bf

[ "MIT" ]

null

data_wrangling.ipynb

73Shivam/Machine_Learning

d338a9a8f0d0f20cbff6f11287886f30fcd640bf

[ "MIT" ]

null

155.341499

25,791

0.590871

[ [ [ "import numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt", "_____no_output_____" ], [ "filename = \"https://s3-api.us-geo.objectstorage.softlayer.net/cf-courses-data/CognitiveClass/DA0101EN/auto.csv\"\nheaders = [\"symboling\",\"normalized-losses\",\"make\",\"fuel-type\",\"aspiration\", \"num-of-doors\",\"body-style\",\n \"drive-wheels\",\"engine-location\",\"wheel-base\", \"length\",\"width\",\"height\",\"curb-weight\",\"engine-type\",\n \"num-of-cylinders\", \"engine-size\",\"fuel-system\",\"bore\",\"stroke\",\"compression-ratio\",\"horsepower\",\n \"peak-rpm\",\"city-mpg\",\"highway-mpg\",\"price\"]", "_____no_output_____" ], [ "df = pd.read_csv(filename, names= headers)\ndf.head(10)", "_____no_output_____" ] ], [ [ "# Table of content\n- Identify and handle missing values\n - Identify missing values\n - Deal with missing values\n - Correct data format\n- Data standardization\n- Data Normalization (centering/scaling)\n- Binning\n- Indicator variable", "_____no_output_____" ], [ "# Identify and handle missing values", "_____no_output_____" ] ], [ [ "df.replace(\"?\", np.nan, inplace=True)", "_____no_output_____" ], [ "df['normalized-losses'].isnull().value_counts()", "_____no_output_____" ], [ "# loop for knowing missing values in each columns\nfor col in df.columns:\n print(df[col].isnull().value_counts())\n print()", "False 205\nName: symboling, dtype: int64\n\nFalse 164\nTrue 41\nName: normalized-losses, dtype: int64\n\nFalse 205\nName: make, dtype: int64\n\nFalse 205\nName: fuel-type, dtype: int64\n\nFalse 205\nName: aspiration, dtype: int64\n\nFalse 203\nTrue 2\nName: num-of-doors, dtype: int64\n\nFalse 205\nName: body-style, dtype: int64\n\nFalse 205\nName: drive-wheels, dtype: int64\n\nFalse 205\nName: engine-location, dtype: int64\n\nFalse 205\nName: wheel-base, dtype: int64\n\nFalse 205\nName: length, dtype: int64\n\nFalse 205\nName: width, dtype: int64\n\nFalse 205\nName: height, dtype: int64\n\nFalse 205\nName: curb-weight, dtype: int64\n\nFalse 205\nName: engine-type, dtype: int64\n\nFalse 205\nName: num-of-cylinders, dtype: int64\n\nFalse 205\nName: engine-size, dtype: int64\n\nFalse 205\nName: fuel-system, dtype: int64\n\nFalse 201\nTrue 4\nName: bore, dtype: int64\n\nFalse 201\nTrue 4\nName: stroke, dtype: int64\n\nFalse 205\nName: compression-ratio, dtype: int64\n\nFalse 203\nTrue 2\nName: horsepower, dtype: int64\n\nFalse 203\nTrue 2\nName: peak-rpm, dtype: int64\n\nFalse 205\nName: city-mpg, dtype: int64\n\nFalse 205\nName: highway-mpg, dtype: int64\n\nFalse 201\nTrue 4\nName: price, dtype: int64\n\n" ] ], [ [ "<h3 id=\"deal_missing_values\">Deal with missing data</h3>\nHow to deal with missing data?\n\n<ol>\n <li>drop data \n a. drop the whole row \n b. drop the whole column\n </li>\n <li>replace data \n a. replace it by mean \n b. replace it by frequency \n c. replace it based on other functions\n </li>\n</ol>", "_____no_output_____" ] ], [ [ "mean_norm_loss = df['normalized-losses'].astype('float').mean()\ndf['normalized-losses'].replace(np.nan, mean_norm_loss, inplace=True)", "_____no_output_____" ], [ "mean_bore_loss = df['bore'].astype('float').mean()\ndf['bore'].replace(np.nan, mean_bore_loss, inplace=True)", "_____no_output_____" ], [ "mean_stroke_loss = df['stroke'].astype('float').mean()\ndf['stroke'].replace(np.nan, mean_stroke_loss, inplace=True)", "_____no_output_____" ], [ "mean_horsepower_loss = df['horsepower'].astype('float').mean()\ndf['horsepower'].replace(np.nan, mean_horsepower_loss, inplace=True)", "_____no_output_____" ], [ "mean_rpm_loss = df['peak-rpm'].astype('float').mean()\ndf['peak-rpm'].replace(np.nan, mean_rpm_loss, inplace=True)", "_____no_output_____" ], [ "df['num-of-doors'].value_counts()", "_____no_output_____" ], [ "df['num-of-doors'].replace(np.nan, 'four', inplace=True)", "_____no_output_____" ], [ "df['peak-rpm'] = df['peak-rpm'].astype('int')", "_____no_output_____" ], [ "df['horsepower'] = df['horsepower'].astype(int)", "_____no_output_____" ], [ "df['stroke'] = df['stroke'].astype(float)", "_____no_output_____" ], [ "df['bore'] = df['bore'].astype(float)", "_____no_output_____" ], [ "df['normalized-losses'] = df['normalized-losses'].astype(int)", "_____no_output_____" ], [ "df.info()", "<class 'pandas.core.frame.DataFrame'>\nRangeIndex: 205 entries, 0 to 204\nData columns (total 26 columns):\n # Column Non-Null Count Dtype \n--- ------ -------------- ----- \n 0 symboling 205 non-null int64 \n 1 normalized-losses 205 non-null int32 \n 2 make 205 non-null object \n 3 fuel-type 205 non-null object \n 4 aspiration 205 non-null object \n 5 num-of-doors 205 non-null object \n 6 body-style 205 non-null object \n 7 drive-wheels 205 non-null object \n 8 engine-location 205 non-null object \n 9 wheel-base 205 non-null float64\n 10 length 205 non-null float64\n 11 width 205 non-null float64\n 12 height 205 non-null float64\n 13 curb-weight 205 non-null int64 \n 14 engine-type 205 non-null object \n 15 num-of-cylinders 205 non-null object \n 16 engine-size 205 non-null int64 \n 17 fuel-system 205 non-null object \n 18 bore 205 non-null float64\n 19 stroke 205 non-null float64\n 20 compression-ratio 205 non-null float64\n 21 horsepower 205 non-null int32 \n 22 peak-rpm 205 non-null int32 \n 23 city-mpg 205 non-null int64 \n 24 highway-mpg 205 non-null int64 \n 25 price 201 non-null object \ndtypes: float64(7), int32(3), int64(5), object(11)\nmemory usage: 39.4+ KB\n" ], [ "df.dropna(subset=['price'], inplace= True)", "_____no_output_____" ], [ "df.price = df.price.astype(int)", "_____no_output_____" ], [ "df.info()", "<class 'pandas.core.frame.DataFrame'>\nInt64Index: 201 entries, 0 to 204\nData columns (total 26 columns):\n # Column Non-Null Count Dtype \n--- ------ -------------- ----- \n 0 symboling 201 non-null int64 \n 1 normalized-losses 201 non-null int32 \n 2 make 201 non-null object \n 3 fuel-type 201 non-null object \n 4 aspiration 201 non-null object \n 5 num-of-doors 201 non-null object \n 6 body-style 201 non-null object \n 7 drive-wheels 201 non-null object \n 8 engine-location 201 non-null object \n 9 wheel-base 201 non-null float64\n 10 length 201 non-null float64\n 11 width 201 non-null float64\n 12 height 201 non-null float64\n 13 curb-weight 201 non-null int64 \n 14 engine-type 201 non-null object \n 15 num-of-cylinders 201 non-null object \n 16 engine-size 201 non-null int64 \n 17 fuel-system 201 non-null object \n 18 bore 201 non-null float64\n 19 stroke 201 non-null float64\n 20 compression-ratio 201 non-null float64\n 21 horsepower 201 non-null int32 \n 22 peak-rpm 201 non-null int32 \n 23 city-mpg 201 non-null int64 \n 24 highway-mpg 201 non-null int64 \n 25 price 201 non-null int32 \ndtypes: float64(7), int32(4), int64(5), object(10)\nmemory usage: 49.3+ KB\n" ], [ "df.to_csv('automobiles.csv')", "_____no_output_____" ] ], [ [ "# data standardization", "_____no_output_____" ] ], [ [ "df['city-mpg'] # needs to be coverted", "_____no_output_____" ], [ "df['city-kml'] = 235/df['city-mpg']\ndf.head()", "_____no_output_____" ] ], [ [ "# data normalization or scaling", "_____no_output_____" ] ], [ [ "df['peak-rpm'] = df['peak-rpm']/ df['peak-rpm'].max()", "_____no_output_____" ], [ "df['city-kml'] = df['city-kml'] / df['city-kml'].max()", "_____no_output_____" ], [ "df.head()", "_____no_output_____" ] ], [ [ "# binning ", "_____no_output_____" ] ], [ [ "group_names = ['low','medium','high']", "_____no_output_____" ], [ "bins = np.linspace(df.horsepower.min(), df.horsepower.max(),4 )", "_____no_output_____" ], [ "df['hp-binned'] = pd.cut(df.horsepower,bins, labels=group_names,include_lowest=True)", "_____no_output_____" ], [ "df['hp-binned'].value_counts()", "_____no_output_____" ], [ "df['hp-binned'].value_counts().plot(kind='bar')", "_____no_output_____" ], [ "df['price'].plot.hist(bins=5)", "_____no_output_____" ] ], [ [ "# indicator variable or dummy variable", "_____no_output_____" ] ], [ [ "df['fuel-type'].unique()", "_____no_output_____" ], [ "dummy_var1 = pd.get_dummies(df['fuel-type'])", "_____no_output_____" ], [ "dummy_var2 = pd.get_dummies(df['make'])", "_____no_output_____" ], [ "# merge the dummy var in df and drop the orignal col\ndf = pd.concat([df, dummy_var1[:-1], dummy_var2[:-1]], axis=1)", "_____no_output_____" ], [ "df.drop(columns=['fuel-type','make'],inplace=True)", "_____no_output_____" ], [ "df.info()", "<class 'pandas.core.frame.DataFrame'>\nInt64Index: 201 entries, 0 to 204\nData columns (total 50 columns):\n # Column Non-Null Count Dtype \n--- ------ -------------- ----- \n 0 symboling 201 non-null int64 \n 1 normalized-losses 201 non-null int32 \n 2 aspiration 201 non-null object \n 3 num-of-doors 201 non-null object \n 4 body-style 201 non-null object \n 5 drive-wheels 201 non-null object \n 6 engine-location 201 non-null object \n 7 wheel-base 201 non-null float64 \n 8 length 201 non-null float64 \n 9 width 201 non-null float64 \n 10 height 201 non-null float64 \n 11 curb-weight 201 non-null int64 \n 12 engine-type 201 non-null object \n 13 num-of-cylinders 201 non-null object \n 14 engine-size 201 non-null int64 \n 15 fuel-system 201 non-null object \n 16 bore 201 non-null float64 \n 17 stroke 201 non-null float64 \n 18 compression-ratio 201 non-null float64 \n 19 horsepower 201 non-null int32 \n 20 peak-rpm 201 non-null float64 \n 21 city-mpg 201 non-null int64 \n 22 highway-mpg 201 non-null int64 \n 23 price 201 non-null int32 \n 24 city-kml 201 non-null float64 \n 25 hp-binned 201 non-null category\n 26 diesel 200 non-null float64 \n 27 gas 200 non-null float64 \n 28 alfa-romero 200 non-null float64 \n 29 audi 200 non-null float64 \n 30 bmw 200 non-null float64 \n 31 chevrolet 200 non-null float64 \n 32 dodge 200 non-null float64 \n 33 honda 200 non-null float64 \n 34 isuzu 200 non-null float64 \n 35 jaguar 200 non-null float64 \n 36 mazda 200 non-null float64 \n 37 mercedes-benz 200 non-null float64 \n 38 mercury 200 non-null float64 \n 39 mitsubishi 200 non-null float64 \n 40 nissan 200 non-null float64 \n 41 peugot 200 non-null float64 \n 42 plymouth 200 non-null float64 \n 43 porsche 200 non-null float64 \n 44 renault 200 non-null float64 \n 45 saab 200 non-null float64 \n 46 subaru 200 non-null float64 \n 47 toyota 200 non-null float64 \n 48 volkswagen 200 non-null float64 \n 49 volvo 200 non-null float64 \ndtypes: category(1), float64(33), int32(3), int64(5), object(8)\nmemory usage: 86.5+ KB\n" ], [ "df['num-of-doors'].nunique() # not neccesssary to convert to dummy if you have 2 values only", "_____no_output_____" ], [ "df['engine-type'].nunique()", "_____no_output_____" ], [ "df['num-of-doors'].replace('two',0,inplace=True)\ndf['num-of-doors'].replace('four',1,inplace=True)\ndf.head()", "_____no_output_____" ], [ "pd.get_dummies(df['engine-type'])", "_____no_output_____" ] ] ]

[ "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad57efd85994953f0ac56de1bad94c0859650af

16,118

ipynb

Jupyter Notebook

pandas/01_Getting_&_Knowing_Your_Data/World Food Facts/Exercises_with_solutions.ipynb

eric999j/Udemy_Python_Hand_On

7a985b3e2c9adfd3648d240af56ac00bb916c3ad

[ "Apache-2.0" ]

1

2020-12-31T18:03:34.000Z

pandas/01_Getting_&_Knowing_Your_Data/World Food Facts/Exercises_with_solutions.ipynb

cntfk2017/Udemy_Python_Hand_On

52f2a5585bfdea95d893f961c8c21844072e93c7

[ "Apache-2.0" ]

null

pandas/01_Getting_&_Knowing_Your_Data/World Food Facts/Exercises_with_solutions.ipynb

cntfk2017/Udemy_Python_Hand_On

52f2a5585bfdea95d893f961c8c21844072e93c7

[ "Apache-2.0" ]

2

2019-09-23T14:26:48.000Z

2020-05-25T07:09:26.000Z

29.903525

199

0.405944

[ [ [ "# Ex1 - Getting and knowing your Data", "_____no_output_____" ], [ "### Step 1. Go to https://www.kaggle.com/openfoodfacts/world-food-facts", "_____no_output_____" ], [ "### Step 2. Download the dataset to your computer and unzip it.", "_____no_output_____" ] ], [ [ "import pandas as pd\nimport numpy as np", "_____no_output_____" ] ], [ [ "### Step 3. Use the csv file and assign it to a dataframe called food", "_____no_output_____" ] ], [ [ "food = pd.read_csv('/Users/guilhermeoliveira/Desktop/world-food-facts/FoodFacts.csv')", "//anaconda/lib/python2.7/site-packages/IPython/core/interactiveshell.py:2723: DtypeWarning: Columns (0,3,5,27,36) have mixed types. Specify dtype option on import or set low_memory=False.\n interactivity=interactivity, compiler=compiler, result=result)\n" ] ], [ [ "### Step 4. See the first 5 entries", "_____no_output_____" ] ], [ [ "food.head()", "_____no_output_____" ] ], [ [ "### Step 5. What is the number of observations in the dataset?", "_____no_output_____" ] ], [ [ "food.shape #will give you both (observations/rows, columns)\nfood.shape[0] #will give you only the observations/rows number", "_____no_output_____" ] ], [ [ "### Step 6. What is the number of columns in the dataset?", "_____no_output_____" ] ], [ [ "print food.shape #will give you both (observations/rows, columns)\nprint food.shape[1] #will give you only the columns number\n\n#OR\n\nfood.info() #Columns: 159 entries", "(65503, 159)\n159\n<class 'pandas.core.frame.DataFrame'>\nRangeIndex: 65503 entries, 0 to 65502\nColumns: 159 entries, code to nutrition_score_uk_100g\ndtypes: float64(103), object(56)\nmemory usage: 79.5+ MB\n" ] ], [ [ "### Step 7. Print the name of all the columns.", "_____no_output_____" ] ], [ [ "food.columns", "_____no_output_____" ] ], [ [ "### Step 8. What is the name of 105th column?", "_____no_output_____" ] ], [ [ "food.columns[104]", "_____no_output_____" ] ], [ [ "### Step 9. What is the type of the observations of the 105th column?", "_____no_output_____" ] ], [ [ "food.dtypes['glucose_100g']", "_____no_output_____" ] ], [ [ "### Step 10. How is the dataset indexed?", "_____no_output_____" ] ], [ [ "food.index", "_____no_output_____" ] ], [ [ "### Step 11. What is the product name of the 19th observation?", "_____no_output_____" ] ], [ [ "food.values[18][7]", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad5848028006acc9dbcaa6e16b2ed522ead54a0

16,169

ipynb

Jupyter Notebook

week-5/webinar-5/wsc_5.ipynb

mattDevigili/dms-smm695

c1789c5667e2854449bd49cf0290e5cee149bdcf

[ "MIT" ]

34

2020-05-14T05:04:34.000Z

2021-11-07T20:37:04.000Z

week-5/webinar-5/wsc_5.ipynb

mattDevigili/dms-smm695

c1789c5667e2854449bd49cf0290e5cee149bdcf

[ "MIT" ]

null

week-5/webinar-5/wsc_5.ipynb

mattDevigili/dms-smm695

c1789c5667e2854449bd49cf0290e5cee149bdcf

[ "MIT" ]

23

2020-05-21T10:24:10.000Z

2022-03-24T23:37:30.000Z

27.266442

133

0.461315

[ [ [ "# Set-up", "_____no_output_____" ] ], [ [ "# libraries\nimport re\nimport numpy as np\nimport pandas as pd\nfrom pymongo import MongoClient", "_____no_output_____" ], [ "# let's connect to the localhost\nclient = MongoClient()\n\n# let's create a database \ndb = client.moma\n\n# collection\nartworks = db.artworks\n\n# print connection\nprint(\"\"\"\nDatabase\n==========\n{}\n\nCollection\n==========\n{}\n\"\"\".format(db, artworks), flush=True\n)", "_____no_output_____" ] ], [ [ "## Data", "_____no_output_____" ], [ "![MoMa](https://images.musement.com/cover/0001/31/moma-museum-of-modern-art-tickets-tours-jpg_header-30520.jpeg?&q=60&fit=crop)", "_____no_output_____" ] ], [ [ "df = pd.read_csv('https://media.githubusercontent.com/media/MuseumofModernArt/collection/master/Artworks.csv')\n\ndf.info()", "_____no_output_____" ] ], [ [ "# Loading", "_____no_output_____" ] ], [ [ "%%time \n# slow loading of data\nd = {}\n\nfor i in df.index:\n d = {\n \"_id\": str(df.loc[i, \"Cataloged\"]) + str(df.loc[i, \"ObjectID\"]),\n \"Title\": df.loc[i, \"Title\"],\n \"Date\": df.loc[i, \"Date\"],\n \"Artist\": {\n \"Name\": df.loc[i, \"Artist\"],\n \"Bio\": df.loc[i, \"ArtistBio\"],\n \"Nationality\": df.loc[i, \"Nationality\"],\n \"Birth\": df.loc[i, \"BeginDate\"],\n \"Death\": df.loc[i, \"EndDate\"],\n \"Gender\": df.loc[i, \"Gender\"]\n },\n \"Characteristics\":{\n \"Medium\": df.loc[i,'Medium'], \n \"Dimensions\": df.loc[i,'Dimensions'],\n \"Circumference\": df.loc[i,'Circumference (cm)'], \n \"Depth\": df.loc[i,'Depth (cm)'], \n \"Diameter\": df.loc[i,'Diameter (cm)'], \n \"Height\": df.loc[i,'Height (cm)'],\n \"Length\": df.loc[i,'Length (cm)'], \n \"Weight\": df.loc[i,'Weight (kg)'], \n \"Width\": df.loc[i,'Width (cm)'], \n \"Seat Height\": df.loc[i,'Seat Height (cm)'],\n \"Duration\": df.loc[i,'Duration (sec.)']\n },\n \"Acquisition\": {\n \"Date\": df.loc[i, \"DateAcquired\"],\n \"CreditLine\": df.loc[i, \"CreditLine\"],\n \"Number\": df.loc[i, \"AccessionNumber\"]\n },\n \"Classification\": df.loc[i, \"Classification\"],\n \"Department\": df.loc[i, \"Department\"],\n \"URL\": df.loc[i, \"URL\"], \n \"ThumbnailURL\": df.loc[i, \"ThumbnailURL\"]\n }\n artworks.insert_one(d)", "_____no_output_____" ], [ "# for further reference https://docs.mongodb.com/manual/reference/command/collStats/\nstats = db.command(\"collstats\", \"artworks\")\ns0 = stats.get('size')/10**6\n\nprint(\"\"\"\nNamespace: {}\n\nDocument Count: {}\n\nSize: {}\n\n\"\"\".format(stats.get('ns'), stats.get('count'), s0), flush=True)", "_____no_output_____" ] ], [ [ "## Cleaning", "_____no_output_____" ] ], [ [ "# get key names\nl = []\nfor i in d.keys():\n try:\n for b in d.get(str(i)).keys():\n l.append(str(i) + '.' + str(b))\n except:\n l.append(i)", "_____no_output_____" ], [ "# unset NaN fields\nfor i in l:\n update = artworks.update_many({str(i):np.nan},{\"$unset\": {str(i):\"\"}})\n print(\"\"\"\n Key: {}\n Matched: {}\n Modified: {}\n ------------\n \"\"\".format(i, update.matched_count, update.modified_count), flush=True)", "_____no_output_____" ], [ "# for further reference https://docs.mongodb.com/manual/reference/command/collStats/\nstats = db.command(\"collstats\", \"artworks\")\ns1 = stats.get('size')/10**6\n\nprint(\"\"\"\nNamespace: {}\n\nDocument Count: {}\n\nSize: {}\n\nVar. Size: {}\n\n\"\"\".format(stats.get('ns'), stats.get('count'), s1, round(s0-s1, 2)), flush=True)", "_____no_output_____" ] ], [ [ "## Further Cleaning", "_____no_output_____" ] ], [ [ "# change data type\nupdate = artworks.update_many({\"Date\":{\"$regex\": '^[0-9]*$'}}, [{ \"$set\": { \"Date\": { \"$toInt\": \"$Date\" } } }])\n\nprint(\"\"\"\n Key: {}\n Matched: {}\n Modified: {}\n ------------\n \"\"\".format(\"Date\", update.matched_count, update.modified_count), flush=True)", "_____no_output_____" ], [ "# create an array field to store ranges\nfor i in artworks.find({\"Date\":{\"$regex\": '^[0-9]{4}-[0-9]{4}$'}}):\n date = i.get('Date').split('-')\n a = int(date[0])\n b = int(date[1])\n id = i.get('_id')\n update = artworks.update_one({\"_id\": str(id)},{\"$set\": {\"Date\": [a, b]}})\n print(update.matched_count, update.modified_count)\n \nfor i in artworks.find({\"Date\":{\"$regex\": '^[0-9]{4}–[0-9]{4}$'}}):\n date = i.get('Date').split('–')\n a = int(date[0])\n b = int(date[1])\n id = i.get('_id')\n update = artworks.update_one({\"_id\": str(id)},{\"$set\": {\"Date\": [a, b]}})\n print(update.matched_count, update.modified_count)\n\nfor i in artworks.find({\"Date\": {\"$regex\": '^[0-9]{4}-[0-9]{2}$'}}, {\"Date\": 1}):\n date = i.get('Date').split('-')\n a = int(date[0])\n b = int(date[0][0] + date[0][1] + date[1])\n id = i.get('_id')\n update = artworks.update_one({\"_id\": str(id)},{\"$set\": {\"Date\": [a, b]}})\n print(update.matched_count, update.modified_count)\n\nfor i in artworks.find({\"Date\": {\"$regex\": '^[0-9]{4}–[0-9]{2}$'}}, {\"Date\": 1}):\n date = i.get('Date').split('–')\n a = int(date[0])\n b = int(date[0][0] + date[0][1]+ date[1])\n id = i.get('_id')\n update = artworks.update_one({\"_id\": str(id)},{\"$set\": {\"Date\": [a, b]}})\n print(update.matched_count, update.modified_count)\n\n# perform some further cleaning\nfor i in artworks.find({\"Date\":{\"$regex\": '^c. [0-9]{4}$'}}):\n date = i.get('Date').split(' ')\n b = int(date[1])\n id = i.get('_id')\n update = artworks.update_one({\"_id\": str(id)},{\"$set\": {\"Date\": b}})\n print(update.matched_count, update.modified_count)", "_____no_output_____" ], [ "# remove Unknown or n.d.\nupdate = artworks.update_many({\"Date\": {\"$in\": [\"n.d.\", \"Unknown\", \"unknown\"]}}, {\"$unset\": {\"Date\": \"\"}})\nprint(\"\"\"\n Matched: {}\n Modified: {}\n \"\"\".format(update.matched_count, update.modified_count), flush=True)", "_____no_output_____" ], [ "for i in artworks.find({\"Date\": {\"$type\": \"string\"}}, {\"Date\":1}):\n print(i)", "_____no_output_____" ] ], [ [ "# Aggregation and loading", "_____no_output_____" ] ], [ [ "# collection\nartw = db.artw\n\n# print connection\nprint(\"\"\"\nDatabase\n==========\n{}\n\nCollection\n==========\n{}\n\"\"\".format(db, artw), flush=True\n)", "_____no_output_____" ], [ "# df to dict\ndf.rename(columns={'Duration (sec.)': 'Duration (sec)'}, inplace=True)\ndd = df.to_dict('records')\n\ndd[0]", "_____no_output_____" ], [ "%%time\n# insert array\ninsert = artw.insert_many(dd)\n\n# define the pipeline\npipeline = [\n {\"$project\": \n {\n \"_id\": {\"$concat\": [\"$Cataloged\", {\"$toString\": \"$ObjectID\"}]},\n \"Title\": \"$Title\",\n \"Date\": \"$Date\",\n \"Artist\": {\n \"Name\": \"$Artist\", \n 'Bio': \"$ArtistBio\",\n 'Nationality': \"$Nationality\",\n \"Birth\": \"$BeginDate\",\n \"Death\": \"$EndDate\", \n \"Gender\": \"$Gender\",\n },\n \"Characteristics\":{\n \"Medium\": '$Medium', \n \"Dimensions\": '$Dimensions',\n \"Circumference\": '$Circumference (cm)', \n \"Depth\": '$Depth (cm)', \n \"Diameter\": '$Diameter (cm)', \n \"Height\": '$Height (cm)',\n \"Length\": '$Length (cm)', \n \"Weight\": '$Weight (kg)', \n \"Width\": '$Width (cm)', \n \"Seat Height\": '$Seat Height (cm)',\n \"Duration\": '$Duration (sec)'\n },\n \"Acquisition\": {\n \"Date\": \"$DateAcquired\",\n \"CreditLine\": \"$CreditLine\",\n \"Number\": \"$AccessionNumber\"\n },\n \"Classification\": \"$Classification\",\n \"Department\": \"$Department\",\n \"URL\": \"$URL\", \n \"ThumbnailURL\": \"$ThumbnailURL\"\n }\n },\n { \"$out\" : \"artw\" }\n]\n\n# perform the aggregation\nagr = artw.aggregate(pipeline)", "_____no_output_____" ], [ "# unset field with null values\n[artw.update_many({str(i):np.nan},{\"$unset\": {str(i):\"\"}}) for i in l]", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code" ] ]

4ad58ca5bb0779e9e6f8659b73d834ebdaf8729d

61,898

ipynb

Jupyter Notebook

Metropolis Hastings.ipynb

sharanry/probabilisticML_experiments

78d8f8c3718634fef5069325927e8fdf76775b56

[ "MIT" ]

null

Metropolis Hastings.ipynb

sharanry/probabilisticML_experiments

78d8f8c3718634fef5069325927e8fdf76775b56

[ "MIT" ]

null

Metropolis Hastings.ipynb

sharanry/probabilisticML_experiments

78d8f8c3718634fef5069325927e8fdf76775b56

[ "MIT" ]

null

240.848249

24,968

0.918398

[ [ [ "# Implementation of Metropolis Hastings Sampling Method", "_____no_output_____" ] ], [ [ "import numpy as np\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nimport sys\nfrom tqdm import tqdm", "_____no_output_____" ], [ "# Defining Normal distribution class with sampling, pdf and log probability functionalities.\nclass Normal():\n def __init__(self, mu=0, sigma=1):\n self.mu = mu\n self.sigma = sigma\n\n def sample(self, length=1):\n x = np.random.uniform(0,1,[2, length])\n #Box-Müller method\n z = np.sqrt(-2*np.log(x[0]))*np.cos(2*np.pi*x[1])\n return z * self.sigma**2 + self.mu\n \n def pdf(self, data):\n return np.exp(-(data-self.mu)**2/(2*self.sigma**2))/np.sqrt(2*np.pi*self.sigma**2)\n \n def logp(self, data):\n #log probability function\n return -(data-self.mu)**2/(2*self.sigma**2) - 0.5*np.log(2*np.pi*self.sigma**2)", "_____no_output_____" ], [ "# defining Metropolis Hastings Sampler which takes in the data, prior distribution, proposal width(sigma) \ndef sampler(data, mu_init, sigma, draws, prior_dist):\n mu_current=mu_init\n posterior = [float(mu_current)]\n with tqdm(total=draws) as pbar:\n for i in range(draws):\n mu_proposed = Normal(mu_current, sigma).sample()\n\n #log likelihoods for acceptance\n \n # the numerator\n log_proposal = Normal(mu_proposed, sigma).logp(data).sum() + prior_dist.logp(mu_proposed)\n \n # the denominator\n log_current = Normal(mu_current, sigma).logp(data).sum() + prior_dist.logp(mu_current)\n\n # log acceptance probability\n log_accept = min(0, log_proposal - log_current)\n\n # condition for accepting proposed mu\n if np.log(np.random.rand()) < log_accept:\n mu_current = mu_proposed\n\n posterior.append(float(mu_current))\n pbar.update(1)\n \n return posterior", "_____no_output_____" ], [ "# Generating Data\ndata = Normal().sample(1000)", "_____no_output_____" ], [ "# defining prior mu distribution\nmu = Normal()\n\n# sampling the postrior\npost = sampler(data=data, mu_init=-1, sigma=0.5, draws=15000, prior_dist=mu)", "100%|██████████| 15000/15000 [00:01<00:00, 14136.19it/s]\n" ], [ "plt.plot(post)", "_____no_output_____" ], [ "plt.plot(post[100:])", "_____no_output_____" ], [ "fig = sns.distplot(post[100:])\nfig.set(xlabel='mu', ylabel='frequency')", "/home/sharan/anaconda3/envs/pymc3/lib/python3.6/site-packages/matplotlib/axes/_axes.py:6462: UserWarning: The 'normed' kwarg is deprecated, and has been replaced by the 'density' kwarg.\n warnings.warn(\"The 'normed' kwarg is deprecated, and has been \"\n" ] ] ]

[ "markdown", "code" ]

[ [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad5a4b422d07243add007fd48fbc950b3b56dcf

12,449

ipynb

Jupyter Notebook

notebooks/simple-atlas-opendata-plot.ipynb

oshadura/opendata-higgs-discovery

47b4ac787a69b9580fc1f72948d3af0bdc20759b

[ "MIT" ]

null

notebooks/simple-atlas-opendata-plot.ipynb

oshadura/opendata-higgs-discovery

47b4ac787a69b9580fc1f72948d3af0bdc20759b

[ "MIT" ]

null

notebooks/simple-atlas-opendata-plot.ipynb

oshadura/opendata-higgs-discovery

47b4ac787a69b9580fc1f72948d3af0bdc20759b

[ "MIT" ]

null

127.030612

10,272

0.892682

[ [ [ "from func_adl_servicex import ServiceXSourceUpROOT\nfrom hist import Hist\nimport mplhep as mpl\nimport awkward as ak\n\n# This is CMS data...\n# mpl.style.use(mpl.style.CMS)", "_____no_output_____" ], [ "files = {\n 'ggH125_ZZ4lep':\n {\n 'files': ['root://eospublic.cern.ch//eos/opendata/atlas/OutreachDatasets/2020-01-22/4lep/MC/mc_345060.ggH125_ZZ4lep.4lep.root'],\n 'treename': 'ggH125_ZZ4lep'\n },\n}", "_____no_output_____" ], [ "data = ServiceXSourceUpROOT(files['ggH125_ZZ4lep']['files'], 'mini', backend_name='open_uproot') \\\n .Select(\"lambda e: {'lep_pt': e['lep_pt']}\") \\\n .AsAwkwardArray() \\\n .value()", "_____no_output_____" ], [ "h = (Hist.new\n .Reg(50, 0, 200, name='mu_pt', label='Muon Track $p_T$')\n .Int64()\n )\nh.fill(ak.flatten(data['lep_pt'])/1000.0)\n_ = h.plot()", "_____no_output_____" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code" ] ]

4ad5acfb7ee4baa708d2c2c0e784118b10650594

77,985

ipynb

Jupyter Notebook

app/python_scripts/deep_dives/conflict_res.ipynb

vlm-wpi/MQP

e9e56740f152f4cde17ae53cb66040514f20343a

[ "MIT" ]

null

app/python_scripts/deep_dives/conflict_res.ipynb

vlm-wpi/MQP

e9e56740f152f4cde17ae53cb66040514f20343a

[ "MIT" ]

null

app/python_scripts/deep_dives/conflict_res.ipynb

vlm-wpi/MQP

e9e56740f152f4cde17ae53cb66040514f20343a

[ "MIT" ]

null

237.759146

57,640

0.88894

[ [ [ "# this is for the data from the three runs where we try \n# each of the conflict resolutions\n\n# imports and reading csv\nimport pandas as pd\nimport numpy as np\nimport matplotlib.pyplot as plt\nfrom matplotlib.gridspec import GridSpec\n\nimport os\nimport sys\n\n#get correct path for files\n__file__ = 'conflict_res'\nabsolutepath = os.path.abspath(__file__)\nprint(absolutepath) #should be MQP/app/python_scripts/deep_dives/conflict_res\nfileDirectory = os.path.dirname(absolutepath)\nprint(fileDirectory) #should be MQP/app/python_scripts/deep_dives\n#Path of parent directory\nparentDirectory = os.path.dirname(fileDirectory)\nprint(parentDirectory) #should be MQP/app/python_scripts\nappDirectory = os.path.dirname(parentDirectory)\nprint(appDirectory) #should be MQP/app\n#Navigate to all_data directory\nall_data = os.path.join(appDirectory, 'all_data') #the directory going to, change if something different\nprint(all_data) #should be MQP/app/all_data\npathOutput1 = os.path.join(all_data, 'data_1.csv') #the file reading from, change if something different\nprint(pathOutput1) #should be MQP/app/all_data/data_1.csv\npathOutput3 = os.path.join(all_data, 'data_3.csv') #the file reading from, change if something different\nprint(pathOutput3) #should be MQP/app/all_data/data_3.csv\npathOutput4 = os.path.join(all_data, 'data_4.csv') #the file reading from, change if something different\nprint(pathOutput4) #should be MQP/app/all_data/data_4.csv\npathOutput12 = os.path.join(all_data, 'data_12.csv') #the file reading from, change if something different\nprint(pathOutput12) #should be MQP/app/all_data/data_12.csv\n\n\ndata1 = pd.read_csv(pathOutput1)\ndata3 = pd.read_csv(pathOutput3)\ndata4 = pd.read_csv(pathOutput4)\ndata12 = pd.read_csv(pathOutput12)\n", "/Users/victoriamirecki/Downloads/MQP/app/python_scripts/deep_dives/conflict_res\n/Users/victoriamirecki/Downloads/MQP/app/python_scripts/deep_dives\n/Users/victoriamirecki/Downloads/MQP/app/python_scripts\n/Users/victoriamirecki/Downloads/MQP/app\n/Users/victoriamirecki/Downloads/MQP/app/all_data\n/Users/victoriamirecki/Downloads/MQP/app/all_data/data_1.csv\n/Users/victoriamirecki/Downloads/MQP/app/all_data/data_3.csv\n/Users/victoriamirecki/Downloads/MQP/app/all_data/data_4.csv\n/Users/victoriamirecki/Downloads/MQP/app/all_data/data_12.csv\n" ], [ "# general initializations\nx=['move, exit, others', 'move, exit and reset, others', 'move, exit', 'none']\n", "_____no_output_____" ], [ "# y1 initializations\ny1=[]\ny1e=[]\n\navg_1_1=np.mean(data1['total_exit_time'])\navg_1_3=np.mean(data3['total_exit_time'])\navg_1_4=np.mean(data4['total_exit_time'])\navg_1_12=np.mean(data12['total_exit_time'])\n\ny1.append(avg_1_1)\ny1.append(avg_1_3)\ny1.append(avg_1_4)\ny1.append(avg_1_12)\n\ny1e.append(np.std(data1['total_exit_time']))\ny1e.append(np.std(data3['total_exit_time']))\ny1e.append(np.std(data4['total_exit_time']))\ny1e.append(np.std(data12['total_exit_time']))\n\n# print('y1', y1)", "_____no_output_____" ], [ "# y2 initializations\ny2=[]\ny2e=[]\n\navg_2_1=np.mean(data1['avg_exit_time'])\navg_2_3=np.mean(data3['avg_exit_time'])\navg_2_4=np.mean(data4['avg_exit_time'])\navg_2_12=np.mean(data12['avg_exit_time'])\n\ny2.append(avg_2_1)\ny2.append(avg_2_3)\ny2.append(avg_2_4)\ny2.append(avg_2_12)\n\ny2e.append(np.std(data1['avg_exit_time']))\ny2e.append(np.std(data3['avg_exit_time']))\ny2e.append(np.std(data4['avg_exit_time']))\ny2e.append(np.std(data12['avg_exit_time']))\n\n# print('y2', y2)", "_____no_output_____" ], [ "# y3 initializations\ny3=[]\ny3e=[]\n\navg_3_1=np.mean(data1['avg_collisions_total'])\navg_3_3=np.mean(data3['avg_collisions_total'])\navg_3_4=np.mean(data4['avg_collisions_total'])\navg_3_12=np.mean(data12['avg_collisions_total'])\n\ny3.append(avg_3_1)\ny3.append(avg_3_3)\ny3.append(avg_3_4)\ny3.append(avg_3_12)\n\ny3e.append(np.std(data1['avg_collisions_total']))\ny3e.append(np.std(data3['avg_collisions_total']))\ny3e.append(np.std(data4['avg_collisions_total']))\ny3e.append(np.std(data12['avg_collisions_total']))\n\n# print('y3', y3)", "_____no_output_____" ], [ "# y4 initializations\ny4=[]\ny4e=[]\n\navg_4_1=np.mean(data1['total_avg_occ_all_time'])\navg_4_3=np.mean(data3['total_avg_occ_all_time'])\navg_4_4=np.mean(data4['total_avg_occ_all_time'])\navg_4_12=np.mean(data12['total_avg_occ_all_time'])\n\ny4.append(avg_4_1)\ny4.append(avg_4_3)\ny4.append(avg_4_4)\ny4.append(avg_4_12)\n\ny4e.append(np.std(data1['total_avg_occ_all_time']))\ny4e.append(np.std(data3['total_avg_occ_all_time']))\ny4e.append(np.std(data4['total_avg_occ_all_time']))\ny4e.append(np.std(data12['total_avg_occ_all_time']))\n\n# print('y4', y4)\n", "_____no_output_____" ], [ "# y5 initializations\ny5=[]\ny5e=[]\n\navg_5_1=np.mean(data1['evaluation_metric'])\navg_5_3=np.mean(data3['evaluation_metric'])\navg_5_4=np.mean(data4['evaluation_metric'])\navg_5_12=np.mean(data12['evaluation_metric'])\n\ny5.append(avg_5_1)\ny5.append(avg_5_3)\ny5.append(avg_5_4)\ny5.append(avg_5_12)\n\ny5e.append(np.std(data1['evaluation_metric']))\ny5e.append(np.std(data3['evaluation_metric']))\ny5e.append(np.std(data4['evaluation_metric']))\ny5e.append(np.std(data12['evaluation_metric']))\n\n# print('y4', y4)\n", "_____no_output_____" ], [ "# building the bar graphs\n\n\nfig=plt.figure(figsize=(30, 12))\ngs = GridSpec(nrows=2, ncols=3)\nax0 = fig.add_subplot(gs[0, 0])\nplt.bar(x,y1,yerr=y1e, align='center', alpha=0.5, ecolor='black', capsize=10)\nplt.title('Average Total Exit Time Based on Conflict Res')\nplt.xlabel('Conflict Resolution Strategy')\nplt.ylabel('Average Total Exit Time')\nax1 = fig.add_subplot(gs[1, 0])\nplt.bar(x,y2,yerr=y2e, align='center', alpha=0.5, ecolor='black', capsize=10)\nplt.title('Mean of Average Exit Time Based on Conflict Res')\nplt.xlabel('Conflict Resolution Strategy')\nplt.ylabel('Mean of Average Exit Time')\nax2 = fig.add_subplot(gs[0, 1])\nplt.bar(x,y3,yerr=y3e, align='center', alpha=0.5, ecolor='black', capsize=10)\nplt.title('Mean of Average Collisions Based on Conflict Res')\nplt.xlabel('Conflict Resolution Strategy')\nplt.ylabel('Mean of Average Collisions')\nax3 = fig.add_subplot(gs[1, 1])\nplt.bar(x,y4,yerr=y4e, align='center', alpha=0.5, ecolor='black', capsize=10)\nplt.title('Mean of Average Area Occupancy Based on Conflict Res')\nplt.xlabel('Conflict Resolution Strategy')\nplt.ylabel('Mean of Average Area Occupancy')\nax3 = fig.add_subplot(gs[:, 2])\nplt.bar(x,y5,yerr=y5e, align='center', alpha=0.5, ecolor='black', capsize=10)\nplt.title('Average Evaluation Metric Based on Conflict Res')\nplt.xlabel('Conflict Resolution Strategy')\nplt.ylabel('Average Evaluation Metric')\n", "_____no_output_____" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad5b4c884564ca12585888cfc5adc49724b46a6

292,987

ipynb

Jupyter Notebook

docs/source/notebooks/GP-smoothing.ipynb

JvParidon/pymc3

f60dc1db74df3c54d707761e9dc054a7fdf0435c

[ "Apache-2.0" ]

2

2020-05-29T07:10:45.000Z

2021-04-07T06:43:52.000Z

docs/source/notebooks/GP-smoothing.ipynb

JvParidon/pymc3

f60dc1db74df3c54d707761e9dc054a7fdf0435c

[ "Apache-2.0" ]

null

docs/source/notebooks/GP-smoothing.ipynb

JvParidon/pymc3

f60dc1db74df3c54d707761e9dc054a7fdf0435c

[ "Apache-2.0" ]

1

2019-01-02T09:02:18.000Z

794.00271

82,482

0.940741

[ [ [ "# Gaussian Process (GP) smoothing\n\nThis example deals with the case when we want to **smooth** the observed data points $(x_i, y_i)$ of some 1-dimensional function $y=f(x)$, by finding the new values $(x_i, y'_i)$ such that the new data is more \"smooth\" (see more on the definition of smoothness through allocation of variance in the model description below) when moving along the $x$ axis. \n\nIt is important to note that we are **not** dealing with the problem of interpolating the function $y=f(x)$ at the unknown values of $x$. Such problem would be called \"regression\" not \"smoothing\", and will be considered in other examples.\n\nIf we assume the functional dependency between $x$ and $y$ is **linear** then, by making the independence and normality assumptions about the noise, we can infer a straight line that approximates the dependency between the variables, i.e. perform a linear regression. We can also fit more complex functional dependencies (like quadratic, cubic, etc), if we know the functional form of the dependency in advance.\n\nHowever, the **functional form** of $y=f(x)$ is **not always known in advance**, and it might be hard to choose which one to fit, given the data. For example, you wouldn't necessarily know which function to use, given the following observed data. Assume you haven't seen the formula that generated it:", "_____no_output_____" ] ], [ [ "%pylab inline\nfigsize(12, 6);", "Populating the interactive namespace from numpy and matplotlib\n" ], [ "import numpy as np\nimport scipy.stats as stats\n\nx = np.linspace(0, 50, 100)\ny = (np.exp(1.0 + np.power(x, 0.5) - np.exp(x/15.0)) + \n np.random.normal(scale=1.0, size=x.shape))\n\nplot(x, y);\nxlabel(\"x\");\nylabel(\"y\");\ntitle(\"Observed Data\");", "_____no_output_____" ] ], [ [ "### Let's try a linear regression first\n\nAs humans, we see that there is a non-linear dependency with some noise, and we would like to capture that dependency. If we perform a linear regression, we see that the \"smoothed\" data is less than satisfactory:", "_____no_output_____" ] ], [ [ "plot(x, y);\nxlabel(\"x\");\nylabel(\"y\");\n\nlin = stats.linregress(x, y)\nplot(x, lin.intercept + lin.slope * x);\ntitle(\"Linear Smoothing\");", "_____no_output_____" ] ], [ [ "### Linear regression model recap\n\nThe linear regression assumes there is a linear dependency between the input $x$ and output $y$, sprinkled with some noise around it so that for each observed data point we have:\n\n$$ y_i = a + b\\, x_i + \\epsilon_i $$\n\nwhere the observation errors at each data point satisfy:\n\n$$ \\epsilon_i \\sim N(0, \\sigma^2) $$\n\nwith the same $\\sigma$, and the errors are independent:\n\n$$ cov(\\epsilon_i, \\epsilon_j) = 0 \\: \\text{ for } i \\neq j $$\n\nThe parameters of this model are $a$, $b$, and $\\sigma$. It turns out that, under these assumptions, the maximum likelihood estimates of $a$ and $b$ don't depend on $\\sigma$. Then $\\sigma$ can be estimated separately, after finding the most likely values for $a$ and $b$.", "_____no_output_____" ], [ "### Gaussian Process smoothing model\n\nThis model allows departure from the linear dependency by assuming that the dependency between $x$ and $y$ is a Brownian motion over the domain of $x$. This doesn't go as far as assuming a particular functional dependency between the variables. Instead, by **controlling the standard deviation of the unobserved Brownian motion** we can achieve different levels of smoothness of the recovered functional dependency at the original data points. \n\nThe particular model we are going to discuss assumes that the observed data points are **evenly spaced** across the domain of $x$, and therefore can be indexed by $i=1,\\dots,N$ without the loss of generality. The model is described as follows:\n\n\\begin{equation}\n\\begin{aligned}\nz_i & \\sim \\mathcal{N}(z_{i-1} + \\mu, (1 - \\alpha)\\cdot\\sigma^2) \\: \\text{ for } i=2,\\dots,N \\\\\nz_1 & \\sim ImproperFlat(-\\infty,\\infty) \\\\\ny_i & \\sim \\mathcal{N}(z_i, \\alpha\\cdot\\sigma^2)\n\\end{aligned}\n\\end{equation}\n\nwhere $z$ is the hidden Brownian motion, $y$ is the observed data, and the total variance $\\sigma^2$ of each ovservation is split between the hidden Brownian motion and the noise in proportions of $1 - \\alpha$ and $\\alpha$ respectively, with parameter $0 < \\alpha < 1$ specifying the degree of smoothing.\n\nWhen we estimate the maximum likelihood values of the hidden process $z_i$ at each of the data points, $i=1,\\dots,N$, these values provide an approximation of the functional dependency $y=f(x)$ as $\\mathrm{E}\\,[f(x_i)] = z_i$ at the original data points $x_i$ only. Therefore, again, the method is called smoothing and not regression.", "_____no_output_____" ], [ "### Let's describe the above GP-smoothing model in PyMC3", "_____no_output_____" ] ], [ [ "import pymc3 as pm\nfrom theano import shared\nfrom pymc3.distributions.timeseries import GaussianRandomWalk\nfrom scipy import optimize", "_____no_output_____" ] ], [ [ "Let's create a model with a shared parameter for specifying different levels of smoothing. We use very wide priors for the \"mu\" and \"tau\" parameters of the hidden Brownian motion, which you can adjust according to your application.", "_____no_output_____" ] ], [ [ "LARGE_NUMBER = 1e5\n\nmodel = pm.Model()\nwith model:\n smoothing_param = shared(0.9)\n mu = pm.Normal(\"mu\", sigma=LARGE_NUMBER)\n tau = pm.Exponential(\"tau\", 1.0/LARGE_NUMBER)\n z = GaussianRandomWalk(\"z\",\n mu=mu,\n tau=tau / (1.0 - smoothing_param), \n shape=y.shape)\n obs = pm.Normal(\"obs\", \n mu=z, \n tau=tau / smoothing_param, \n observed=y)", "_____no_output_____" ] ], [ [ "Let's also make a helper function for inferring the most likely values of $z$:", "_____no_output_____" ] ], [ [ "def infer_z(smoothing):\n with model:\n smoothing_param.set_value(smoothing)\n res = pm.find_MAP(vars=[z], fmin=optimize.fmin_l_bfgs_b)\n return res['z']", "_____no_output_____" ] ], [ [ "Please note that in this example, we are only looking at the MAP estimate of the unobserved variables. We are not really interested in inferring the posterior distributions. Instead, we have a control parameter $\\alpha$ which lets us allocate the variance between the hidden Brownian motion and the noise. Other goals and/or different models may require sampling to obtain the posterior distributions, but for our goal a MAP estimate will suffice.\n\n### Exploring different levels of smoothing\n\nLet's try to allocate 50% variance to the noise, and see if the result matches our expectations.", "_____no_output_____" ] ], [ [ "smoothing = 0.5\nz_val = infer_z(smoothing)\n\nplot(x, y);\nplot(x, z_val);\ntitle(\"Smoothing={}\".format(smoothing));", "_____no_output_____" ] ], [ [ "It appears that the variance is split evenly between the noise and the hidden process, as expected. \n\nLet's try gradually increasing the smoothness parameter to see if we can obtain smoother data:", "_____no_output_____" ] ], [ [ "smoothing = 0.9\nz_val = infer_z(smoothing)\n\nplot(x, y);\nplot(x, z_val);\ntitle(\"Smoothing={}\".format(smoothing));", "_____no_output_____" ] ], [ [ "### Smoothing \"to the limits\"\n\nBy increading the smoothing parameter, we can gradually make the inferred values of the hidden Brownian motion approach the average value of the data. This is because as we increase the smoothing parameter, we allow less and less of the variance to be allocated to the Brownian motion, so eventually it aproaches the process which almost doesn't change over the domain of $x$:", "_____no_output_____" ] ], [ [ "fig, axes = subplots(2, 2)\n\nfor ax, smoothing in zip(axes.ravel(), [0.95, 0.99, 0.999, 0.9999]):\n\n z_val = infer_z(smoothing)\n\n ax.plot(x, y)\n ax.plot(x, z_val)\n ax.set_title('Smoothing={:05.4f}'.format(smoothing))", "_____no_output_____" ] ], [ [ "This example originally contributed by: Andrey Kuzmenko, http://github.com/akuz", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ] ]

4ad5bf557c87478c0b82d1dcceec088f69a76a79

42,443

ipynb

Jupyter Notebook

machine_learning/logistic_regression/sklearn_lr.ipynb

ningchi/book_notes

c6f8001f7d5f873896c4b3a8b1409b21ef33c328

[ "CC0-1.0" ]

7

2017-12-31T12:10:42.000Z

2021-11-10T15:49:34.000Z

machine_learning/logistic_regression/sklearn_lr.ipynb

ningchi/book_notes

c6f8001f7d5f873896c4b3a8b1409b21ef33c328

[ "CC0-1.0" ]

1

2017-12-05T13:04:14.000Z

2017-12-07T16:24:50.000Z

machine_learning/logistic_regression/sklearn_lr.ipynb

ningchi/book_notes

c6f8001f7d5f873896c4b3a8b1409b21ef33c328

[ "CC0-1.0" ]

2

2017-06-27T07:19:28.000Z

2017-11-19T08:57:35.000Z

116.282192

28,884

0.601654

[ [ [ "# %load /Users/facai/Study/book_notes/preconfig.py\n%matplotlib inline\n\nimport matplotlib.pyplot as plt\nimport seaborn as sns\n\nfrom IPython.display import SVG", "_____no_output_____" ] ], [ [ "逻辑回归在scikit-learn中的实现简介\n==============================", "_____no_output_____" ], [ "分析用的代码版本信息：\n\n```bash\n~/W/g/scikit-learn ❯❯❯ git log -n 1\ncommit d161bfaa1a42da75f4940464f7f1c524ef53484f\nAuthor: John B Nelson <[email protected]>\nDate: Thu May 26 18:36:37 2016 -0400\n\n Add missing double quote (#6831)\n```", "_____no_output_____" ], [ "### 0. 总纲\n\n下面是sklearn中逻辑回归的构成情况:", "_____no_output_____" ] ], [ [ "SVG(\"./res/sklearn_lr.svg\")", "_____no_output_____" ] ], [ [ "如[逻辑回归在spark中的实现简介](./spark_ml_lr.ipynb)中分析一样，主要把精力定位到算法代码上，即寻优算子和损失函数。", "_____no_output_____" ], [ "### 1. 寻优算子\n\nsklearn支持liblinear, sag, lbfgs和newton-cg四种寻优算子，其中lbfgs属于scipy包，liblinear属于LibLinear库，剩下两种由sklearn自己实现。代码很好定位，逻辑也很明了，不多说：\n\n```python\n 704 if solver == 'lbfgs':\n 705 try:\n 706 w0, loss, info = optimize.fmin_l_bfgs_b(\n 707 func, w0, fprime=None,\n 708 args=(X, target, 1. / C, sample_weight),\n 709 iprint=(verbose > 0) - 1, pgtol=tol, maxiter=max_iter)\n 710 except TypeError:\n 711 # old scipy doesn't have maxiter\n 712 w0, loss, info = optimize.fmin_l_bfgs_b(\n 713 func, w0, fprime=None,\n 714 args=(X, target, 1. / C, sample_weight),\n 715 iprint=(verbose > 0) - 1, pgtol=tol)\n 716 if info[\"warnflag\"] == 1 and verbose > 0:\n 717 warnings.warn(\"lbfgs failed to converge. Increase the number \"\n 718 \"of iterations.\")\n 719 try:\n 720 n_iter_i = info['nit'] - 1\n 721 except:\n 722 n_iter_i = info['funcalls'] - 1\n 723 elif solver == 'newton-cg':\n 724 args = (X, target, 1. / C, sample_weight)\n 725 w0, n_iter_i = newton_cg(hess, func, grad, w0, args=args,\n 726 maxiter=max_iter, tol=tol)\n 727 elif solver == 'liblinear':\n 728 coef_, intercept_, n_iter_i, = _fit_liblinear(\n 729 X, target, C, fit_intercept, intercept_scaling, None,\n 730 penalty, dual, verbose, max_iter, tol, random_state,\n 731 sample_weight=sample_weight)\n 732 if fit_intercept:\n 733 w0 = np.concatenate([coef_.ravel(), intercept_])\n 734 else:\n 735 w0 = coef_.ravel()\n 736\n 737 elif solver == 'sag':\n 738 if multi_class == 'multinomial':\n 739 target = target.astype(np.float64)\n 740 loss = 'multinomial'\n 741 else:\n 742 loss = 'log'\n 743\n 744 w0, n_iter_i, warm_start_sag = sag_solver(\n 745 X, target, sample_weight, loss, 1. / C, max_iter, tol,\n 746 verbose, random_state, False, max_squared_sum, warm_start_sag)\n```", "_____no_output_____" ], [ "### 2. 损失函数\n\n#### 2.1 二分类\n\n二分类的损失函数和导数由`_logistic_loss_and_grad`实现，运算逻辑和[逻辑回归算法简介和Python实现](./demo.ipynb)是相同的，不多说。", "_____no_output_____" ], [ "#### 2.2 多分类\n\nsklearn的多分类支持ovr (one vs rest，一对多)和multinominal两种方式。\n\n\n##### 2.2.0 ovr\n默认是ovr，它会对毎个标签训练一个二分类的分类器，即总共$K$个。训练代码在\n\n```python\n1230 fold_coefs_ = Parallel(n_jobs=self.n_jobs, verbose=self.verbose,\n1231 backend=backend)(\n1232 path_func(X, y, pos_class=class_, Cs=[self.C],\n1233 fit_intercept=self.fit_intercept, tol=self.tol,\n1234 verbose=self.verbose, solver=self.solver, copy=False,\n1235 multi_class=self.multi_class, max_iter=self.max_iter,\n1236 class_weight=self.class_weight, check_input=False,\n1237 random_state=self.random_state, coef=warm_start_coef_,\n1238 max_squared_sum=max_squared_sum,\n1239 sample_weight=sample_weight)\n1240 for (class_, warm_start_coef_) in zip(classes_, warm_start_coef))\n```\n\n注意，1240L的`for class_ in classes`配合1232L的`pos_class=class`，就是逐个取标签来训练的逻辑。", "_____no_output_____" ], [ "##### 2.2.1 multinominal\n\n前面讲到ovr会遍历标签，逐个训练。为了兼容这段逻辑，真正的二分类问题需要做变化：\n\n```python\n1201 if len(self.classes_) == 2:\n1202 n_classes = 1\n1203 classes_ = classes_[1:]\n```\n\n同样地，multinominal需要一次对全部标签做处理，也需要做变化：\n\n```python\n1217 # Hack so that we iterate only once for the multinomial case.\n1218 if self.multi_class == 'multinomial':\n1219 classes_ = [None]\n1220 warm_start_coef = [warm_start_coef]\n```\n\n好，接下来，我们看multinoinal的损失函数和导数计算代码，它是`_multinomial_loss_grad`这个函数。\n\nsklearn里多分类的代码使用的公式和[逻辑回归算法简介和Python实现](./demo.ipynb)里一致，即：\n\n\\begin{align}\n L(\\beta) &= \\log(\\sum_i e^{\\beta_{i0} + \\beta_i x)}) - (\\beta_{k0} + \\beta_k x) \\\\\n \\frac{\\partial L}{\\partial \\beta} &= x \\left ( \\frac{e^{\\beta_{k0} + \\beta_k x}}{\\sum_i e^{\\beta_{i0} + \\beta_i x}} - I(y = k) \\right ) \\\\\n\\end{align}\n\n具体到损失函数：\n\n```python\n244 def _multinomial_loss(w, X, Y, alpha, sample_weight):\n 245 #+-- 37 lines: \"\"\"Computes multinomial loss and class probabilities.---\n 282 n_classes = Y.shape[1]\n 283 n_features = X.shape[1]\n 284 fit_intercept = w.size == (n_classes * (n_features + 1))\n 285 w = w.reshape(n_classes, -1)\n 286 sample_weight = sample_weight[:, np.newaxis]\n 287 if fit_intercept:\n 288 intercept = w[:, -1]\n 289 w = w[:, :-1]\n 290 else:\n 291 intercept = 0\n 292 p = safe_sparse_dot(X, w.T)\n 293 p += intercept\n 294 p -= logsumexp(p, axis=1)[:, np.newaxis]\n 295 loss = -(sample_weight * Y * p).sum()\n 296 loss += 0.5 * alpha * squared_norm(w)\n 297 p = np.exp(p, p)\n 298 return loss, p, w\n```\n\n+ 292L-293L是计算$\\beta_{i0} + \\beta_i x$。\n+ 294L是计算 $L(\\beta)$。注意，这里防止计算溢出，是在`logsumexp`函数里作的，原理和[逻辑回归在spark中的实现简介](./spark_ml_lr.ipynb)一样。\n+ 295L是加总（注意，$Y$毎列是单位向量，所以起了选标签对应$k$的作用）。\n+ 296L加上L2正则。\n+ 注意，297L是p变回了$\\frac{e^{\\beta_{k0} + \\beta_k x}}{\\sum_i e^{\\beta_{i0} + \\beta_i x}}$，为了计算导数时直接用。\n\n好，再看导数的计算：\n\n```python\n 301 def _multinomial_loss_grad(w, X, Y, alpha, sample_weight):\n 302 #+-- 37 lines: \"\"\"Computes the multinomial loss, gradient and class probabilities.---\n 339 n_classes = Y.shape[1]\n 340 n_features = X.shape[1]\n 341 fit_intercept = (w.size == n_classes * (n_features + 1))\n 342 grad = np.zeros((n_classes, n_features + bool(fit_intercept)))\n 343 loss, p, w = _multinomial_loss(w, X, Y, alpha, sample_weight)\n 344 sample_weight = sample_weight[:, np.newaxis]\n 345 diff = sample_weight * (p - Y)\n 346 grad[:, :n_features] = safe_sparse_dot(diff.T, X)\n 347 grad[:, :n_features] += alpha * w\n 348 if fit_intercept:\n 349 grad[:, -1] = diff.sum(axis=0)\n 350 return loss, grad.ravel(), p\n```\n\n+ 345L-346L，对应了导数的计算式；\n+ 347L是加上L2的导数；\n+ 348L-349L，是对intercept的计算。", "_____no_output_____" ], [ "#### 2.3 Hessian\n\n注意，sklearn支持牛顿法，需要用到Hessian阵，定义见维基[Hessian matrix](https://en.wikipedia.org/wiki/Hessian_matrix)，\n\n\\begin{equation}\n{\\mathbf H}={\\begin{bmatrix}{\\dfrac {\\partial ^{2}f}{\\partial x_{1}^{2}}}&{\\dfrac {\\partial ^{2}f}{\\partial x_{1}\\,\\partial x_{2}}}&\\cdots &{\\dfrac {\\partial ^{2}f}{\\partial x_{1}\\,\\partial x_{n}}}\\\\[2.2ex]{\\dfrac {\\partial ^{2}f}{\\partial x_{2}\\,\\partial x_{1}}}&{\\dfrac {\\partial ^{2}f}{\\partial x_{2}^{2}}}&\\cdots &{\\dfrac {\\partial ^{2}f}{\\partial x_{2}\\,\\partial x_{n}}}\\\\[2.2ex]\\vdots &\\vdots &\\ddots &\\vdots \\\\[2.2ex]{\\dfrac {\\partial ^{2}f}{\\partial x_{n}\\,\\partial x_{1}}}&{\\dfrac {\\partial ^{2}f}{\\partial x_{n}\\,\\partial x_{2}}}&\\cdots &{\\dfrac {\\partial ^{2}f}{\\partial x_{n}^{2}}}\\end{bmatrix}}.\n\\end{equation}\n\n其实就是各点位的二阶偏导。具体推导就不写了，感兴趣可以看[Logistic Regression - Jia Li](http://sites.stat.psu.edu/~jiali/course/stat597e/notes2/logit.pdf)或[Logistic regression: a simple ANN Nando de Freitas](https://www.cs.ox.ac.uk/people/nando.defreitas/machinelearning/lecture6.pdf)。\n\n基本公式是$\\mathbf{H} = \\mathbf{X}^T \\operatorname{diag}(\\pi_i (1 - \\pi_i)) \\mathbf{X}$，其中$\\pi_i = \\operatorname{sigm}(x_i \\beta)$。\n\n```python\n 167 def _logistic_grad_hess(w, X, y, alpha, sample_weight=None):\n 168 #+-- 33 lines: \"\"\"Computes the gradient and the Hessian, in the case of a logistic loss.\n 201 w, c, yz = _intercept_dot(w, X, y)\n 202 #+-- 4 lines: if sample_weight is None:---------\n 206 z = expit(yz)\n 207 #+-- 8 lines: z0 = sample_weight * (z - 1) * y---\n 215 # The mat-vec product of the Hessian\n 216 d = sample_weight * z * (1 - z)\n 217 if sparse.issparse(X):\n 218 dX = safe_sparse_dot(sparse.dia_matrix((d, 0),\n 219 shape=(n_samples, n_samples)), X)\n 220 else:\n 221 # Precompute as much as possible\n 222 dX = d[:, np.newaxis] * X\n 223\n 224 if fit_intercept:\n 225 # Calculate the double derivative with respect to intercept\n 226 # In the case of sparse matrices this returns a matrix object.\n 227 dd_intercept = np.squeeze(np.array(dX.sum(axis=0)))\n 228\n 229 def Hs(s):\n 230 ret = np.empty_like(s)\n 231 ret[:n_features] = X.T.dot(dX.dot(s[:n_features]))\n 232 ret[:n_features] += alpha * s[:n_features]\n 233\n 234 # For the fit intercept case.\n 235 if fit_intercept:\n 236 ret[:n_features] += s[-1] * dd_intercept\n 237 ret[-1] = dd_intercept.dot(s[:n_features])\n 238 ret[-1] += d.sum() * s[-1]\n 239 return ret\n 240\n 241 return grad, Hs\n```\n\n+ 201L, 206L, 和216L是计算中间的$\\pi_i (1 - \\pi_i)$。\n+ 217L-222L，对中间参数变为对角阵后，预算公式后半部份，配合231L就是整个式子了。\n\n这里我也只知其然，以后有时间再深挖下吧。", "_____no_output_____" ], [ "### 3. 小结\n\n本文简单介绍了sklearn中逻辑回归的实现，包括二分类和多分类的具体代码和公式对应。", "_____no_output_____" ] ] ]

[ "code", "markdown", "code", "markdown" ]

[ [ "code" ], [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown" ] ]

4ad5c62472705b31d2d279b66ed0e874f2008fff

9,071

ipynb

Jupyter Notebook

docs/source/demos/fraud.ipynb

skvorekn/evalml

2cbfa344ec3fdc0fb0f4a0f1093811135b9b97d8

[ "BSD-3-Clause" ]

null

docs/source/demos/fraud.ipynb

skvorekn/evalml

2cbfa344ec3fdc0fb0f4a0f1093811135b9b97d8

[ "BSD-3-Clause" ]

null

docs/source/demos/fraud.ipynb

skvorekn/evalml

2cbfa344ec3fdc0fb0f4a0f1093811135b9b97d8

[ "BSD-3-Clause" ]

null

30.337793

413

0.598721

[ [ [ "# Building a Fraud Prediction Model with EvalML\n\nIn this demo, we will build an optimized fraud prediction model using EvalML. To optimize the pipeline, we will set up an objective function to minimize the percentage of total transaction value lost to fraud. At the end of this demo, we also show you how introducing the right objective during the training is over 4x better than using a generic machine learning metric like AUC.", "_____no_output_____" ] ], [ [ "import evalml\nfrom evalml import AutoMLSearch\nfrom evalml.objectives import FraudCost", "_____no_output_____" ] ], [ [ "## Configure \"Cost of Fraud\" \n\nTo optimize the pipelines toward the specific business needs of this model, we can set our own assumptions for the cost of fraud. These parameters are\n\n* `retry_percentage` - what percentage of customers will retry a transaction if it is declined?\n* `interchange_fee` - how much of each successful transaction do you collect?\n* `fraud_payout_percentage` - the percentage of fraud will you be unable to collect\n* `amount_col` - the column in the data the represents the transaction amount\n\nUsing these parameters, EvalML determines attempt to build a pipeline that will minimize the financial loss due to fraud.", "_____no_output_____" ] ], [ [ "fraud_objective = FraudCost(retry_percentage=.5,\n interchange_fee=.02,\n fraud_payout_percentage=.75,\n amount_col='amount')", "_____no_output_____" ] ], [ [ "## Search for best pipeline", "_____no_output_____" ], [ "In order to validate the results of the pipeline creation and optimization process, we will save some of our data as the holdout set.", "_____no_output_____" ] ], [ [ "X, y = evalml.demos.load_fraud(n_rows=1000)", "_____no_output_____" ] ], [ [ "EvalML natively supports one-hot encoding. Here we keep 1 out of the 6 categorical columns to decrease computation time.", "_____no_output_____" ] ], [ [ "cols_to_drop = ['datetime', 'expiration_date', 'country', 'region', 'provider']\nfor col in cols_to_drop:\n X.pop(col)\n\nX_train, X_holdout, y_train, y_holdout = evalml.preprocessing.split_data(X, y, problem_type='binary', test_size=0.2, random_seed=0)\n\nprint(X.types)", "_____no_output_____" ] ], [ [ "Because the fraud labels are binary, we will use `AutoMLSearch(X_train=X_train, y_train=y_train, problem_type='binary')`. When we call `.search()`, the search for the best pipeline will begin. ", "_____no_output_____" ] ], [ [ "automl = AutoMLSearch(X_train=X_train, y_train=y_train,\n problem_type='binary', \n objective=fraud_objective,\n additional_objectives=['auc', 'f1', 'precision'],\n max_batches=1,\n optimize_thresholds=True)\n\nautoml.search()", "_____no_output_____" ] ], [ [ "### View rankings and select pipelines\n\nOnce the fitting process is done, we can see all of the pipelines that were searched, ranked by their score on the fraud detection objective we defined.", "_____no_output_____" ] ], [ [ "automl.rankings", "_____no_output_____" ] ], [ [ "To select the best pipeline we can call `automl.best_pipeline`.", "_____no_output_____" ] ], [ [ "best_pipeline = automl.best_pipeline", "_____no_output_____" ] ], [ [ "### Describe pipelines\n\nWe can get more details about any pipeline created during the search process, including how it performed on other objective functions, by calling the `describe_pipeline` method and passing the `id` of the pipeline of interest.", "_____no_output_____" ] ], [ [ "automl.describe_pipeline(automl.rankings.iloc[1][\"id\"])", "_____no_output_____" ] ], [ [ "## Evaluate on holdout data\n\nFinally, since the best pipeline is already trained, we evaluate it on the holdout data.", "_____no_output_____" ], [ "Now, we can score the pipeline on the holdout data using both our fraud cost objective and the AUC (Area under the ROC Curve) objective.", "_____no_output_____" ] ], [ [ "best_pipeline.score(X_holdout, y_holdout, objectives=[\"auc\", fraud_objective])", "_____no_output_____" ] ], [ [ "## Why optimize for a problem-specific objective?\n\nTo demonstrate the importance of optimizing for the right objective, let's search for another pipeline using AUC, a common machine learning metric. After that, we will score the holdout data using the fraud cost objective to see how the best pipelines compare.", "_____no_output_____" ] ], [ [ "automl_auc = AutoMLSearch(X_train=X_train, y_train=y_train,\n problem_type='binary',\n objective='auc',\n additional_objectives=['f1', 'precision'],\n max_batches=1,\n optimize_thresholds=True)\n\nautoml_auc.search()", "_____no_output_____" ] ], [ [ "Like before, we can look at the rankings of all of the pipelines searched and pick the best pipeline.", "_____no_output_____" ] ], [ [ "automl_auc.rankings", "_____no_output_____" ], [ "best_pipeline_auc = automl_auc.best_pipeline", "_____no_output_____" ], [ "# get the fraud score on holdout data\nbest_pipeline_auc.score(X_holdout, y_holdout, objectives=[\"auc\", fraud_objective])", "_____no_output_____" ], [ "# fraud score on fraud optimized again\nbest_pipeline.score(X_holdout, y_holdout, objectives=[\"auc\", fraud_objective])", "_____no_output_____" ] ], [ [ "When we optimize for AUC, we can see that the AUC score from this pipeline is better than the AUC score from the pipeline optimized for fraud cost. However, the losses due to fraud are over 3% of the total transaction amount when optimized for AUC and under 1% when optimized for fraud cost. As a result, we lose more than 2% of the total transaction amount by not optimizing for fraud cost specifically.\n\nThis happens because optimizing for AUC does not take into account the user-specified `retry_percentage`, `interchange_fee`, `fraud_payout_percentage` values. Thus, the best pipelines may produce the highest AUC but may not actually reduce the amount loss due to your specific type fraud.\n\nThis example highlights how performance in the real world can diverge greatly from machine learning metrics.", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ] ]

4ad5cc381437434f418c46260b64d3c6c039a0f2

14,460

ipynb

Jupyter Notebook

end_to_end_tests/notebook_tests/basic_tests.ipynb

gbrault/jp_proxy_widget

2bd44001b0216f30e599cc33eb04912db2c49ae8

[ "BSD-2-Clause" ]

53

2017-11-03T13:04:45.000Z

2022-03-19T15:19:13.000Z

end_to_end_tests/notebook_tests/basic_tests.ipynb

gbrault/jp_proxy_widget

2bd44001b0216f30e599cc33eb04912db2c49ae8

[ "BSD-2-Clause" ]

22

2018-08-03T20:22:44.000Z

2022-03-21T11:37:59.000Z

end_to_end_tests/notebook_tests/basic_tests.ipynb

gbrault/jp_proxy_widget

2bd44001b0216f30e599cc33eb04912db2c49ae8

[ "BSD-2-Clause" ]

12

2018-08-24T14:06:24.000Z

2022-02-06T14:05:15.000Z

30.442105

133

0.559544

[ [ [ "# Basic functionality tests.\n\nIf the notebook cells complete with no exception the tests have passed.\n\nThe tests must be run in the full `jupyter notebook` or `jupyter lab` environment.\n\n*Note:* I couldn't figure out to make the validation tests run correctly \nat top level cell evaluation using `Run all`\nbecause the widgets initialize after later cells have executed, causing spurious\nfailures. Consequently the automated validation steps involve an extra round trip using\na widget at the bottom of the notebook which is guaranteed to render last.", "_____no_output_____" ] ], [ [ "# Some test artifacts used below:\n\nimport jp_proxy_widget\nfrom jp_proxy_widget import notebook_test_helpers\n\nvalidators = notebook_test_helpers.ValidationSuite()\n\nimport time\n\nclass PythonClass:\n \n class_attribute = \"initial class attribute value\"\n \n def __init__(self):\n self.set_instance_attribute(\"initial instance attribute value\")\n \n def set_instance_attribute(self, value):\n self.instance_attribute = value\n \n @classmethod\n def set_class_attribute(cls, value):\n cls.class_attribute = value\n\nnotebook_test_helpers\n\njp_proxy_widget\n", "_____no_output_____" ], [ "python_instance = PythonClass()\n\ndef python_function(value1, value2):\n python_instance.new_attribute = \"value1=%s and value2=%s\" % (value1, value2)", "_____no_output_____" ] ], [ [ "# pong: test that a proxy widget can call back to Python", "_____no_output_____" ] ], [ [ "import jp_proxy_widget", "_____no_output_____" ], [ "pong = jp_proxy_widget.JSProxyWidget()", "_____no_output_____" ], [ "def validate_pong():\n # check that the Python callbacks were called.\n assert python_instance.instance_attribute == \"instance\"\n assert PythonClass.class_attribute == \"class\"\n assert python_instance.new_attribute == 'value1=1 and value2=3'\n assert pong.error_msg == 'No error'\n print (\"pong says\", pong.error_msg)\n print (\"Pong callback test succeeded!\")\n\npong.js_init(\"\"\"\n//debugger;\ninstance_method(\"instance\");\nclass_method(\"class\");\npython_function(1, 3);\nelement.html(\"Callback test widget: nothing interesting to see here\")\n//validate()\n\"\"\", \n instance_method=python_instance.set_instance_attribute,\n class_method=PythonClass.set_class_attribute,\n python_function=python_function,\n #validate=validate_pong\n )\n\n#widget_validator_list.append([pong, validate_pong])\nvalidators.add_validation(pong, validate_pong)\n\n#pong.debugging_display()\npong", "_____no_output_____" ], [ "\n# set the mainloop check to True if running cells one at a time\nmainloop_check = False\n\nif mainloop_check:\n # At this time this fails on \"run all\"\n validate_pong()", "_____no_output_____" ] ], [ [ "# pingpong: test that Python can call in to a widget\n\n... use a widget callback to pass the value back", "_____no_output_____" ] ], [ [ "pingpong_list = \"just some strings\".split()\n\ndef pingpong_python_fn(argument1, argument2):\n print(\"called pingpong_python_fn\") # this print goes nowhere?\n pingpong_list[:] = [argument1, argument2]\n\ndef validate_pingpong():\n # check that the callback got the right values\n assert pingpong_list == [\"testing\", 123]\n print (\"ping pong test callback got \", pingpong_list)\n print (\"ping pong test succeeded!\")\n \npingpong = jp_proxy_widget.JSProxyWidget()\npingpong.js_init(\"\"\"\nelement.html(\"Ping pong test -- no call yet.\")\nelement.call_in_to_the_widget = function (argument1, argument2) {\n element.html(\" Call in sent \" + argument1 + \" and \" + argument2 + \"\")\n call_back_to_python(argument1, argument2);\n}\nelement.validate = validate;\n\"\"\", call_back_to_python=pingpong_python_fn, validate=validate_pingpong)\n\n#widget_validator_list.append([pingpong, validate_pingpong])\nvalidators.add_validation(pingpong, validate_pingpong)\n\n#pingpong.debugging_display()\npingpong", "_____no_output_____" ], [ "# call in to javascript\npingpong.element.call_in_to_the_widget(\"testing\", 123)\n# call in to javascript and back to python to validate\npingpong.element.validate()\n\nif mainloop_check:\n validate_pingpong()\n ", "_____no_output_____" ] ], [ [ "# roundtrip: datatype round trip\n\nTest that values can be passed in to the proxy widget and back out again.", "_____no_output_____" ] ], [ [ "binary = bytearray(b\"\\x12\\xff binary bytes\")\nstring_value = \"just a string\"\nint_value = -123\nfloat_value = 45.6\njson_dictionary = {\"keys\": None, \"must\": 321, \"be\": [6, 12], \"strings\": \"values\", \"can\": [\"be\", \"any json\"]}\nlist_value = [9, string_value, json_dictionary]\n\nroundtrip_got_values = []\n\nfrom jp_proxy_widget import hex_codec\n\nfrom pprint import pprint\n\ndef get_values_back(binary, string_value, int_value, float_value, json_dictionary, list_value):\n # NOTE: binary values must be converted explicitly from hex string encoding!\n binary = hex_codec.hex_to_bytearray(binary)\n roundtrip_got_values[:] = [binary, string_value, int_value, float_value, json_dictionary, list_value]\n print (\"GOT VALUES BACK\")\n pprint(roundtrip_got_values)\n\nroundtrip_names = \"binary string_value int_value float_value json_dictionary list_value\".split()\n\ndef validate_roundtrip():\n #assert roundtrip_got_values == [string_value, int_value, float_value, json_dictionary, list_value]\n expected_values = [binary, string_value, int_value, float_value, json_dictionary, list_value]\n if len(expected_values) != len(roundtrip_got_values):\n print (\"bad lengths\", len(expected_values), len(roundtrip_got_values))\n pprint(expected_values)\n pprint(roundtrip_got_values)\n assert len(expected_values) == len(roundtrip_got_values)\n for (name, got, expected) in zip(roundtrip_names, roundtrip_got_values, expected_values):\n if (got != expected):\n print(name, \"BAD MATCH got\")\n pprint(got)\n print(\" ... expected\")\n pprint(expected)\n assert got == expected, \"values don't match: \" + repr((name, got, expected))\n print (\"roundtrip values match!\")\n \nroundtrip = jp_proxy_widget.JSProxyWidget()\nroundtrip.js_init(r\"\"\"\nelement.all_values = [binary, string_value, int_value, float_value, json_dictionary, list_value];\nhtml = [\"<pre> Binary values sent as bytearrays appear in Javascript as Uint8Arrays\"]\nfor (var i=0; i<names.length; i++) {\n html.push(names[i]);\n var v = element.all_values[i];\n if (v instanceof Uint8Array) {\n html.push(\" Uint8Array\")\n } else {\n html.push(\" type: \" + (typeof v))\n }\n html.push(\" value: \" + v);\n}\nhtml.push(\"</pre>\");\nelement.html(html.join(\"\\n\"));\n\n// send the values back\ncallback(binary, string_value, int_value, float_value, json_dictionary, list_value);\n\"\"\",\n binary=binary,\n string_value=string_value,\n int_value=int_value,\n float_value=float_value,\n json_dictionary=json_dictionary,\n list_value=list_value,\n names=roundtrip_names,\n callback=get_values_back,\n # NOTE: must up the callable level!\n callable_level=4\n )\n\nroundtrip.debugging_display()", "_____no_output_____" ], [ "validators.add_validation(roundtrip, validate_roundtrip)\n\nif mainloop_check:\n validate_roundtrip()\n \n#validate_roundtrip()", "_____no_output_____" ] ], [ [ "# loadCSS -- test load of simple CSS file.\n\nWe want to load this css file", "_____no_output_____" ] ], [ [ "from jp_proxy_widget import js_context\nstyle_fn=\"js/simple.css\"\nprint(js_context.get_text_from_file_name(style_fn))", "_____no_output_____" ], [ "loadCSS = jp_proxy_widget.JSProxyWidget()\n\n# load the file\nloadCSS.load_css(style_fn)\n\n# callback for storing the styled element color\nloadCSSstyle = {}\n\ndef color_callback(color):\n loadCSSstyle[\"color\"] = color\n\n# initialize the element using the style and callback to report the color.\nloadCSS.js_init(\"\"\"\nelement.html('<div>Styled widget element.</div>')\n\nvar e = document.getElementById(\"loadCSSelement\");\nvar style = window.getComputedStyle(e);\ncolor_callback(style[\"color\"]);\n\"\"\", color_callback=color_callback)\n\ndef validate_loadCSS():\n expect = 'rgb(216, 50, 61)'\n assert expect == loadCSSstyle[\"color\"], repr((expect, loadCSSstyle))\n print (\"Loaded CSS color is correct!\")\n\nloadCSS", "_____no_output_____" ], [ "validators.add_validation(loadCSS, validate_loadCSS)\n\nif mainloop_check:\n validate_loadCSS()\n ", "_____no_output_____" ] ], [ [ "# loadJS -- load a javascript file (once only per interpreter)\n\nWe want to load this javascript file:", "_____no_output_____" ] ], [ [ "js_fn=\"js/simple.js\"\nprint(js_context.get_text_from_file_name(js_fn))", "_____no_output_____" ], [ "loadJS = jp_proxy_widget.JSProxyWidget()\n\n# load the file\nloadJS.load_js_files([js_fn], force=True)\n\n# callback for storing the styled element color\nloadJSinfo = {}\n\ndef answer_callback(answer):\n loadJSinfo[\"answer\"] = answer\n\nloadJS.js_init(\"\"\"\n element.html('The answer is ' + window.the_answer + '')\n answer_callback(window.the_answer);\n\"\"\", answer_callback=answer_callback, js_fn=js_fn)\n\ndef validate_loadJS():\n expect = 42\n assert expect == loadJSinfo[\"answer\"], repr((expect, loadJSinfo))\n print (\"Loaded JS value is correct!\")\n\nloadJS", "_____no_output_____" ], [ "validators.add_validation(loadJS, validate_loadJS)\n\nif mainloop_check:\n validate_loadJS()\n \nloadJS.print_status()", "_____no_output_____" ], [ "delay_ms = 1000\nvalidators.run_all_in_widget(delay_ms=delay_ms)", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code" ] ]

4ad5ced8dee57c04454bd4eaf83b3a2f08fe543e

5,055

ipynb

Jupyter Notebook

Midterm_Exam_PYTHON_FUNDAMENTALS.ipynb

fernandoBjr/CPEN-21A-ECE-2-2

fda37ab438d43b390d8b1e5832bbbed208183ef5

[ "Apache-2.0" ]

null

Midterm_Exam_PYTHON_FUNDAMENTALS.ipynb

fernandoBjr/CPEN-21A-ECE-2-2

fda37ab438d43b390d8b1e5832bbbed208183ef5

[ "Apache-2.0" ]

null

Midterm_Exam_PYTHON_FUNDAMENTALS.ipynb

fernandoBjr/CPEN-21A-ECE-2-2

fda37ab438d43b390d8b1e5832bbbed208183ef5

[ "Apache-2.0" ]

null

25.149254

257

0.410287

[ [ [ "<a href=\"https://colab.research.google.com/github/fernandoBjr/CPEN-21A-ECE-2-2/blob/main/Midterm_Exam_PYTHON_FUNDAMENTALS.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>", "_____no_output_____" ], [ "#**PROBLEM STATEMENT 1**", "_____no_output_____" ] ], [ [ "name = \"Name: Buño, Fernando Jr, L.\"\nstudentnumber = \"Student Number: 202013292\"\nage = \"Age: 11\"\nbirthday = \"Birthday: April 11, 2001\"\naddress = \"Addres: B 8 L 1 Brazil Street Langkaan II Dasma Cavite\"\ncourse = \"Course: Electrical Communication Engineering\"\ngwa = \"GWA: 1.75\"\n\nprint(name)\nprint(studentnumber)\nprint(age)\nprint(birthday)\nprint(address)\nprint(course)\nprint(gwa)\n", "Name: Buño, Fernando Jr, L.\nStudent Number: 202013292\nAge: 11\nBirthday: April 11, 2001\nAddres: B 8 L 1 Brazil Street Langkaan II Dasma Cavite\nCourse: Electrical Communication Engineering\nGWA: 1.75\n" ] ], [ [ "#**PROBLEM STATEMENT 2**", "_____no_output_____" ] ], [ [ "n = 4\nansw = \"Y\"\n\nprint(2<n) and (n<6)\nprint(2<n) or (n==6)\nprint(not(2<n)) or (n==6)\nprint(not(n<6))\nprint(answ==\"Y\") or (answ==\"y\")\nprint(answ==\"Y\") and (answ==\"y\")\nprint(not(answ==\"y\"))\nprint((2<n)and(n==5+1)) or (answ==\"No\")\nprint(n==2)and(n==7) or (answ==\"Y\")\nprint(n==2) and (n==7) or(answ==\"Y\")", "True\nTrue\nFalse\nFalse\nTrue\nTrue\nTrue\nFalse\nFalse\nFalse\n" ] ], [ [ "##**PROBLEM STATEMENT 3**", "_____no_output_____" ] ], [ [ "x = 2\ny = -3\nw = 7\nz = -10\n\nprint(x/y)\nprint(w/y/x)\nprint(z/y%x)\nprint(x%-y*w)\nprint(x%y)\nprint(z%w-y/x*5+5)\nprint(9-x%(2+y))\nprint(z//w)\nprint((2+y)**2)\nprint(w/x*2)", "-0.6666666666666666\n-1.1666666666666667\n1.3333333333333335\n14\n-1\n16.5\n9\n-2\n1\n7.0\n" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad5d03c55c4e38e6f99494525f13ee81b84f3c0

10,892

ipynb

Jupyter Notebook

examples/notebooks/jiant_MNLI_Diagnostic_Example.ipynb

a7b23/jiant

b9e59fab42d71463fd23950d16abd8d828acaab9

[ "MIT" ]

1

2021-03-11T21:30:52.000Z

examples/notebooks/jiant_MNLI_Diagnostic_Example.ipynb

a7b23/jiant

b9e59fab42d71463fd23950d16abd8d828acaab9

[ "MIT" ]

null

examples/notebooks/jiant_MNLI_Diagnostic_Example.ipynb

a7b23/jiant

b9e59fab42d71463fd23950d16abd8d828acaab9

[ "MIT" ]

null

24.810934

114

0.539111

[ [ [ "# MNLI Diagnostic Example", "_____no_output_____" ], [ "## Setup", "_____no_output_____" ], [ "#### Install dependencies", "_____no_output_____" ] ], [ [ "%%capture\n!git clone https://github.com/jiant-dev/jiant.git", "_____no_output_____" ], [ "%%capture\n# This Colab notebook already has its CUDA-runtime compatible versions of torch and torchvision installed\n!sed -e /\"torch==1.5.0\"/d -i jiant/requirements.txt\n!sed -e /\"torchvision==0.6.0\"/d -i jiant/requirements.txt\n!pip install -r jiant/requirements.txt", "_____no_output_____" ] ], [ [ "#### Download data", "_____no_output_____" ] ], [ [ "%%capture\n# Download/preprocess MNLI and RTE data\n!wget https://raw.githubusercontent.com/huggingface/transformers/master/utils/download_glue_data.py\n!python download_glue_data.py \\\n --data_dir ./raw_data \\\n --tasks \"MNLI,diagnostic\"\n!PYTHONPATH=/content/jiant python jiant/jiant/scripts/preproc/export_glue_data.py \\\n --input_base_path=./raw_data \\\n --output_base_path=./tasks/ \\\n --task_name_ls \"mnli,glue_diagnostic\"", "_____no_output_____" ] ], [ [ "## `jiant` Pipeline", "_____no_output_____" ] ], [ [ "import sys\nsys.path.insert(0, \"/content/jiant\")", "_____no_output_____" ], [ "import jiant.proj.main.tokenize_and_cache as tokenize_and_cache\nimport jiant.proj.main.export_model as export_model\nimport jiant.proj.main.scripts.configurator as configurator\nimport jiant.proj.main.runscript as main_runscript\nimport jiant.shared.caching as caching\nimport jiant.utils.python.io as py_io\nimport jiant.utils.display as display\nimport os\nimport torch", "_____no_output_____" ] ], [ [ "#### Task config", "_____no_output_____" ] ], [ [ "# Write MNLI task config\npy_io.write_json({\n \"task\": \"mnli\",\n \"name\": \"mnli\",\n \"paths\": {\n \"train\": \"/content/tasks/data/mnli/train.jsonl\",\n \"val\": \"/content/tasks/data/mnli/val.jsonl\",\n },\n}, path=\"./tasks/configs/mnli_config.json\")\n\n# Write MNLI-mismatched task config\npy_io.write_json({\n \"task\": \"mnli\",\n \"name\": \"mnli_mismatched\",\n \"paths\": {\n \"val\": \"/content/tasks/data/mnli/val_mismatched.jsonl\",\n },\n}, path=\"./tasks/configs/mnli_mismatched_config.json\")\n\n# Write GLUE diagnostic task config\npy_io.write_json({\n \"task\": \"glue_diagnostics\",\n \"name\": \"glue_diagnostics\",\n \"paths\": {\n \"test\": \"/content/tasks/data/glue_diagnostics/test.jsonl\",\n },\n}, path=\"./tasks/configs/glue_diagnostics_config.json\")", "_____no_output_____" ] ], [ [ "#### Download model", "_____no_output_____" ] ], [ [ "export_model.lookup_and_export_model(\n model_type=\"roberta-base\",\n output_base_path=\"./models/roberta-base\",\n)", "_____no_output_____" ] ], [ [ "#### Tokenize and cache\n", "_____no_output_____" ] ], [ [ "# Tokenize and cache each task\ntokenize_and_cache.main(tokenize_and_cache.RunConfiguration(\n task_config_path=f\"./tasks/configs/mnli_config.json\",\n model_type=\"roberta-base\",\n model_tokenizer_path=\"./models/roberta-base/tokenizer\",\n output_dir=f\"./cache/mnli\",\n phases=[\"train\", \"val\"],\n))\n\ntokenize_and_cache.main(tokenize_and_cache.RunConfiguration(\n task_config_path=f\"./tasks/configs/mnli_mismatched_config.json\",\n model_type=\"roberta-base\",\n model_tokenizer_path=\"./models/roberta-base/tokenizer\",\n output_dir=f\"./cache/mnli_mismatched\",\n phases=[\"val\"],\n))\n\ntokenize_and_cache.main(tokenize_and_cache.RunConfiguration(\n task_config_path=f\"./tasks/configs/glue_diagnostics_config.json\",\n model_type=\"roberta-base\",\n model_tokenizer_path=\"./models/roberta-base/tokenizer\",\n output_dir=f\"./cache/glue_diagnostics\",\n phases=[\"test\"],\n))", "_____no_output_____" ], [ "row = caching.ChunkedFilesDataCache(\"./cache/mnli/train\").load_chunk(0)[0][\"data_row\"]\nprint(row.input_ids)\nprint(row.tokens)", "_____no_output_____" ], [ "row = caching.ChunkedFilesDataCache(\"./cache/mnli_mismatched/val\").load_chunk(0)[0][\"data_row\"]\nprint(row.input_ids)\nprint(row.tokens)", "_____no_output_____" ], [ "row = caching.ChunkedFilesDataCache(\"./cache/glue_diagnostics/test\").load_chunk(0)[0][\"data_row\"]\nprint(row.input_ids)\nprint(row.tokens)", "_____no_output_____" ] ], [ [ "#### Writing a run config", "_____no_output_____" ] ], [ [ "jiant_run_config = configurator.SimpleAPIMultiTaskConfigurator(\n task_config_base_path=\"./tasks/configs\",\n task_cache_base_path=\"./cache\",\n train_task_name_list=[\"mnli\"],\n val_task_name_list=[\"mnli\", \"mnli_mismatched\"],\n test_task_name_list=[\"glue_diagnostics\"],\n train_batch_size=8,\n eval_batch_size=16,\n epochs=0.1,\n num_gpus=1,\n).create_config()\ndisplay.show_json(jiant_run_config)", "_____no_output_____" ] ], [ [ "Configure all three tasks to use an `mnli` head.", "_____no_output_____" ] ], [ [ "jiant_run_config[\"taskmodels_config\"][\"task_to_taskmodel_map\"] = {\n \"mnli\": \"mnli\",\n \"mnli_mismatched\": \"mnli\",\n \"glue_diagnostics\": \"glue_diagnostics\",\n}\nos.makedirs(\"./run_configs/\", exist_ok=True)\npy_io.write_json(jiant_run_config, \"./run_configs/jiant_run_config.json\")", "_____no_output_____" ] ], [ [ "#### Start training", "_____no_output_____" ] ], [ [ "run_args = main_runscript.RunConfiguration(\n jiant_task_container_config_path=\"./run_configs/jiant_run_config.json\",\n output_dir=\"./runs/run1\",\n model_type=\"roberta-base\",\n model_path=\"./models/roberta-base/model/roberta-base.p\",\n model_config_path=\"./models/roberta-base/model/roberta-base.json\",\n model_tokenizer_path=\"./models/roberta-base/tokenizer\",\n learning_rate=1e-5,\n eval_every_steps=500,\n do_train=True,\n do_val=True,\n do_save=True,\n write_test_preds=True,\n force_overwrite=True,\n)\nmain_runscript.run_loop(run_args)", "_____no_output_____" ], [ "test_preds = torch.load(\"./runs/run1/test_preds.p\")\ntest_preds[\"glue_diagnostics\"]", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown", "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ] ]

4ad5e604178a3589a995590913f7ec2e270ba3bf

5,188

ipynb

Jupyter Notebook

04-2017-11-10/01_html_liste.ipynb

xldrkp/jupyter-notebooks

28b556d398547b0faec311a202d27ce769cf44da

[ "MIT" ]

1

2020-08-06T18:31:30.000Z

04-2017-11-10/01_html_liste.ipynb

xldrkp/jupyter-notebooks

28b556d398547b0faec311a202d27ce769cf44da

[ "MIT" ]

null

04-2017-11-10/01_html_liste.ipynb

xldrkp/jupyter-notebooks

28b556d398547b0faec311a202d27ce769cf44da

[ "MIT" ]

null

22.362069

238

0.538358

[ [ [ "# Listen in Python: Ungeordnete HTML-Liste generieren\n\nIn dieser Einheit werden wir uns mit Listen in Python auseinandersetzen. Dazu definieren wir eine Liste mit diversen Tieren.", "_____no_output_____" ] ], [ [ "tiere = [\"Hund\", \"Katze\", \"Giraffe\", \"Maus\", \"Frosch\"]", "_____no_output_____" ] ], [ [ "Wir können den Inhalt der Variablen `tiere` einfach wieder ausgeben.", "_____no_output_____" ] ], [ [ "print(tiere)", "['Hund', 'Katze', 'Giraffe', 'Maus', 'Frosch']\n" ] ], [ [ "Um nun die Tiere einzeln wieder ausgeben zu können, iterieren wir über die Liste mit einer Schleife.", "_____no_output_____" ] ], [ [ "for tier in tiere:\n print(tier)", "Hund\nKatze\nGiraffe\nMaus\nFrosch\n" ] ], [ [ "In HTML ist eine Liste z.B. ein `<ul><li></li></ul>`-Konstrukt. Dieses können wir mit Python herstellen.", "_____no_output_____" ] ], [ [ "html_liste = '<ul>'\n\nfor tier in tiere:\n html_liste = html_liste + '<li>{}</li>'.format(tier)\n\nhtml_liste += '</ul>'", "_____no_output_____" ] ], [ [ "**Erklärung:** In Z1 speichern wir den ersten Teil der HTML-Liste in der Variablen `liste`. Ab Z3 beginnt unsere Schleife über die Inhalte der `tiere` zu iterieren. \n\nIn Z4 hängen wir mit dem Operator `+=` an den Inhalt der Variablen `html_liste` eine weitere Zeichenkette dran. Dabei sind die `{}` ein Platzhalter, der mit `.format(tier)` durch das aktuelle Tier aus der Iteration besetzt wird. \n\nNachdem die Schleife durchgelaufen ist, wird an den String noch das Endtag der ungeordneten Liste, `</ul>`, drangehängt.", "_____no_output_____" ], [ "Die Ausgabe von `liste` zeigt uns einen String.", "_____no_output_____" ] ], [ [ "print(html_liste)", "<ul><li>Hund</li><li>Katze</li><li>Giraffe</li><li>Maus</li><li>Frosch</li></ul>\n" ] ], [ [ "## Liste als HTML ausgeben\n\nUm das generierte Konstrukt nun als HTML ausgeben zu können (in diesem Notebook), importieren wir noch eine besondere Funktion.", "_____no_output_____" ] ], [ [ "# Nur notwendig, um hier, in diesem Notebook, \n# das HTML ausgeben zu können.\nfrom IPython.display import HTML\nHTML(html_liste)", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad5fc1d1a139c8612cb838f1853838fc64a8c56

1,416

ipynb

Jupyter Notebook

lessons/3.ipynb

BobNobrain/matstat-labs

bf7c95b29a6dc8b65675a780871400fd729f9e35

[ "MIT" ]

null

lessons/3.ipynb

BobNobrain/matstat-labs

bf7c95b29a6dc8b65675a780871400fd729f9e35

[ "MIT" ]

null

lessons/3.ipynb

BobNobrain/matstat-labs

bf7c95b29a6dc8b65675a780871400fd729f9e35

[ "MIT" ]

7

2018-11-18T06:31:49.000Z

2020-11-09T18:36:40.000Z

21.454545

89

0.481638

[ [ [ "# 3 семинар. Проверка гипотез о параметрах нормально распределённой совокупности\n\n```\nАуд.: 212(199)-220(207), 224(211), 234(221), 241(228)\nД/З: 221(208), 223(210), 228(215), 229(216), 238(225), 249(236)\n```\n\n$$α = P[Z \\in V_k | H_0]$$\n$$β = P[Z \\in V \\\\ V_k | H_1]$$\n$$D(X_{mean} | (x_1..x_n) \\leftarrow X \\sim N(m, σ)) = σ^2 / n$$\n", "_____no_output_____" ] ], [ [ "from scipy import stats \n\nq = stats.norm(0, 1).ppf(0.9)\n# [stats.norm(0, 1).ppf(1 - 0.9), stats.norm(0, 1).ppf(0.9)]\n(q / 0.2) ** 2", "_____no_output_____" ] ] ]

[ "markdown", "code" ]

[ [ "markdown" ], [ "code" ] ]

4ad5fdb8e3c7fa8da187e8347df4d13da25d2c5a

30,538

ipynb

Jupyter Notebook

Network.ipynb

DavidOWade/MNIST-Neural-Network

d5f246c0d204c869baf92e931e0d77a7b9244d83

[ "MIT" ]

null

Network.ipynb

DavidOWade/MNIST-Neural-Network

d5f246c0d204c869baf92e931e0d77a7b9244d83

[ "MIT" ]

null

Network.ipynb

DavidOWade/MNIST-Neural-Network

d5f246c0d204c869baf92e931e0d77a7b9244d83

[ "MIT" ]

null

100.12459

17,648

0.795828

[ [ [ "# MNIST Digit Classification Neural Network\n---\n\nA neural network is a system of interconnected nodes, or artificial neurons that perform some task by learning from a dataset and incrementally improving its own performance. These artificial neurons are organised into multiple layers including an input layer, where data is fed forward through the network's successive layers, until it produces some output in the final layer.\n\nNetworks \"learn\" by analyzing a dataset of training inputs, where each training example is classified by a label. Through a process called backpropagation, the network adjusts the \"weights\" connecting each neuron (which can be thought of as the synapses connecting neurons in a human brain) based on how close the output produced from traning examples, which classifies each training example, is to the actual classification of those examples. Biases for each neuron are also updated accordingly.\n\n### The MNIST Dataset\n\nThis project produces a neural network that classifies images of handwritten digits ranged from 0-9. These images are gathered from the MNIST database - a large set of images of handwritten digits commonly used for training neural networks like this one. This is my first attempt at building a neural network from scratch and I plan to continually update this project as I improve my code.\n\nEach image is input as a 784-dimensional vector, with each vector component representing the greyscale value of a pixel in the image. The network has one hidden layer composed of 25 neurons and a final output layer of 10 neurons. Output in the network can be viewed as the \"activation\" of these output neurons, or the degree to which a neuron is affected by the input of the system. For example, with an input representing the digit 0, the output neuron of index 0 (so, the first neuron) would have a higher value (or activation) associated with it, while other neurons would have comparably lower activations. \n\nHere are some other important features about my network:\n- It uses the sigmoid activation function\n- The number of epochs (a mini-batch of 100 training examples) and learning rates can be cusomised. These values are set to 800 and 1 by default.\n- Currently, my network has an average training accuracy of 85%.\n\n---\n\nThe following code implements my neural network", "_____no_output_____" ] ], [ [ "import numpy as np\nimport math\n\n# Sigmoid activation function returns a value between 0 and 1\n# based on the degree to which the input varies from 0\ndef sigmoid(x):\n if x.size == 1:\n return 1 / (1 + math.exp(-x))\n else:\n return np.array([(1 / (1 + math.exp(-i))) for i in x])\n\ndef sigmoid_derivative(x):\n if x.size == 1:\n return math.exp(-x) / ((1 + math.exp(-x))**2)\n else:\n return np.array([((math.exp(-i))/(1 + math.exp(-i))**2) for i in x])\n\nclass NNetwork:\n # The network is initialised with the training and testing sets as input\n def __init__(self, X_train, Y_train, X_test, Y_test):\n self.X_train = X_train\n self.Y_train = Y_train\n self.X_test = X_test\n self.Y_test = Y_test\n \n self.input = np.zeros(784)\n self.output = np.zeros(10)\n self.y = np.zeros(10)\n \n # Weights and biases are initialised as random values between -1 and 1\n self.weights2 = np.random.uniform(low=-1.0, high=1.0, size=(25,784))\n self.weights3 = np.random.uniform(low=-1.0, high=1.0, size=(10,25))\n self.bias2 = np.random.uniform(low=-1.0, high=1.0, size=25)\n self.bias3 = np.random.uniform(low=-1.0, high=1.0, size=10)\n \n def train(self, epochs, lr):\n for i in range(epochs):\n d_weights2 = np.zeros(self.weights2.shape)\n d_weights3 = np.zeros(self.weights3.shape)\n d_bias2 = np.zeros(self.bias2.shape)\n d_bias3 = np.zeros(self.bias3.shape)\n for j in range(100):\n self.input = self.X_train[(i * 100) + j,:]\n self.y[self.Y_train[(i * 100) + j]] = 1\n self.feedforward()\n updates = self.backprop() # The gradient of the cost function\n d_weights2 += updates[0]\n d_weights3 += updates[1]\n d_bias2 += updates[2]\n d_bias3 += updates[3]\n self.y = np.zeros(10)\n d_weights2 /= 100\n d_weights3 /= 100\n d_bias2 /= 100\n d_bias3 /= 100\n \n # The average negative value of the change in the cost with respect to the change \n # in each weight & bias for 100 training examples is calculated and added to the\n # current value of each weight and bias\n self.weights2 += -1 * lr * d_weights2\n self.weights3 += -1 * lr * d_weights3\n self.bias2 += -1 * lr * d_bias2\n self.bias3 += -1 * lr * d_bias3\n print(\"Training complete!\")\n \n # This function classifies a single image\n def classify(self, x):\n self.input = x\n self.feedforward()\n return np.argmax(self.output)\n \n def test(self):\n acc = 0\n for i in range(10000):\n x = X_test[i,:]\n y = Y_test[i]\n yHAT = self.classify(x)\n if y == yHAT:\n acc += 1\n print(\"Testing accuracy: \" + str((acc / 10000) * 100) + \"%\")\n \n # This function uses the sigmoid activation function to \n # feed an input forward, producing the values of the neurons\n # in the second layer and the final layer\n def feedforward(self):\n self.layer2 = sigmoid(np.dot(self.input, self.weights2.T) + self.bias2)\n self.output = sigmoid(np.dot(self.layer2, self.weights3.T) + self.bias3)\n \n # This function calculates the gradient of the cost function, where each\n # component of the cost gradient is associated with a single weight or bias\n def backprop(self):\n d_weights2 = np.zeros(self.weights2.shape)\n d_weights3 = np.zeros(self.weights3.shape)\n \n d_bias2 = np.zeros(self.bias2.shape)\n d_bias3 = np.zeros(self.bias3.shape)\n\n d_weights2 = self.input * (sigmoid_derivative(np.dot(self.input, self.weights2.T) + self.bias2)[:, np.newaxis] * np.sum((self.weights3.T * (sigmoid_derivative(np.dot(self.layer2, self.weights3.T) + self.bias3)) * 2 * (self.output - self.y)), axis=1)[:, np.newaxis])\n d_weights3 = np.tile(self.layer2,(10,1)) * sigmoid_derivative(np.dot(self.layer2, self.weights3.T) + self.bias3)[:, np.newaxis] * (2 * (self.output - self.y))[:, np.newaxis]\n \n d_bias2 = sigmoid_derivative(np.dot(self.input, self.weights2.T) + self.bias2) * (d_bias2 + np.sum((self.weights3.T * (sigmoid_derivative(np.dot(self.layer2, self.weights3.T) + self.bias3)) * 2 * (self.output - self.y)), axis=1))\n d_bias3 = sigmoid_derivative(np.dot(self.layer2, self.weights3.T) + self.bias3) * (d_bias3 + 2 * (self.output - self.y))\n \n return d_weights2, d_weights3, d_bias2, d_bias3\n \n \n ", "_____no_output_____" ] ], [ [ "The following code downloads the mnist dataset and converts it to input for the network. This code is based on hsjeong5's github project [MNIST-for-Numpy](https://github.com/hsjeong5/MNIST-for-Numpy).", "_____no_output_____" ] ], [ [ "import mnist\n\nmnist.init()\nX_train, Y_train, X_test, Y_test = mnist.load()\nX_train = X_train / 255\nX_test = X_test / 255", "Downloading train-images-idx3-ubyte.gz...\nDownloading t10k-images-idx3-ubyte.gz...\nDownloading train-labels-idx1-ubyte.gz...\nDownloading t10k-labels-idx1-ubyte.gz...\nDownload complete.\nSave complete.\n" ] ], [ [ "The following code uses the above input data to train & test the accuracy of a neural network", "_____no_output_____" ] ], [ [ "network = NNetwork(X_train, Y_train, X_test, Y_test)\nnetwork.train(600, 1)\nnetwork.test()", "Training complete!\nTesting accuracy: 81.8%\n" ] ], [ [ "Run the code below to test my network on three random images", "_____no_output_____" ] ], [ [ "import matplotlib.pyplot as plt\n\nimgIndex = np.random.randint(low=0, high=10000, size=4)\n\nfig = plt.figure(figsize=(8,6))\nfig.subplots_adjust(wspace=0.3, hspace=0.3)\n\nax1 = fig.add_subplot(221)\nax1.set_title(\"Testing image index \" + str(imgIndex[0]))\nplt.imshow(X_test[imgIndex[0]].reshape(28, 28), cmap='gray')\n\nax2 = fig.add_subplot(222)\nax2.set_title(\"Testing image index \" + str(imgIndex[1]))\nplt.imshow(X_test[imgIndex[1]].reshape(28, 28), cmap='gray')\n\nax3 = fig.add_subplot(223)\nax3.set_title(\"Testing image index \" + str(imgIndex[2]))\nplt.imshow(X_test[imgIndex[2]].reshape(28, 28), cmap='gray')\n\nax4 = fig.add_subplot(224)\nax4.set_title(\"Testing image index \" + str(imgIndex[3]))\nplt.imshow(X_test[imgIndex[3]].reshape(28, 28), cmap='gray')\n\nprint(\"Image \" + str(imgIndex[0]) + \" value: \" + str(Y_test[imgIndex[0]]) + \" Classified by network as: \" + str(network.classify(X_test[imgIndex[0],:])))\nprint(\"Image \" + str(imgIndex[1]) + \" value: \" + str(Y_test[imgIndex[1]]) + \" Classified by network as: \" + str(network.classify(X_test[imgIndex[1],:])))\nprint(\"Image \" + str(imgIndex[2]) + \" value: \" + str(Y_test[imgIndex[2]]) + \" Classified by network as: \" + str(network.classify(X_test[imgIndex[2],:])))\nprint(\"Image \" + str(imgIndex[3]) + \" value: \" + str(Y_test[imgIndex[3]]) + \" Classified by network as: \" + str(network.classify(X_test[imgIndex[3],:])))\nprint()\n\nplt.show()", "Image 5533 value: 7 Classified by network as: 7\nImage 4278 value: 0 Classified by network as: 0\nImage 5643 value: 2 Classified by network as: 2\nImage 9791 value: 0 Classified by network as: 0\n\n" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad6001986db72cd79b55080be7045224f00ba5e

73,050

ipynb

Jupyter Notebook

NewYorkCityTaxiFarePrediction/NewYorkCityTaxiFarePrediction-TF-DNN-V2.ipynb

subhachandrachandra/Kaggle_Notebooks

79f2d1aa806b57f489e7e7d3e37242f82ce2b1a4

[ "MIT" ]

null

NewYorkCityTaxiFarePrediction/NewYorkCityTaxiFarePrediction-TF-DNN-V2.ipynb

subhachandrachandra/Kaggle_Notebooks

79f2d1aa806b57f489e7e7d3e37242f82ce2b1a4

[ "MIT" ]

null

NewYorkCityTaxiFarePrediction/NewYorkCityTaxiFarePrediction-TF-DNN-V2.ipynb

subhachandrachandra/Kaggle_Notebooks

79f2d1aa806b57f489e7e7d3e37242f82ce2b1a4

[ "MIT" ]

null

33.356164

652

0.508186

[ [ [ "import numpy as np\nimport pandas as pd\nfrom sklearn import metrics\nfrom sklearn import model_selection\nfrom sklearn import preprocessing\nimport dask.dataframe as ddf\nimport dask\n\nimport tensorflow as tf", "_____no_output_____" ] ], [ [ "### Load NYC Taxi fare prepped data", "_____no_output_____" ] ], [ [ "train_df = (pd\n .read_parquet(\"../../datasets/kaggle/new-york-city-taxi-fare-prediction/train_full_2\")\n .sample(frac=0.80, random_state=42)\n )", "_____no_output_____" ], [ "print(train_df.shape)\nprint(train_df.columns)\nprint(train_df.info())", "(43397466, 46)\nIndex(['fare_amount', 'pickup_datetime', 'pickup_longitude', 'pickup_latitude',\n 'dropoff_longitude', 'dropoff_latitude', 'passenger_count',\n 'distance_miles', 'is_pickup_JFK_new', 'is_dropoff_JFK_new',\n 'is_pickup_EWR_new', 'is_dropoff_EWR_new', 'is_pickup_LGA_new',\n 'is_dropoff_LGA_new', 'is_to_from_JFK_new', 'distance_to_center',\n 'hour', 'weekday', 'year_2009', 'year_2010', 'year_2011', 'year_2012',\n 'year_2013', 'year_2014', 'year_2015', 'hour_period_4', 'hour_period_0',\n 'hour_period_1', 'hour_period_2', 'hour_period_5', 'hour_period_3',\n 'distance_to_center_dropoff', 'pickup_zone_2', 'pickup_zone_0',\n 'pickup_zone_1', 'pickup_zone_3', 'pickup_zone_4', 'pickup_zone_5',\n 'pickup_zone_6', 'dropoff_zone_2', 'dropoff_zone_0', 'dropoff_zone_1',\n 'dropoff_zone_3', 'dropoff_zone_4', 'dropoff_zone_5', 'dropoff_zone_6'],\n dtype='object')\n<class 'pandas.core.frame.DataFrame'>\nInt64Index: 43397466 entries, 1598883 to 12212406\nData columns (total 46 columns):\nfare_amount float32\npickup_datetime datetime64[ns]\npickup_longitude float32\npickup_latitude float32\ndropoff_longitude float32\ndropoff_latitude float32\npassenger_count uint8\ndistance_miles float32\nis_pickup_JFK_new bool\nis_dropoff_JFK_new bool\nis_pickup_EWR_new bool\nis_dropoff_EWR_new bool\nis_pickup_LGA_new bool\nis_dropoff_LGA_new bool\nis_to_from_JFK_new bool\ndistance_to_center float32\nhour int64\nweekday int64\nyear_2009 uint8\nyear_2010 uint8\nyear_2011 uint8\nyear_2012 uint8\nyear_2013 uint8\nyear_2014 uint8\nyear_2015 uint8\nhour_period_4 uint8\nhour_period_0 uint8\nhour_period_1 uint8\nhour_period_2 uint8\nhour_period_5 uint8\nhour_period_3 uint8\ndistance_to_center_dropoff float32\npickup_zone_2 uint8\npickup_zone_0 uint8\npickup_zone_1 uint8\npickup_zone_3 uint8\npickup_zone_4 uint8\npickup_zone_5 uint8\npickup_zone_6 uint8\ndropoff_zone_2 uint8\ndropoff_zone_0 uint8\ndropoff_zone_1 uint8\ndropoff_zone_3 uint8\ndropoff_zone_4 uint8\ndropoff_zone_5 uint8\ndropoff_zone_6 uint8\ndtypes: bool(7), datetime64[ns](1), float32(8), int64(2), uint8(28)\nmemory usage: 4.0 GB\nNone\n" ], [ "cols = [\n \"distance_miles\",\n \"passenger_count\",\n \"is_pickup_JFK_new\",\n \"is_dropoff_JFK_new\",\n #\"is_pickup_EWR_new\",\n \"is_dropoff_EWR_new\",\n #\"is_pickup_LGA_new\",\n #\"is_dropoff_LGA_new\",\n #\"is_to_from_JFK_new\",\n \"distance_to_center\",\n \"distance_to_center_dropoff\",\n \"year_2009\",\n \"year_2010\",\n \"year_2011\",\n \"year_2012\",\n \"year_2013\",\n \"year_2014\",\n \"year_2015\",\n #\"hour_period_0\",\n #\"hour_period_1\",\n #\"hour_period_2\",\n #\"hour_period_3\",\n \"hour_period_4\",\n #\"hour_period_5\",\n \"pickup_zone_0\",\n \"pickup_zone_1\",\n \"pickup_zone_2\",\n \"pickup_zone_3\",\n \"pickup_zone_4\",\n \"pickup_zone_5\",\n \"pickup_zone_6\",\n \"dropoff_zone_0\",\n \"dropoff_zone_1\",\n \"dropoff_zone_2\",\n \"dropoff_zone_3\",\n \"dropoff_zone_4\",\n \"dropoff_zone_5\",\n \"dropoff_zone_6\",\n]\nx_train = train_df[cols].values\ny_train = train_df[['fare_amount']].values", "_____no_output_____" ], [ "import gc\ndel train_df\ngc.collect()", "_____no_output_____" ], [ "#scaler = preprocessing.StandardScaler()\nscaler = preprocessing.MinMaxScaler()\nx_train_norm = scaler.fit_transform(x_train)", "/home/subhachandra/anaconda3/envs/tensorflow/lib/python3.6/site-packages/sklearn/utils/validation.py:475: DataConversionWarning: Data with input dtype object was converted to float64 by MinMaxScaler.\n warnings.warn(msg, DataConversionWarning)\n" ], [ "del x_train\ngc.collect()", "_____no_output_____" ], [ "feature_columns = [\n tf.feature_column.numeric_column('x', shape=np.array(x_train_norm).shape[1:])]", "_____no_output_____" ], [ "from datetime import datetime\nprint(datetime.now())\nglobal_step = tf.Variable(0, trainable=False)\n\ntrain_input_fn = tf.estimator.inputs.numpy_input_fn(\n x={'x': x_train_norm}, y=y_train, batch_size=10000, num_epochs=5, shuffle=True) \nregressor = tf.estimator.DNNRegressor(\n feature_columns=feature_columns, hidden_units=[1000, 250, 60, 15],\n optimizer=tf.train.AdamOptimizer(learning_rate=0.0005, beta1=0.9, beta2=0.999, epsilon=0.1),\n batch_norm=True)\nregressor.train(input_fn=train_input_fn)\nprint(datetime.now())", "2018-09-25 22:41:30.577171\nINFO:tensorflow:Using default config.\nWARNING:tensorflow:Using temporary folder as model directory: /tmp/tmp8kx1d9wm\nINFO:tensorflow:Using config: {'_model_dir': '/tmp/tmp8kx1d9wm', '_tf_random_seed': None, '_save_summary_steps': 100, '_save_checkpoints_steps': None, '_save_checkpoints_secs': 600, '_session_config': None, '_keep_checkpoint_max': 5, '_keep_checkpoint_every_n_hours': 10000, '_log_step_count_steps': 100, '_train_distribute': None, '_device_fn': None, '_service': None, '_cluster_spec': <tensorflow.python.training.server_lib.ClusterSpec object at 0x7fa4abff4710>, '_task_type': 'worker', '_task_id': 0, '_global_id_in_cluster': 0, '_master': '', '_evaluation_master': '', '_is_chief': True, '_num_ps_replicas': 0, '_num_worker_replicas': 1}\nINFO:tensorflow:Calling model_fn.\nINFO:tensorflow:Done calling model_fn.\nINFO:tensorflow:Create CheckpointSaverHook.\nINFO:tensorflow:Graph was finalized.\nINFO:tensorflow:Running local_init_op.\nINFO:tensorflow:Done running local_init_op.\nINFO:tensorflow:Saving checkpoints for 0 into /tmp/tmp8kx1d9wm/model.ckpt.\nINFO:tensorflow:loss = 2100894.5, step = 0\nINFO:tensorflow:global_step/sec: 23.3797\nINFO:tensorflow:loss = 1588676.1, step = 100 (4.275 sec)\nINFO:tensorflow:global_step/sec: 24.6418\nINFO:tensorflow:loss = 1340190.1, step = 200 (4.058 sec)\nINFO:tensorflow:global_step/sec: 23.8619\nINFO:tensorflow:loss = 1249795.9, step = 300 (4.191 sec)\nINFO:tensorflow:global_step/sec: 25.0258\nINFO:tensorflow:loss = 1217116.8, step = 400 (3.996 sec)\nINFO:tensorflow:global_step/sec: 24.7377\nINFO:tensorflow:loss = 906687.9, step = 500 (4.043 sec)\nINFO:tensorflow:global_step/sec: 24.404\nINFO:tensorflow:loss = 797817.1, step = 600 (4.098 sec)\nINFO:tensorflow:global_step/sec: 24.4694\nINFO:tensorflow:loss = 642253.56, step = 700 (4.087 sec)\nINFO:tensorflow:global_step/sec: 25.4534\nINFO:tensorflow:loss = 626624.3, step = 800 (3.929 sec)\nINFO:tensorflow:global_step/sec: 25.1432\nINFO:tensorflow:loss = 442386.6, step = 900 (3.977 sec)\nINFO:tensorflow:global_step/sec: 26.4125\nINFO:tensorflow:loss = 379201.22, step = 1000 (3.786 sec)\nINFO:tensorflow:global_step/sec: 23.7181\nINFO:tensorflow:loss = 374420.78, step = 1100 (4.217 sec)\nINFO:tensorflow:global_step/sec: 24.1379\nINFO:tensorflow:loss = 257228.69, step = 1200 (4.142 sec)\nINFO:tensorflow:global_step/sec: 24.206\nINFO:tensorflow:loss = 247167.1, step = 1300 (4.131 sec)\nINFO:tensorflow:global_step/sec: 23.2165\nINFO:tensorflow:loss = 231606.47, step = 1400 (4.307 sec)\nINFO:tensorflow:global_step/sec: 25.1209\nINFO:tensorflow:loss = 171092.22, step = 1500 (3.981 sec)\nINFO:tensorflow:global_step/sec: 25.0705\nINFO:tensorflow:loss = 262952.78, step = 1600 (3.989 sec)\nINFO:tensorflow:global_step/sec: 24.4339\nINFO:tensorflow:loss = 183970.11, step = 1700 (4.093 sec)\nINFO:tensorflow:global_step/sec: 24.384\nINFO:tensorflow:loss = 201980.48, step = 1800 (4.100 sec)\nINFO:tensorflow:global_step/sec: 22.8031\nINFO:tensorflow:loss = 188921.6, step = 1900 (4.386 sec)\nINFO:tensorflow:global_step/sec: 24.8342\nINFO:tensorflow:loss = 239098.06, step = 2000 (4.027 sec)\nINFO:tensorflow:global_step/sec: 24.6229\nINFO:tensorflow:loss = 200540.88, step = 2100 (4.061 sec)\nINFO:tensorflow:global_step/sec: 24.7353\nINFO:tensorflow:loss = 186977.1, step = 2200 (4.043 sec)\nINFO:tensorflow:global_step/sec: 26.0812\nINFO:tensorflow:loss = 182712.78, step = 2300 (3.834 sec)\nINFO:tensorflow:global_step/sec: 24.627\nINFO:tensorflow:loss = 221842.72, step = 2400 (4.061 sec)\nINFO:tensorflow:global_step/sec: 24.547\nINFO:tensorflow:loss = 199860.28, step = 2500 (4.074 sec)\nINFO:tensorflow:global_step/sec: 24.4057\nINFO:tensorflow:loss = 167536.78, step = 2600 (4.097 sec)\nINFO:tensorflow:global_step/sec: 24.3356\nINFO:tensorflow:loss = 212020.58, step = 2700 (4.109 sec)\nINFO:tensorflow:global_step/sec: 24.1589\nINFO:tensorflow:loss = 214286.42, step = 2800 (4.139 sec)\nINFO:tensorflow:global_step/sec: 24.6555\nINFO:tensorflow:loss = 196224.3, step = 2900 (4.056 sec)\nINFO:tensorflow:global_step/sec: 24.7534\nINFO:tensorflow:loss = 158533.44, step = 3000 (4.040 sec)\nINFO:tensorflow:global_step/sec: 24.6528\nINFO:tensorflow:loss = 196333.12, step = 3100 (4.057 sec)\nINFO:tensorflow:global_step/sec: 24.133\nINFO:tensorflow:loss = 146130.5, step = 3200 (4.144 sec)\nINFO:tensorflow:global_step/sec: 25.2115\nINFO:tensorflow:loss = 199063.34, step = 3300 (3.980 sec)\nINFO:tensorflow:global_step/sec: 24.3842\nINFO:tensorflow:loss = 316612.3, step = 3400 (4.088 sec)\nINFO:tensorflow:global_step/sec: 25.2684\nINFO:tensorflow:loss = 199040.06, step = 3500 (3.957 sec)\nINFO:tensorflow:global_step/sec: 24.502\nINFO:tensorflow:loss = 213671.53, step = 3600 (4.081 sec)\nINFO:tensorflow:global_step/sec: 24.8003\nINFO:tensorflow:loss = 167795.62, step = 3700 (4.032 sec)\nINFO:tensorflow:global_step/sec: 24.6228\nINFO:tensorflow:loss = 172772.16, step = 3800 (4.061 sec)\nINFO:tensorflow:global_step/sec: 23.8539\nINFO:tensorflow:loss = 252507.66, step = 3900 (4.192 sec)\nINFO:tensorflow:global_step/sec: 24.8628\nINFO:tensorflow:loss = 942760.75, step = 4000 (4.022 sec)\nINFO:tensorflow:global_step/sec: 23.3657\nINFO:tensorflow:loss = 157312.02, step = 4100 (4.280 sec)\nINFO:tensorflow:global_step/sec: 25.2674\nINFO:tensorflow:loss = 252486.06, step = 4200 (3.957 sec)\nINFO:tensorflow:global_step/sec: 24.3134\nINFO:tensorflow:loss = 210227.88, step = 4300 (4.113 sec)\nINFO:tensorflow:global_step/sec: 24.8007\nINFO:tensorflow:loss = 166473.67, step = 4400 (4.032 sec)\nINFO:tensorflow:global_step/sec: 22.5815\nINFO:tensorflow:loss = 208260.72, step = 4500 (4.428 sec)\nINFO:tensorflow:global_step/sec: 24.7108\nINFO:tensorflow:loss = 200092.44, step = 4600 (4.047 sec)\nINFO:tensorflow:global_step/sec: 24.3366\nINFO:tensorflow:loss = 167874.92, step = 4700 (4.109 sec)\nINFO:tensorflow:global_step/sec: 25.0432\nINFO:tensorflow:loss = 182072.22, step = 4800 (3.993 sec)\nINFO:tensorflow:global_step/sec: 24.4943\nINFO:tensorflow:loss = 159979.19, step = 4900 (4.082 sec)\nINFO:tensorflow:global_step/sec: 24.8842\nINFO:tensorflow:loss = 333354.38, step = 5000 (4.019 sec)\nINFO:tensorflow:global_step/sec: 24.3327\nINFO:tensorflow:loss = 187568.56, step = 5100 (4.111 sec)\nINFO:tensorflow:global_step/sec: 23.3797\nINFO:tensorflow:loss = 191646.6, step = 5200 (4.276 sec)\nINFO:tensorflow:global_step/sec: 25.153\nINFO:tensorflow:loss = 182431.7, step = 5300 (3.976 sec)\nINFO:tensorflow:global_step/sec: 23.4958\nINFO:tensorflow:loss = 167369.86, step = 5400 (4.257 sec)\nINFO:tensorflow:global_step/sec: 23.7227\nINFO:tensorflow:loss = 184402.61, step = 5500 (4.214 sec)\nINFO:tensorflow:global_step/sec: 23.4563\nINFO:tensorflow:loss = 186296.03, step = 5600 (4.264 sec)\nINFO:tensorflow:global_step/sec: 24.3707\nINFO:tensorflow:loss = 203567.67, step = 5700 (4.103 sec)\nINFO:tensorflow:global_step/sec: 23.9675\nINFO:tensorflow:loss = 192690.33, step = 5800 (4.172 sec)\nINFO:tensorflow:global_step/sec: 24.5756\nINFO:tensorflow:loss = 165997.27, step = 5900 (4.069 sec)\nINFO:tensorflow:global_step/sec: 24.4804\nINFO:tensorflow:loss = 212627.19, step = 6000 (4.085 sec)\nINFO:tensorflow:global_step/sec: 24.5344\nINFO:tensorflow:loss = 183376.9, step = 6100 (4.076 sec)\nINFO:tensorflow:global_step/sec: 22.9439\nINFO:tensorflow:loss = 220011.53, step = 6200 (4.358 sec)\nINFO:tensorflow:global_step/sec: 25.8403\nINFO:tensorflow:loss = 229227.62, step = 6300 (3.870 sec)\nINFO:tensorflow:global_step/sec: 23.8219\nINFO:tensorflow:loss = 166229.53, step = 6400 (4.198 sec)\nINFO:tensorflow:global_step/sec: 24.6293\nINFO:tensorflow:loss = 165056.47, step = 6500 (4.060 sec)\nINFO:tensorflow:global_step/sec: 23.6756\nINFO:tensorflow:loss = 182097.08, step = 6600 (4.224 sec)\nINFO:tensorflow:global_step/sec: 24.8445\nINFO:tensorflow:loss = 192676.38, step = 6700 (4.025 sec)\nINFO:tensorflow:global_step/sec: 23.9207\nINFO:tensorflow:loss = 184007.3, step = 6800 (4.181 sec)\nINFO:tensorflow:global_step/sec: 24.1643\nINFO:tensorflow:loss = 217440.89, step = 6900 (4.138 sec)\nINFO:tensorflow:global_step/sec: 24.5124\nINFO:tensorflow:loss = 169177.8, step = 7000 (4.080 sec)\nINFO:tensorflow:global_step/sec: 24.2017\nINFO:tensorflow:loss = 346739.34, step = 7100 (4.132 sec)\nINFO:tensorflow:global_step/sec: 25.161\nINFO:tensorflow:loss = 201689.31, step = 7200 (3.974 sec)\n" ], [ "val_df = (pd\n .read_parquet(\"../../datasets/kaggle/new-york-city-taxi-fare-prediction/train_full_2\")\n .sample(frac=0.20, random_state=420)\n )", "_____no_output_____" ], [ "print(val_df.shape)\nprint(val_df.columns)\nprint(val_df.info())", "(10849366, 46)\nIndex(['fare_amount', 'pickup_datetime', 'pickup_longitude', 'pickup_latitude',\n 'dropoff_longitude', 'dropoff_latitude', 'passenger_count',\n 'distance_miles', 'is_pickup_JFK_new', 'is_dropoff_JFK_new',\n 'is_pickup_EWR_new', 'is_dropoff_EWR_new', 'is_pickup_LGA_new',\n 'is_dropoff_LGA_new', 'is_to_from_JFK_new', 'distance_to_center',\n 'hour', 'weekday', 'year_2009', 'year_2010', 'year_2011', 'year_2012',\n 'year_2013', 'year_2014', 'year_2015', 'hour_period_4', 'hour_period_0',\n 'hour_period_1', 'hour_period_2', 'hour_period_5', 'hour_period_3',\n 'distance_to_center_dropoff', 'pickup_zone_2', 'pickup_zone_0',\n 'pickup_zone_1', 'pickup_zone_3', 'pickup_zone_4', 'pickup_zone_5',\n 'pickup_zone_6', 'dropoff_zone_2', 'dropoff_zone_0', 'dropoff_zone_1',\n 'dropoff_zone_3', 'dropoff_zone_4', 'dropoff_zone_5', 'dropoff_zone_6'],\n dtype='object')\n<class 'pandas.core.frame.DataFrame'>\nInt64Index: 10849366 entries, 17761954 to 18905951\nData columns (total 46 columns):\nfare_amount float32\npickup_datetime datetime64[ns]\npickup_longitude float32\npickup_latitude float32\ndropoff_longitude float32\ndropoff_latitude float32\npassenger_count uint8\ndistance_miles float32\nis_pickup_JFK_new bool\nis_dropoff_JFK_new bool\nis_pickup_EWR_new bool\nis_dropoff_EWR_new bool\nis_pickup_LGA_new bool\nis_dropoff_LGA_new bool\nis_to_from_JFK_new bool\ndistance_to_center float32\nhour int64\nweekday int64\nyear_2009 uint8\nyear_2010 uint8\nyear_2011 uint8\nyear_2012 uint8\nyear_2013 uint8\nyear_2014 uint8\nyear_2015 uint8\nhour_period_4 uint8\nhour_period_0 uint8\nhour_period_1 uint8\nhour_period_2 uint8\nhour_period_5 uint8\nhour_period_3 uint8\ndistance_to_center_dropoff float32\npickup_zone_2 uint8\npickup_zone_0 uint8\npickup_zone_1 uint8\npickup_zone_3 uint8\npickup_zone_4 uint8\npickup_zone_5 uint8\npickup_zone_6 uint8\ndropoff_zone_2 uint8\ndropoff_zone_0 uint8\ndropoff_zone_1 uint8\ndropoff_zone_3 uint8\ndropoff_zone_4 uint8\ndropoff_zone_5 uint8\ndropoff_zone_6 uint8\ndtypes: bool(7), datetime64[ns](1), float32(8), int64(2), uint8(28)\nmemory usage: 1.0 GB\nNone\n" ], [ "x_val = val_df[cols].values\ny_val = val_df[['fare_amount']].values", "_____no_output_____" ], [ "del val_df\ngc.collect()", "_____no_output_____" ], [ "x_val_norm = scaler.transform(x_val)", "_____no_output_____" ], [ "del x_val\ngc.collect()", "_____no_output_____" ], [ "val_input_fn = tf.estimator.inputs.numpy_input_fn(\n x={'x': x_val_norm}, y=y_val, num_epochs=1, shuffle=False)\nscores = regressor.evaluate(input_fn=val_input_fn)\nprint('MSE (tensorflow): {0:f}'.format(scores['average_loss']))", "INFO:tensorflow:Calling model_fn.\nINFO:tensorflow:Done calling model_fn.\nINFO:tensorflow:Starting evaluation at 2018-09-26-05:56:52\nINFO:tensorflow:Graph was finalized.\nINFO:tensorflow:Restoring parameters from /tmp/tmp8kx1d9wm/model.ckpt-21699\nINFO:tensorflow:Running local_init_op.\nINFO:tensorflow:Done running local_init_op.\nINFO:tensorflow:Finished evaluation at 2018-09-26-05:57:53\nINFO:tensorflow:Saving dict for global step 21699: average_loss = 26.408344, global_step = 21699, label/mean = 11.322498, loss = 3380.255, prediction/mean = 11.31766\nINFO:tensorflow:Saving 'checkpoint_path' summary for global step 21699: /tmp/tmp8kx1d9wm/model.ckpt-21699\nMSE (tensorflow): 26.408344\n" ], [ "predictions = regressor.predict(input_fn=val_input_fn)\ny_predicted = np.array(list(p['predictions'] for p in predictions))\ny_predicted = y_predicted.reshape(np.array(y_val).shape)\nscore_sklearn = metrics.mean_squared_error(y_predicted, y_val)\nprint('MSE (sklearn): {0:f}'.format(score_sklearn))\ny_predicted.shape", "INFO:tensorflow:Calling model_fn.\nINFO:tensorflow:Done calling model_fn.\nINFO:tensorflow:Graph was finalized.\nINFO:tensorflow:Restoring parameters from /tmp/tmp8kx1d9wm/model.ckpt-21699\nINFO:tensorflow:Running local_init_op.\nINFO:tensorflow:Done running local_init_op.\nMSE (sklearn): 26.408339\n" ], [ "y_predicted[0:2]", "_____no_output_____" ], [ "test_df = (pd\n .read_parquet(\"../../datasets/kaggle/new-york-city-taxi-fare-prediction/test_full_2\")\n )", "_____no_output_____" ], [ "x_test = test_df[cols].values\nx_test_norm = scaler.transform(x_test)", "_____no_output_____" ], [ "test_input_fn = tf.estimator.inputs.numpy_input_fn(\n x={'x': x_test_norm}, num_epochs=1, shuffle=False)\ntest_predictions = regressor.predict(input_fn=test_input_fn)\ny_predicted = np.array(list(p['predictions'] for p in test_predictions))\ny_predicted = y_predicted.reshape((9914,1))\n#y_predicted = np.array(list(test_predictions))\n#y_predicted = y_predicted.reshape((9914,1))\ny_predicted.shape", "INFO:tensorflow:Calling model_fn.\nINFO:tensorflow:Done calling model_fn.\nINFO:tensorflow:Graph was finalized.\nINFO:tensorflow:Restoring parameters from /tmp/tmp8kx1d9wm/model.ckpt-21699\nINFO:tensorflow:Running local_init_op.\nINFO:tensorflow:Done running local_init_op.\n" ], [ "preds = [p[0] for p in y_predicted]\npreds", "_____no_output_____" ], [ "# Write the predictions to a CSV file which we can submit to the competition.\nsubmission = pd.DataFrame(\n {'key': test_df.key, 'fare_amount': preds},\n columns = ['key', 'fare_amount'])\nsubmission.to_csv('../../datasets/kaggle/new-york-city-taxi-fare-prediction/submission.csv', index = False)", "_____no_output_____" ], [ "submission.describe()", "_____no_output_____" ] ] ]

[ "code", "markdown", "code" ]

[ [ "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad61a01e9a15a8a0c295182238cc674e42685c7

52,957

ipynb

Jupyter Notebook

2nd-ML100Days/homework/D-085/Day085_CB_EarlyStop.ipynb

qwerzxcv98/100Day-ML-Marathon

3c86cd083b086a1e4b7a55a41b127909c7860f79

[ "MIT" ]

3

2019-08-22T15:19:11.000Z

2019-08-24T00:54:54.000Z

2nd-ML100Days/homework/D-085/Day085_CB_EarlyStop.ipynb

magikerwin1993/100Day-ML-Marathon

3c86cd083b086a1e4b7a55a41b127909c7860f79

[ "MIT" ]

null

2nd-ML100Days/homework/D-085/Day085_CB_EarlyStop.ipynb

magikerwin1993/100Day-ML-Marathon

3c86cd083b086a1e4b7a55a41b127909c7860f79

[ "MIT" ]

1

2019-07-18T01:52:04.000Z

147.924581

20,896

0.838643

[ [ [ "## 範例重點\n* 學習如何在 keras 中加入 EarlyStop\n* 知道如何設定監控目標\n* 比較有無 earlystopping 對 validation 的影響", "_____no_output_____" ] ], [ [ "import os\nfrom tensorflow import keras\n\n# 本範例不需使用 GPU, 將 GPU 設定為 \"無\"\nos.environ[\"CUDA_VISIBLE_DEVICES\"] = \"0\"", "_____no_output_____" ], [ "train, test = keras.datasets.cifar10.load_data()", "_____no_output_____" ], [ "## 資料前處理\ndef preproc_x(x, flatten=True):\n x = x / 255.\n if flatten:\n x = x.reshape((len(x), -1))\n return x\n\ndef preproc_y(y, num_classes=10):\n if y.shape[-1] == 1:\n y = keras.utils.to_categorical(y, num_classes)\n return y ", "_____no_output_____" ], [ "x_train, y_train = train\nx_test, y_test = test\n\n# 資料前處理 - X 標準化\nx_train = preproc_x(x_train)\nx_test = preproc_x(x_test)\n\n# 資料前處理 -Y 轉成 onehot\ny_train = preproc_y(y_train)\ny_test = preproc_y(y_test)", "_____no_output_____" ], [ "from tensorflow.keras.layers import BatchNormalization\n\n\"\"\"\n建立神經網路，並加入 BN layer\n\"\"\"\ndef build_mlp(input_shape, output_units=10, num_neurons=[256, 128, 64]):\n input_layer = keras.layers.Input(input_shape)\n \n for i, n_units in enumerate(num_neurons):\n if i == 0:\n x = keras.layers.Dense(units=n_units, \n activation=\"relu\", \n name=\"hidden_layer\"+str(i+1))(input_layer)\n x = BatchNormalization()(x)\n else:\n x = keras.layers.Dense(units=n_units, \n activation=\"relu\", \n name=\"hidden_layer\"+str(i+1))(x)\n x = BatchNormalization()(x)\n \n out = keras.layers.Dense(units=output_units, activation=\"softmax\", name=\"output\")(x)\n \n model = keras.models.Model(inputs=[input_layer], outputs=[out])\n return model", "_____no_output_____" ], [ "## 超參數設定\nLEARNING_RATE = 1e-3\nEPOCHS = 50\nBATCH_SIZE = 1024\nMOMENTUM = 0.95", "_____no_output_____" ], [ "\"\"\"\n# 載入 Callbacks, 並將 monitor 設定為監控 validation loss\n\"\"\"\nfrom tensorflow.keras.callbacks import EarlyStopping\n\nearlystop = EarlyStopping(monitor=\"val_loss\", \n patience=5, \n verbose=1\n )", "_____no_output_____" ], [ "model = build_mlp(input_shape=x_train.shape[1:])\nmodel.summary()\noptimizer = keras.optimizers.SGD(lr=LEARNING_RATE, nesterov=True, momentum=MOMENTUM)\nmodel.compile(loss=\"categorical_crossentropy\", metrics=[\"accuracy\"], optimizer=optimizer)\n\nmodel.fit(x_train, y_train, \n epochs=EPOCHS, \n batch_size=BATCH_SIZE, \n validation_data=(x_test, y_test), \n shuffle=True,\n callbacks=[earlystop]\n )", "WARNING: Logging before flag parsing goes to stderr.\nW0720 17:57:39.040111 10944 deprecation.py:323] From c:\\users\\qwerz\\miniconda3\\envs\\ml100\\lib\\site-packages\\tensorflow\\python\\ops\\math_grad.py:1250: add_dispatch_support.<locals>.wrapper (from tensorflow.python.ops.array_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse tf.where in 2.0, which has the same broadcast rule as np.where\n" ], [ "# Collect results\ntrain_loss = model.history.history[\"loss\"]\nvalid_loss = model.history.history[\"val_loss\"]\ntrain_acc = model.history.history[\"accuracy\"]\nvalid_acc = model.history.history[\"val_accuracy\"]", "_____no_output_____" ], [ "import matplotlib.pyplot as plt\n%matplotlib inline\n\nplt.plot(range(len(train_loss)), train_loss, label=\"train loss\")\nplt.plot(range(len(valid_loss)), valid_loss, label=\"valid loss\")\nplt.legend()\nplt.title(\"Loss\")\nplt.show()\n\nplt.plot(range(len(train_acc)), train_acc, label=\"train accuracy\")\nplt.plot(range(len(valid_acc)), valid_acc, label=\"valid accuracy\")\nplt.legend()\nplt.title(\"Accuracy\")\nplt.show()", "_____no_output_____" ] ], [ [ "## Work\n1. 試改變 monitor \"Validation Accuracy\" 並比較結果\n2. 調整 earlystop 的等待次數至 10, 25 並比較結果", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown" ]

[ [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ] ]

4ad61ea66edbb683f2b72266b0e8a2d8fec9477f

23,835

ipynb

Jupyter Notebook

2 preliminaries/.ipynb_checkpoints/autograd-checkpoint.ipynb

pedro-abundio-wang/d2l-numpy

59e3e536f81d355f10a99a4e936d2b3e68201f1d

[ "Apache-2.0" ]

null

2 preliminaries/.ipynb_checkpoints/autograd-checkpoint.ipynb

pedro-abundio-wang/d2l-numpy

59e3e536f81d355f10a99a4e936d2b3e68201f1d

[ "Apache-2.0" ]

15

2019-10-10T13:01:15.000Z

2022-02-10T00:21:14.000Z

2 preliminaries/.ipynb_checkpoints/autograd-checkpoint.ipynb

pedro-abundio-wang/d2l-numpy

59e3e536f81d355f10a99a4e936d2b3e68201f1d

[ "Apache-2.0" ]

null

29.57196

471

0.568827

[ [ [ "# Automatic Differentiation\n\n:label:`chapter_autograd`\n\n\nIn machine learning, we *train* models, updating them successively \nso that they get better and better as they see more and more data. \nUsually, *getting better* means minimizing a *loss function*, \na score that answers the question \"how *bad* is our model?\"\nThis question is more subtle than it appears.\nUltimately, what we really care about \nis producing a model that performs well on data \nthat we have never seen before.\nBut we can only fit the model to data that we can actually see.\nThus we can decompose the task of fitting models into two key concerns:\n*optimization* the process of fitting our models to observed data\nand *generalization* the mathematical principles and practitioners wisdom\nthat guide as to how to produce models whose validity extends \nbeyond the exact set of datapoints used to train it. \n\nThis section addresses the calculation of derivatives,\na crucial step in nearly all deep learning optimization algorithms.\nWith neural networks, we typically choose loss functions \nthat are differentiable with respect to our model's parameters.\nPut simply, this means that for each parameter, \nwe can determine how rapidly the loss would increase or decrease,\nwere we to *increase* or *decrease* that parameter\nby an infinitessimally small amount.\nWhile the calculations for taking these derivatives are straightforward,\nrequiring only some basic calculus, \nfor complex models, working out the updates by hand\ncan be a pain (and often error-prone).\n\nThe autograd package expedites this work \nby automatically calculating derivatives. \nAnd while many other libraries require \nthat we compile a symbolic graph to take automatic derivatives, \n`autograd` allows us to take derivatives \nwhile writing ordinary imperative code. \nEvery time we pass data through our model, \n`autograd` builds a graph on the fly, \ntracking which data combined through \nwhich operations to produce the output. \nThis graph enables `autograd` \nto subsequently backpropagate gradients on command. \nHere, *backpropagate* simply means to trace through the compute graph, \nfilling in the partial derivatives with respect to each parameter. \nIf you are unfamiliar with some of the math, \ne.g., gradients, please refer to :numref:`chapter_math`.", "_____no_output_____" ] ], [ [ "from mxnet import autograd, np, npx\nnpx.set_np()", "_____no_output_____" ] ], [ [ "## A Simple Example\n\nAs a toy example, say that we are interested \nin differentiating the mapping \n$y = 2\\mathbf{x}^{\\top}\\mathbf{x}$ \nwith respect to the column vector $\\mathbf{x}$. \nTo start, let's create the variable `x` and assign it an initial value.", "_____no_output_____" ] ], [ [ "x = np.arange(4)\nx", "_____no_output_____" ] ], [ [ "Note that before we even calculate the gradient \nof ``y`` with respect to ``x``, \nwe will need a place to store it. \nIt's important that we do not allocate new memory\nevery time we take a derivative with respect to a parameter\nbecause we will often update the same parameters \nthousands or millions of times \nand could quickly run out of memory.\n\nNote also that a gradient with respect to a vector $x$ \nis itself vector-valued and has the same shape as $x$.\nThus it is intuitive that in code, \nwe will access a gradient taken with respect to `x` \nas an attribute the `ndarray` `x` itself.\nWe allocate memory for an `ndarray`'s gradient\nby invoking its ``attach_grad()`` method.", "_____no_output_____" ] ], [ [ "x.attach_grad()", "_____no_output_____" ] ], [ [ "After we calculate a gradient taken with respect to `x`, \nwe will be able to access it via the `.grad` attribute. \nAs a safe default, `x.grad` initializes as an array containing all zeros.\nThat's sensible because our most common use case \nfor taking gradient in deep learning is to subsequently \nupdate parameters by adding (or subtracting) the gradient\nto maximize (or minimize) the differentiated function.\nBy initializing the gradient to $\\mathbf{0}$,\nwe ensure that any update accidentally exectuted \nbefore a gradient has actually been calculated\nwill not alter the variable's value.", "_____no_output_____" ] ], [ [ "x.grad", "_____no_output_____" ] ], [ [ "Now let's calculate ``y``. \nBecause we wish to subsequently calculate gradients \nwe want MXNet to generate a computation graph on the fly. \nWe could imagine that MXNet would be turning on a recording device \nto capture the exact path by which each variable is generated.\n\nNote that building the computation graph \nrequires a nontrivial amount of computation. \nSo MXNet will only build the graph when explicitly told to do so. \nWe can invoke this behavior by placing our code \ninside a ``with autograd.record():`` block.", "_____no_output_____" ] ], [ [ "with autograd.record():\n y = 2.0 * np.dot(x, x)\ny", "_____no_output_____" ] ], [ [ "Since `x` is an `ndarray` of length 4, \n`np.dot` will perform an inner product of `x` and `x`,\nyielding the scalar output that we assign to `y`. \nNext, we can automatically calculate the gradient of `y`\nwith respect to each component of `x` \nby calling `y`'s `backward` function.", "_____no_output_____" ] ], [ [ "y.backward()", "_____no_output_____" ] ], [ [ "If we recheck the value of `x.grad`, we will find its contents overwritten by the newly calculated gradient.", "_____no_output_____" ] ], [ [ "x.grad", "_____no_output_____" ] ], [ [ "The gradient of the function $y = 2\\mathbf{x}^{\\top}\\mathbf{x}$ \nwith respect to $\\mathbf{x}$ should be $4\\mathbf{x}$. \nLet's quickly verify that our desired gradient was calculated correctly.\nIf the two `ndarray`s are indeed the same, \nthen their difference should consist of all zeros.", "_____no_output_____" ] ], [ [ "x.grad - 4 * x", "_____no_output_____" ] ], [ [ "If we subsequently compute the gradient of another variable\nwhose value was calculated as a function of `x`, \nthe contents of `x.grad` will be overwritten.", "_____no_output_____" ] ], [ [ "with autograd.record():\n y = x.sum()\ny.backward()\nx.grad", "_____no_output_____" ] ], [ [ "## Backward for Non-scalar Variable\n\nTechnically, when `y` is not a scalar, \nthe most natural interpretation of the gradient of `y` (a vector of length $m$)\nwith respect to `x` (a vector of length $n$) is the Jacobian (an $m\\times n$ matrix).\nFor higher-order and higher-dimensional $y$ and $x$, \nthe Jacobian could be a gnarly high order tensor \nand complex to compute (refer to :numref:`chapter_math`). \n\nHowever, while these more exotic objects do show up \nin advanced machine learning (including in deep learning),\nmore often when we are calling backward on a vector,\nwe are trying to calculate the derivatives of the loss functions\nfor each constitutent of a *batch* of training examples.\nHere, our intent is not to calculate the Jacobian\nbut rather the sum of the partial derivatives \ncomputed individuall for each example in the batch.\n\nThus when we invoke backwards on a vector-valued variable,\nMXNet assumes that we want the sum of the gradients.\nIn short, MXNet, will create a new scalar variable \nby summing the elements in `y`,\nand compute the gradient of that variable with respect to `x`.", "_____no_output_____" ] ], [ [ "with autograd.record(): # y is a vector\n y = x * x\ny.backward()\n\nu = x.copy()\nu.attach_grad()\nwith autograd.record(): # v is scalar\n v = (u * u).sum()\nv.backward()\n\nx.grad - u.grad", "_____no_output_____" ] ], [ [ "## Advanced Autograd\n\nAlready you know enough to employ `autograd` and `ndarray` \nsuccessfully to develop many practical models. \nWhile the rest of this section is not necessary just yet,\nwe touch on a few advanced topics for completeness. \n\n### Detach Computations\n\nSometimes, we wish to move some calculations \noutside of the recorded computation graph. \nFor example, say that `y` was calculated as a function of `x`.\nAnd that subsequently `z` was calcatated a function of both `y` and `x`. \nNow, imagine that we wanted to calculate \nthe gradient of `z` with respect to `x`,\nbut wanted for some reason to treat `y` as a constant,\nand only take into account the role \nthat `x` played after `y` was calculated.\n\nHere, we can call `u = y.detach()` to return a new variable \nthat has the same values as `y` but discards any information\nabout how `u` was computed. \nIn other words, the gradient will not flow backwards through `u` to `x`. \nThis will provide the same functionality as if we had\ncalculated `u` as a function of `x` outside of the `autograd.record` scope, \nyielding a `u` that will be treated as a constant in any called to `backward`. \nThe following backward computes $\\partial (u \\odot x)/\\partial x$ \ninstead of $\\partial (x \\odot x \\odot x) /\\partial x$,\nwhere $\\odot$ stands for element-wise multiplication.", "_____no_output_____" ] ], [ [ "with autograd.record():\n y = x * x\n u = y.detach()\n z = u * x\nz.backward()\nx.grad - u", "_____no_output_____" ] ], [ [ "Since the computation of $y$ was recorded, \nwe can subsequently call `y.backward()` to get $\\partial y/\\partial x = 2x$.", "_____no_output_____" ] ], [ [ "y.backward()\nx.grad - 2*x", "_____no_output_____" ] ], [ [ "## Attach Gradients to Internal Variables\n\nAttaching gradients to a variable `x` implicitly calls `x=x.detach()`. \nIf `x` is computed based on other variables, \nthis part of computation will not be used in the backward function.", "_____no_output_____" ] ], [ [ "y = np.ones(4) * 2\ny.attach_grad()\nwith autograd.record():\n u = x * y\n u.attach_grad() # implicitly run u = u.detach()\n z = u + x\nz.backward()\nprint(x.grad, '\\n', u.grad, '\\n', y.grad)", "[1. 1. 1. 1.] \n [1. 1. 1. 1.] \n [0. 0. 0. 0.]\n" ] ], [ [ "## Head gradients\n\nDetaching allows to breaks the computation into several parts. We could use chain rule :numref:`chapter_math` to compute the gradient for the whole computation. Assume $u = f(x)$ and $z = g(u)$, by chain rule we have $\\frac{dz}{dx} = \\frac{dz}{du} \\frac{du}{dx}.$ To compute $\\frac{dz}{du}$, we can first detach $u$ from the computation and then call `z.backward()` to compute the first term.", "_____no_output_____" ] ], [ [ "y = np.ones(4) * 2\ny.attach_grad()\nwith autograd.record():\n u = x * y\n v = u.detach() # u still keeps the computation graph\n v.attach_grad()\n z = v + x\nz.backward()\nprint(x.grad, '\\n', y.grad)", "[1. 1. 1. 1.] \n [0. 0. 0. 0.]\n" ] ], [ [ "Subsequently, we can call `u.backward()` to compute the second term, \nbut pass the first term as the head gradients to multiply both terms \nso that `x.grad` will contains $\\frac{dz}{dx}$ instead of $\\frac{du}{dx}$.", "_____no_output_____" ] ], [ [ "u.backward(v.grad)\nprint(x.grad, '\\n', y.grad)", "[2. 2. 2. 2.] \n [0. 1. 2. 3.]\n" ] ], [ [ "## Computing the Gradient of Python Control Flow\n\nOne benefit of using automatic differentiation \nis that even if building the computational graph of a function \nrequired passing through a maze of Python control flow \n(e.g. conditionals, loops, and arbitrary function calls), \nwe can still calculate the gradient of the resulting variable. \nIn the following snippet, note that \nthe number of iterations of the `while` loop \nand the evaluation of the `if` statement\nboth depend on the value of the input `b`.", "_____no_output_____" ] ], [ [ "def f(a):\n b = a * 2\n while np.abs(b).sum() < 1000:\n b = b * 2\n if b.sum() > 0:\n c = b\n else:\n c = 100 * b\n return c", "_____no_output_____" ] ], [ [ "Again to compute gradients, we just need to `record` the calculation\nand then call the `backward` function.", "_____no_output_____" ] ], [ [ "a = np.random.normal()\na.attach_grad()\nwith autograd.record():\n d = f(a)\nd.backward()", "_____no_output_____" ] ], [ [ "We can now analyze the `f` function defined above. \nNote that it is piecewise linear in its input `a`. \nIn other words, for any `a` there exists some constant \nsuch that for a given range `f(a) = g * a`. \nConsequently `d / a` allows us to verify that the gradient is correct:", "_____no_output_____" ] ], [ [ "print(a.grad == (d / a))", "1.0\n" ] ], [ [ "## Training Mode and Prediction Mode\n\nAs we have seen, after we call `autograd.record`, \nMXNet logs the operations in the following block. \nThere is one more subtle detail to be aware of.\nAdditionally, `autograd.record` will change \nthe running mode from *prediction* mode to *training* mode. \nWe can verify this behavior by calling the `is_training` function.", "_____no_output_____" ] ], [ [ "print(autograd.is_training())\nwith autograd.record():\n print(autograd.is_training())", "False\nTrue\n" ] ], [ [ "When we get to complicated deep learning models,\nwe will encounter some algorithms where the model\nbehaves differently during training and \nwhen we subsequently use it to make predictions. \nThe popular neural network techniques *dropout* :numref:`chapter_dropout` \nand *batch normalization* :numref:`chapter_batch_norm`\nboth exhibit this characteristic.\nIn other cases, our models may store auxiliary variables in *training* mode \nfor purposes of make computing gradients easier \nthat are not necessary at prediction time. \nWe will cover these differences in detail in later chapters. \n\n\n## Summary\n\n* MXNet provides an `autograd` package to automate the calculation of derivatives. To use it, we first attach gradients to those variables with respect to which we desire partial derivartives. We then record the computation of our target value, executed its backward function, and access the resulting gradient via our variable's `grad` attribute.\n* We can detach gradients and pass head gradients to the backward function to control the part of the computation will be used in the backward function.\n* The running modes of MXNet include *training mode* and *prediction mode*. We can determine the running mode by calling `autograd.is_training()`.\n\n## Exercises\n\n1. Try to run `y.backward()` twice.\n1. In the control flow example where we calculate the derivative of `d` with respect to `a`, what would happen if we changed the variable `a` to a random vector or matrix. At this point, the result of the calculation `f(a)` is no longer a scalar. What happens to the result? How do we analyze this?\n1. Redesign an example of finding the gradient of the control flow. Run and analyze the result.\n1. In a second-price auction (such as in eBay or in computational advertising), the winning bidder pays the second-highest price. Compute the gradient of the final price with respect to the winning bidder's bid using `autograd`. What does the result tell you about the mechanism? If you are curious to learn more about second-price auctions, check out this paper by [Edelman, Ostrovski and Schwartz, 2005](https://www.benedelman.org/publications/gsp-060801.pdf).\n1. Why is the second derivative much more expensive to compute than the first derivative?\n1. Derive the head gradient relationship for the chain rule. If you get stuck, use the [\"Chain rule\" article on Wikipedia](https://en.wikipedia.org/wiki/Chain_rule).\n1. Assume $f(x) = \\sin(x)$. Plot $f(x)$ and $\\frac{df(x)}{dx}$ on a graph, where you computed the latter without any symbolic calculations, i.e. without exploiting that $f'(x) = \\cos(x)$.\n\n## Scan the QR Code to [Discuss](https://discuss.mxnet.io/t/2318)\n\n![](../img/qr_autograd.svg)", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ] ]

4ad621ddde4afc4d3d6f4a509da5e0be5068128f

440,152

ipynb

Jupyter Notebook

07/CS480_Assignment_7.ipynb

Yiming-S/cs480student

c03bf6e2fe6e1501322ec8a34f51f76945ef1d06

[ "MIT" ]

null

07/CS480_Assignment_7.ipynb

Yiming-S/cs480student

c03bf6e2fe6e1501322ec8a34f51f76945ef1d06

[ "MIT" ]

null

07/CS480_Assignment_7.ipynb

Yiming-S/cs480student

c03bf6e2fe6e1501322ec8a34f51f76945ef1d06

[ "MIT" ]

1

2021-09-27T01:20:51.000Z

654.988095

367,236

0.949086

[ [ [ "![CS480_w.png](data:image/png;base64,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)\n", "_____no_output_____" ], [ "# Assignment 7", "_____no_output_____" ], [ "**Yiming Shen**\n\n**5/08/2021**\n\n**HW Topic: maging Data Commons**", "_____no_output_____" ] ], [ [ "# In this assignment, we will explore the Imaging Data Commons,\n# a new service from the National Cancer Institute.\n#\n# We will find and explore a PET dataset and then perform a MIP reconstruction.", "_____no_output_____" ] ], [ [ "![idc.png](data:image/png;base64,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)", "_____no_output_____" ], [ "## **Task 1**: Browse the IDC datasets and find a combined CT or MRI with PET! \n[15 Points]", "_____no_output_____" ] ], [ [ "#\n# Visit https://portal.imaging.datacommons.cancer.gov/ and find a dataset\n# that includes structural imaging such as CT or MRI combined with a PET scan.\n#\n# You can start by selecting the type of a cancer case such as Bladder, Brain etc.\n# Hint: Head and Neck include many PET scans.\n#", "_____no_output_____" ], [ "# TODO: Open the integrated OHIF viewer to view the data.\n#\n# 1) Select a collection, then a case, then a study in the IDC interface.\n# 2) To open the OHIF viewer, click on the view icon.\n# 3) Paste a screenshot of the viewer after the data was loaded below.\n#", "_____no_output_____" ] ], [ [ "![p1](https://raw.githubusercontent.com/Yiming-S/cs480student/main/07/1.png)", "_____no_output_____" ], [ "## **Task 2**: Visualize structural and PET side-by-side! \n[30 Points]", "_____no_output_____" ] ], [ [ "# In the OHIF viewer, please change the layout to show the structural image\n# next to the PET scan.\n#\n# TODO: Use the LAYOUT functionality in the toolbar to switch to a 2-column layout.", "_____no_output_____" ], [ "# TODO: Find the cancerous lesion.\n#\n# 1) Navigate through the PET image to find the cancerous lesion.", "_____no_output_____" ], [ "# TODO: Sync the structural scan (CT or MR) to the PET scan with the lesion.\n#\n# Hint: The \"Loc\" label in the bottom left of both scans should roughly match.\n# Hint 2: The Levels tool allows Window/Level adjustment to increase contrast.\n#", "_____no_output_____" ], [ "# TODO: Paste a screenshot showing the 2-column layout below.", "_____no_output_____" ] ], [ [ "![p2](https://github.com/Yiming-S/cs480student/blob/main/07/2.png?raw=true)", "_____no_output_____" ], [ "## **Task 3**: Use an MIP to reslice the PET scan! \n[35 Points]", "_____no_output_____" ] ], [ [ "# We will now use the 2D MPR functionality to look at the lesion from different\n# orientations.\n#\n# Hint: This works best in Google Chrome.", "_____no_output_____" ] ], [ [ "#### Question 1): What is MPR?\n", "_____no_output_____" ] ], [ [ "# TODO: YOUR ANSWER", "_____no_output_____" ] ], [ [ "**Multiplanar reformation or reconstruction (MPR)** involves the process of converting data from an imaging modality acquired in a certain plane, usually axial, into another plane 1. It is most commonly performed with thin-slice data from volumetric CT in the axial plane, but it may be accomplished with scanning in any plane and whichever modality capable of cross-sectional imaging, including magnetic resonance imaging (MRI), PET and SPECT.", "_____no_output_____" ], [ "#### Question 2): What is the difference between MIP, MinIP, AvgIP?", "_____no_output_____" ] ], [ [ "# TODO: YOUR ANSWER", "_____no_output_____" ] ], [ [ "**MIP : Maximum Intensity Projection** \nMaximum Intensity Projection (MIP) is a volume rendering technique for 3D images that projects in the visualization plane the voxels with maximum intensity that fall in the way of parallel rays traced from the viewpoint to the plane of projection.\n\n**MinIP : Minimum Intensity Projection** \nMinimum-intensity projection (MinIP) is multiplanar slab images produced by displaying only the lowest attenuation value encountered along a ray cast through an object toward the viewer's eye. It is a data visualization method that enables detection of low density structures in a given volume.\n\n**AvgIP :Average Intensity Projection** \nThe average intensity projection simply averages all of the pixel values in the stacks to make the final projected image. For example, the values for the pixel at (0,0), with three layers' values(190,200 and 195), will become (190 + 200 + 195) / 3 = 195.", "_____no_output_____" ] ], [ [ "# Now, please navigate to the lesion in all 3 orientations.\n#\n# TODO: Paste a screenshot below.\n#", "_____no_output_____" ] ], [ [ "![p4](https://github.com/Yiming-S/cs480student/blob/main/07/4.png?raw=true)", "_____no_output_____" ], [ "TODO: Add screenshot", "_____no_output_____" ], [ "#### Question 3): What does the Slab Thickness slider do?", "_____no_output_____" ] ], [ [ "# TODO: YOUR ANSWER", "_____no_output_____" ] ], [ [ "It can change the **thickness** of a show area.", "_____no_output_____" ], [ "## **Task 4**: Let's access the pixel data! [20 Points]", "_____no_output_____" ] ], [ [ "# Please EXIT THE 2D MPR.\n#\n# We will now grab the pixel data of the currently displayed slice\n# using the Javascript console.\n#", "_____no_output_____" ], [ "# 1) Open the Developer Tools of your browser and access the JS console.\n#", "_____no_output_____" ], [ "# 2) The following code allows to access the pixel data of the slice that\n# is currently displayed.\n#\n# element = cornerstone.getEnabledElements()[0];\n# pixels = element.image.getPixelData();\n#", "_____no_output_____" ] ], [ [ "#### Question 4): What is the maximum pixel value of the current slice?\nHint: There are multiple ways of doing this according to\nhttps://medium.com/coding-at-dawn/the-fastest-way-to-find-minimum-and-maximum-values-in-an-array-in-javascript-2511115f8621", "_____no_output_____" ] ], [ [ "# TODO: YOUR ANSWER", "_____no_output_____" ] ], [ [ "```JavaScript\nelement = cornerstone.getEnabledElements()[0];\npixels = element.image.getPixelData();\n\nUint16Array(262144) [27, 24, 13, 6, 14, 28, 28, 24, 22, 16, 17, 22, 25, 27, 26, 25, 19, 18, 25, 29, 25, 19, 24, 27, 23, 18, 19, 20, 16, 20, 24, 22, 23, 20, 21, 22, 22, 20, 21, 22, 24, 24, 24, 27, 25, 19, 12, 11, 19, 28, 33, 35, 31, 19, 21, 29, 22, 17, 21, 21, 29, 33, 20, 12, 15, 20, 21, 27, 28, 18, 20, 23, 19, 20, 18, 20, 28, 30, 31, 36, 23, 10, 18, 22, 20, 24, 25, 22, 25, 26, 25, 25, 24, 23, 25, 27, 21, 18, 18, 23, …]\n\npixels.length\n262144\n\nmax = 0;\n\nfor(i=0;i < pixels.length; i++){}\n\nfor(i=0;i < pixels.length; i++){if (max < pixels[i]){max = pixels[i]}}\n\nmax\n2585\n```", "_____no_output_____" ], [ "The Maximum pixel value is **2585**.", "_____no_output_____" ], [ "![p5](https://github.com/Yiming-S/cs480student/blob/main/07/5.png?raw=true)", "_____no_output_____" ], [ "## **Bonus**: Apply a filter mask to the current slice! [33 Points]\n", "_____no_output_____" ] ], [ [ "# Here we will modify the pixel data!\n#\n# The following code allows you to set the pixels of the current slice.\n#\n# pixels.set(new_pixels);\n#\n# Write code to APPLY A FILTER MASK to the pixels.\n#\n# Hint: You can use the following code to get the dimensions of the current slice.\n# w = element.image.width;\n# h = element.image.height;\n#\n# You can decide which filter mask to apply (e.g., Gaussian blur, edge detection..).\n#\n# ** IMPORTANT **\n# Hint 2: Triggering a REDRAW programmatically might be hard but you can use\n# the LEVELS tool to manually trigger a redraw to see filtered pixels.\n# ** IMPORTANT **", "_____no_output_____" ], [ "#\n#\n# TODO: YOUR CODE GOES HERE\n#\n#", "_____no_output_____" ] ], [ [ "```JavaScript \nfunction convert2d(array, width, height) {\n const newArr = [];\n for (let i = 0; i < height; ++i) \n newArr.push(array.slice(i*width, i*width + width))\n return newArr;}\n\nfunction uniform(len, value) {\n let arr = new Array(len); \n for (let i=0; i<len; ++i) \n arr[i] = Array.isArray(value) ? [...value] : value;\n return arr;}\n\nfunction conv_2d(kernel, array){\n var result = uniform(array.length, uniform(array[0].length, 0));\n var kRows = kernel.length;\n var kCols = kernel[0].length;\n var rows = array.length;\n var cols = array[0].length;\n var kCenterX = Math.floor(kCols/2);\n var kCenterY = Math.floor(kRows/2);\n var i, j, m, n, ii, jj;\n for(i=0; i < rows; ++i){ \n for(j=0; j < cols; ++j){ \n for(m=0; m < kRows; ++m){ \n for(n=0; n < kCols; ++n){\n ii = i + (m - kCenterY);\n jj = j + (n - kCenterX);\n if(ii >= 0 \n && ii < rows \n && jj >= 0 \n && jj < cols){\n result[i][j] += array[ii][jj] * kernel[m][n];};};};};};\n return result;}\n\nfunction sobelFilter(image) {\n const si = [[1,2,1],[0,0,0],[-1,-2,-1]];\n const sj = [[1,0,-1],[2,0,-2],[1,0,-1]];\n const new_i = conv_2d(si, image);\n const new_j = conv_2d(sj, image);\n var result = uniform(image.length, uniform(image[0].length, 0));\n var rows = image.length;\n var cols = image[0].length;\n for (let i = 0; i < rows; i++) {\n for (let j = 0; j < cols; j++) {\n result[i][j] = Math.sqrt(\n Math.pow(new_i[i][j], 2) \n + Math.pow(new_j[i][j], 2))}}\n return result;}\n\n\nelement = cornerstone.getEnabledElements()[0]\nw = element.image.width\nh = element.image.height\npixels = element.image.getPixelData()\nimage = convert2d(pixels, w, h)\nfiltered = sobelFilter(image)\nnew_pixels = [].concat(...filtered)\npixels.set(new_pixels)\n\n```", "_____no_output_____" ] ], [ [ "# TODO: Please add a screenshot after filtering the current slice.", "_____no_output_____" ] ], [ [ "TODO: ADD SCREENSHOT", "_____no_output_____" ], [ "![p7](https://github.com/Yiming-S/cs480student/blob/main/07/7.png?raw=true)", "_____no_output_____" ] ], [ [ "#\n# SUPER BONUS (+33 extra): Can you trigger the redraw programmatically?\n#", "_____no_output_____" ], [ "# TODO: YOUR CODE FOR REDRAW GOES HERE", "_____no_output_____" ], [ "#\n# THANK YOU!!!\n#\n# .--..--..--..--..--..--.\n# .' \\ (`._ (_) _ \\\n# .' | '._) (_) |\n# \\ _.')\\ .----..---. /\n# |(_.' | / .-\\-. \\ |\n# \\ 0| | ( O| O) | o|\n# | _ | .--.____.'._.-. |\n# \\ (_) | o -` .-` |\n# | \\ |`-._ _ _ _ _\\ /\n# \\ | | `. |_||_| |\n# | o | \\_ \\ | -. .-.\n# |.-. \\ `--..-' O | `.`-' .'\n# _.' .' | `-.-' /-.__ ' .-'\n# .' `-.` '.|='=.='=.='=.='=|._/_ `-'.'\n# `-._ `. |________/\\_____| `-.'\n# .' ).| '=' '='\\/ '=' |\n# `._.` '---------------'\n# //___\\ //___\\\n# || ||\n# LGB ||_.-. ||_.-.\n# (_.--__) (_.--__)\n#\n#", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code", "code", "code" ] ]

4ad62e358897c0a951b5e5bcf3af13927eeef907

894,836

ipynb

Jupyter Notebook

code/Rover_Project_Test_Notebook.ipynb

Aslanfmh65/NASA_autonomous_rover

4e8b7b711bbe2f5eea8503954dca440aad721787

[ "MIT" ]

null

code/Rover_Project_Test_Notebook.ipynb

Aslanfmh65/NASA_autonomous_rover

4e8b7b711bbe2f5eea8503954dca440aad721787

[ "MIT" ]

null

code/Rover_Project_Test_Notebook.ipynb

Aslanfmh65/NASA_autonomous_rover

4e8b7b711bbe2f5eea8503954dca440aad721787

[ "MIT" ]

null

1,099.307125

498,732

0.954125

[ [ [ "## Rover Project Test Notebook\nThis notebook contains the functions from the lesson and provides the scaffolding you need to test out your mapping methods. The steps you need to complete in this notebook for the project are the following:\n\n* First just run each of the cells in the notebook, examine the code and the results of each.\n* Run the simulator in \"Training Mode\" and record some data. Note: the simulator may crash if you try to record a large (longer than a few minutes) dataset, but you don't need a ton of data, just some example images to work with. \n* Change the data directory path (2 cells below) to be the directory where you saved data\n* Test out the functions provided on your data\n* Write new functions (or modify existing ones) to report and map out detections of obstacles and rock samples (yellow rocks)\n* Populate the `process_image()` function with the appropriate steps/functions to go from a raw image to a worldmap.\n* Run the cell that calls `process_image()` using `moviepy` functions to create video output\n* Once you have mapping working, move on to modifying `perception.py` and `decision.py` to allow your rover to navigate and map in autonomous mode!\n\n**Note: If, at any point, you encounter frozen display windows or other confounding issues, you can always start again with a clean slate by going to the \"Kernel\" menu above and selecting \"Restart & Clear Output\".**\n\n**Run the next cell to get code highlighting in the markdown cells.**", "_____no_output_____" ] ], [ [ "%%HTML\n<style> code {background-color : orange !important;} </style>", "_____no_output_____" ], [ "import sys\nprint(sys.executable)", "/anaconda2/envs/py36-test/bin/python3\n" ], [ "%matplotlib inline\n#%matplotlib qt # Choose %matplotlib qt to plot to an interactive window (note it may show up behind your browser)\n# Make some of the relevant imports\nimport cv2 # OpenCV for perspective transform\nimport numpy as np\nimport matplotlib.image as mpimg\nimport matplotlib.pyplot as plt\nimport scipy.misc # For saving images as needed\nimport glob # For reading in a list of images from a folder\nimport imageio\nimageio.plugins.ffmpeg.download()\n", "_____no_output_____" ] ], [ [ "## Quick Look at the Data\nThere's some example data provided in the `test_dataset` folder. This basic dataset is enough to get you up and running but if you want to hone your methods more carefully you should record some data of your own to sample various scenarios in the simulator. \n\nNext, read in and display a random image from the `test_dataset` folder", "_____no_output_____" ] ], [ [ "path = '../training_dataset/IGM/IMG/*'\npath = '/Users/aslanfeng/Documents/Lifelong Learning/robotics/RoboND-Rover-Project-master/training_dataset/IGM/IMG/*'\nimg_list = glob.glob(path)\n# Grab a random image and display it\nidx = np.random.randint(0, len(img_list)-1)\nimage = mpimg.imread(img_list[idx])\nplt.imshow(image)", "_____no_output_____" ] ], [ [ "## Calibration Data\nRead in and display example grid and rock sample calibration images. You'll use the grid for perspective transform and the rock image for creating a new color selection that identifies these samples of interest. ", "_____no_output_____" ] ], [ [ "# In the simulator you can toggle on a grid on the ground for calibration\n# You can also toggle on the rock samples with the 0 (zero) key. \n# Here's an example of the grid and one of the rocks\nexample_grid = '../calibration_images/example_grid1.jpg'\nexample_rock = '../calibration_images/example_rock1.jpg'\ngrid_img = mpimg.imread(example_grid)\nrock_img = mpimg.imread(example_rock)\n\nfig = plt.figure(figsize=(12,3))\nplt.subplot(121)\nplt.imshow(grid_img)\nplt.subplot(122)\nplt.imshow(rock_img)", "_____no_output_____" ] ], [ [ "## Perspective Transform\n\nDefine the perspective transform function from the lesson and test it on an image.", "_____no_output_____" ] ], [ [ "# Define a function to perform a perspective transform\n# I've used the example grid image above to choose source points for the\n# grid cell in front of the rover (each grid cell is 1 square meter in the sim)\n# Define a function to perform a perspective transform\ndef perspect_transform(img, src, dst):\n \n M = cv2.getPerspectiveTransform(src, dst)\n warped = cv2.warpPerspective(img, M, (img.shape[1], img.shape[0]))# keep same size as input image\n mask = cv2.warpPerspective(np.ones_like(img[:,:,0]), M, (img.shape[1], img.shape[0]))\n \n return warped, mask\n\n\n# Define calibration box in source (actual) and destination (desired) coordinates\n# These source and destination points are defined to warp the image\n# to a grid where each 10x10 pixel square represents 1 square meter\n# The destination box will be 2*dst_size on each side\ndst_size = 5 \n# Set a bottom offset to account for the fact that the bottom of the image \n# is not the position of the rover but a bit in front of it\n# this is just a rough guess, feel free to change it!\nbottom_offset = 6\nsource = np.float32([[14, 140], [301 ,140],[200, 96], [118, 96]])\ndestination = np.float32([[image.shape[1]/2 - dst_size, image.shape[0] - bottom_offset],\n [image.shape[1]/2 + dst_size, image.shape[0] - bottom_offset],\n [image.shape[1]/2 + dst_size, image.shape[0] - 2*dst_size - bottom_offset], \n [image.shape[1]/2 - dst_size, image.shape[0] - 2*dst_size - bottom_offset],\n ])\nwarped, mask = perspect_transform(grid_img, source, destination)\nplt.imshow(mask, cmap='gray')\n#scipy.misc.imsave('../output/warped_example.jpg', warped)", "_____no_output_____" ] ], [ [ "## Color Thresholding\nDefine the color thresholding function from the lesson and apply it to the warped image\n\n**TODO:** Ultimately, you want your map to not just include navigable terrain but also obstacles and the positions of the rock samples you're searching for. Modify this function or write a new function that returns the pixel locations of obstacles (areas below the threshold) and rock samples (yellow rocks in calibration images), such that you can map these areas into world coordinates as well. \n**Hints and Suggestion:** \n* For obstacles you can just invert your color selection that you used to detect ground pixels, i.e., if you've decided that everything above the threshold is navigable terrain, then everthing below the threshold must be an obstacle!\n\n\n* For rocks, think about imposing a lower and upper boundary in your color selection to be more specific about choosing colors. You can investigate the colors of the rocks (the RGB pixel values) in an interactive matplotlib window to get a feel for the appropriate threshold range (keep in mind you may want different ranges for each of R, G and B!). Feel free to get creative and even bring in functions from other libraries. Here's an example of [color selection](http://opencv-python-tutroals.readthedocs.io/en/latest/py_tutorials/py_imgproc/py_colorspaces/py_colorspaces.html) using OpenCV. \n\n* **Beware However:** if you start manipulating images with OpenCV, keep in mind that it defaults to `BGR` instead of `RGB` color space when reading/writing images, so things can get confusing.", "_____no_output_____" ] ], [ [ "# Identify pixels above the threshold\n# Threshold of RGB > 160 does a nice job of identifying ground pixels only\ndef color_thresh(img, rgb_thresh=(160, 160, 160)):\n # Create an array of zeros same xy size as img, but single channel\n color_select = np.zeros_like(img[:,:,0])\n # Require that each pixel be above all three threshold values in RGB\n # above_thresh will now contain a boolean array with \"True\"\n # where threshold was met\n above_thresh = (img[:,:,0] > rgb_thresh[0]) \\\n & (img[:,:,1] > rgb_thresh[1]) \\\n & (img[:,:,2] > rgb_thresh[2])\n # Index the array of zeros with the boolean array and set to 1\n color_select[above_thresh] = 1\n # Return the binary image\n return color_select\n\nthreshed = color_thresh(warped)\nplt.imshow(threshed, cmap='gray')\n#scipy.misc.imsave('../output/warped_threshed.jpg', threshed*255)", "_____no_output_____" ] ], [ [ "## Coordinate Transformations\nDefine the functions used to do coordinate transforms and apply them to an image.", "_____no_output_____" ] ], [ [ "# Define a function to convert from image coords to rover coords\ndef rover_coords(binary_img):\n # Identify nonzero pixels\n ypos, xpos = binary_img.nonzero()\n # Calculate pixel positions with reference to the rover position being at the \n # center bottom of the image. \n x_pixel = -(ypos - binary_img.shape[0]).astype(np.float)\n y_pixel = -(xpos - binary_img.shape[1]/2 ).astype(np.float)\n return x_pixel, y_pixel\n\n# Define a function to convert to radial coords in rover space\ndef to_polar_coords(x_pixel, y_pixel):\n # Convert (x_pixel, y_pixel) to (distance, angle) \n # in polar coordinates in rover space\n # Calculate distance to each pixel\n dist = np.sqrt(x_pixel**2 + y_pixel**2)\n # Calculate angle away from vertical for each pixel\n angles = np.arctan2(y_pixel, x_pixel)\n return dist, angles\n\n# Define a function to map rover space pixels to world space\ndef rotate_pix(xpix, ypix, yaw):\n # Convert yaw to radians\n yaw_rad = yaw * np.pi / 180\n xpix_rotated = (xpix * np.cos(yaw_rad)) - (ypix * np.sin(yaw_rad))\n \n ypix_rotated = (xpix * np.sin(yaw_rad)) + (ypix * np.cos(yaw_rad))\n # Return the result \n return xpix_rotated, ypix_rotated\n\ndef translate_pix(xpix_rot, ypix_rot, xpos, ypos, scale): \n # Apply a scaling and a translation\n xpix_translated = (xpix_rot / scale) + xpos\n ypix_translated = (ypix_rot / scale) + ypos\n # Return the result \n return xpix_translated, ypix_translated\n\n\n# Define a function to apply rotation and translation (and clipping)\n# Once you define the two functions above this function should work\ndef pix_to_world(xpix, ypix, xpos, ypos, yaw, world_size, scale):\n # Apply rotation\n xpix_rot, ypix_rot = rotate_pix(xpix, ypix, yaw)\n # Apply translation\n xpix_tran, ypix_tran = translate_pix(xpix_rot, ypix_rot, xpos, ypos, scale)\n # Perform rotation, translation and clipping all at once\n x_pix_world = np.clip(np.int_(xpix_tran), 0, world_size - 1)\n y_pix_world = np.clip(np.int_(ypix_tran), 0, world_size - 1)\n # Return the result\n return x_pix_world, y_pix_world\n\n# Grab another random image\nidx = np.random.randint(0, len(img_list)-1)\nimage = mpimg.imread(img_list[idx])\nwarped, mask = perspect_transform(image, source, destination)\nthreshed = color_thresh(warped)\n\n# Calculate pixel values in rover-centric coords and distance/angle to all pixels\nxpix, ypix = rover_coords(threshed)\ndist, angles = to_polar_coords(xpix, ypix)\nmean_dir = np.mean(angles)\n\n# Do some plotting\nfig = plt.figure(figsize=(12,9))\nplt.subplot(221)\nplt.imshow(image)\nplt.subplot(222)\nplt.imshow(warped)\nplt.subplot(223)\nplt.imshow(threshed, cmap='gray')\nplt.subplot(224)\nplt.plot(xpix, ypix, '.')\nplt.ylim(-160, 160)\nplt.xlim(0, 160)\narrow_length = 100\nx_arrow = arrow_length * np.cos(mean_dir)\ny_arrow = arrow_length * np.sin(mean_dir)\nplt.arrow(0, 0, x_arrow, y_arrow, color='red', zorder=2, head_width=10, width=2)\n\n", "_____no_output_____" ], [ "def find_rocks(img, rgb_thresh=(110, 110, 50)):\n color_select = np.zeros_like(img[:,:,0])\n rockpixel = (img[:,:,0] > rgb_thresh[0]) \\\n & (img[:,:,1] > rgb_thresh[1]) \\\n & (img[:,:,2] < rgb_thresh[2])\n color_select[rockpixel] = 1\n \n return color_select\n\nrock_map = find_rocks(rock_img)\nfig = plt.figure(figsize=(12, 3))\nplt.subplot(121)\nplt.imshow(rock_img)\nplt.subplot(122)\nplt.imshow(rock_map, cmap='gray')", "_____no_output_____" ] ], [ [ "## Read in saved data and ground truth map of the world\nThe next cell is all setup to read your saved data into a `pandas` dataframe. Here you'll also read in a \"ground truth\" map of the world, where white pixels (pixel value = 1) represent navigable terrain. \n\nAfter that, we'll define a class to store telemetry data and pathnames to images. When you instantiate this class (`data = Databucket()`) you'll have a global variable called `data` that you can refer to for telemetry and map data within the `process_image()` function in the following cell. \n", "_____no_output_____" ] ], [ [ "# Import pandas and read in csv file as a dataframe\nimport pandas as pd\n# Change the path below to your data directory\n# If you are in a locale (e.g., Europe) that uses ',' as the decimal separator\n# change the '.' to ','\npath = '/Users/aslanfeng/Documents/Lifelong Learning/robotics/RoboND-Rover-Project-master/training_dataset/robot_log.csv'\ndf = pd.read_csv('../training_dataset/robot_log.csv', delimiter=';', decimal='.')\ndf = pd.read_csv(path, delimiter=';',decimal='.')\ncsv_img_list = df[\"Path\"].tolist() # Create list of image pathnames\n# Read in ground truth map and create a 3-channel image with it\nground_truth = mpimg.imread('../calibration_images/map_bw.png')\nground_truth_3d = np.dstack((ground_truth*0, ground_truth*255, ground_truth*0)).astype(np.float)\n\n# Creating a class to be the data container\n# Will read in saved data from csv file and populate this object\n# Worldmap is instantiated as 200 x 200 grids corresponding \n# to a 200m x 200m space (same size as the ground truth map: 200 x 200 pixels)\n# This encompasses the full range of output position values in x and y from the sim\nclass Databucket():\n def __init__(self):\n self.images = csv_img_list \n self.xpos = df[\"X_Position\"].values\n self.ypos = df[\"Y_Position\"].values\n self.yaw = df[\"Yaw\"].values\n self.count = 0 # This will be a running index\n self.worldmap = np.zeros((200, 200, 3)).astype(np.float)\n self.ground_truth = ground_truth_3d # Ground truth worldmap\n\n# Instantiate a Databucket().. this will be a global variable/object\n# that you can refer to in the process_image() function below\ndata = Databucket()\n", "_____no_output_____" ] ], [ [ "## Write a function to process stored images\n\nModify the `process_image()` function below by adding in the perception step processes (functions defined above) to perform image analysis and mapping. The following cell is all set up to use this `process_image()` function in conjunction with the `moviepy` video processing package to create a video from the images you saved taking data in the simulator. \n\nIn short, you will be passing individual images into `process_image()` and building up an image called `output_image` that will be stored as one frame of video. You can make a mosaic of the various steps of your analysis process and add text as you like (example provided below). \n\n\n\nTo start with, you can simply run the next three cells to see what happens, but then go ahead and modify them such that the output video demonstrates your mapping process. Feel free to get creative!", "_____no_output_____" ] ], [ [ "\n# Define a function to pass stored images to\n# reading rover position and yaw angle from csv file\n# This function will be used by moviepy to create an output video\ndef process_image(img):\n # Example of how to use the Databucket() object defined above\n # to print the current x, y and yaw values \n # print(data.xpos[data.count], data.ypos[data.count], data.yaw[data.count])\n\n # TODO: \n # 1) Define source and destination points for perspective transform\n # 2) Apply perspective transform\n # 3) Apply color threshold to identify navigable terrain/obstacles/rock samples\n # 4) Convert thresholded image pixel values to rover-centric coords\n # 5) Convert rover-centric pixel values to world coords\n # 6) Update worldmap (to be displayed on right side of screen)\n # Example: data.worldmap[obstacle_y_world, obstacle_x_world, 0] += 1\n # data.worldmap[rock_y_world, rock_x_world, 1] += 1\n # data.worldmap[navigable_y_world, navigable_x_world, 2] += 1\n \n warped, mask = perspect_transform(img, source, destination)\n threshed = color_thresh(warped)\n obs_map = np.absolute(np.float32(threshed) - 1) * mask\n xpix, ypix = rover_coords(threshed)\n world_size = data.worldmap.shape[0]\n scale = 2 * dst_size\n xpos = data.xpos[data.count]\n ypos = data.ypos[data.count]\n yaw = data.yaw[data.count]\n x_world, y_world = pix_to_world(xpix, ypix, xpos, ypos, yaw, world_size, scale)\n obsxpix, obsypix = rover_coords(obs_map)\n obs_x_world, obs_y_world = pix_to_world(obsxpix, obsypix, xpos, ypos, yaw, world_size, scale)\n \n data.worldmap[y_world, x_world, 2] = 255\n data.worldmap[obs_y_world, obs_x_world, 0] = 255\n nav_pix = data.worldmap[:,:,2] > 0\n data.worldmap[nav_pix,0] = 0\n \n rock_map = find_rocks(warped)\n \n if rock_map.any():\n rock_x, rock_y = rover_coords(rock_map)\n rock_x_world, rock_y_world = pix_to_world(rock_x, rock_y, xpos, ypos, yaw, world_size, scale)\n data.worldmap[rock_y_world, rock_x_world,:] = 255\n \n\n # 7) Make a mosaic image, below is some example code\n # First create a blank image (can be whatever shape you like)\n output_image = np.zeros((img.shape[0] + data.worldmap.shape[0], img.shape[1]*2, 3))\n # Next you can populate regions of the image with various output\n # Here I'm putting the original image in the upper left hand corner\n output_image[0:img.shape[0], 0:img.shape[1]] = img\n\n # Let's create more images to add to the mosaic, first a warped image\n # Add the warped image in the upper right hand corner\n output_image[0:img.shape[0], img.shape[1]:] = warped\n\n # Overlay worldmap with ground truth map\n map_add = cv2.addWeighted(data.worldmap, 1, data.ground_truth, 0.5, 0)\n # Flip map overlay so y-axis points upward and add to output_image \n output_image[img.shape[0]:, 0:data.worldmap.shape[1]] = np.flipud(map_add)\n\n\n # Then putting some text over the image\n cv2.putText(output_image,\"Populate this image with your analyses to make a video!\", (20, 20), \n cv2.FONT_HERSHEY_COMPLEX, 0.4, (255, 255, 255), 1)\n if data.count < len(data.images) - 1:\n data.count += 1 # Keep track of the index in the Databucket()\n \n return output_image", "_____no_output_____" ] ], [ [ "## Make a video from processed image data\nUse the [moviepy](https://zulko.github.io/moviepy/) library to process images and create a video.\n ", "_____no_output_____" ] ], [ [ "# Import everything needed to edit/save/watch video clips\nfrom moviepy.editor import VideoFileClip\nfrom moviepy.editor import ImageSequenceClip\n\n\n# Define pathname to save the output video\noutput = '../output/test_mapping.mp4'\ndata = Databucket() # Re-initialize data in case you're running this cell multiple times\nclip = ImageSequenceClip(data.images, fps=60) # Note: output video will be sped up because \n # recording rate in simulator is fps=25\nnew_clip = clip.fl_image(process_image) #NOTE: this function expects color images!!\n%time new_clip.write_videofile(output, audio=False)", "[MoviePy] >>>> Building video ../output/test_mapping.mp4\n[MoviePy] Writing video ../output/test_mapping.mp4\n" ] ], [ [ "### This next cell should function as an inline video player\nIf this fails to render the video, try running the following cell (alternative video rendering method). You can also simply have a look at the saved mp4 in your `/output` folder", "_____no_output_____" ] ], [ [ "\nfrom IPython.display import HTML\nHTML(\"\"\"\n<video width=\"960\" height=\"540\" controls>\n <source src=\"{0}\">\n</video>\n\"\"\".format(output))", "_____no_output_____" ] ], [ [ "### Below is an alternative way to create a video in case the above cell did not work.", "_____no_output_____" ] ], [ [ "import io\nimport base64\nvideo = io.open(output, 'r+b').read()\nencoded_video = base64.b64encode(video)\nHTML(data='''<video alt=\"test\" controls>\n <source src=\"data:video/mp4;base64,{0}\" type=\"video/mp4\" />\n </video>'''.format(encoded_video.decode('ascii')))", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad63703057d5199f3b98250157f48a3713fda42

2,256

ipynb

Jupyter Notebook

Trail of 0's in factorial.ipynb

kj2411-2411/Python

b1c6e28faca9f219261702bf9aef5e2f6613359a

[ "CC0-1.0" ]

null

Trail of 0's in factorial.ipynb

kj2411-2411/Python

b1c6e28faca9f219261702bf9aef5e2f6613359a

[ "CC0-1.0" ]

null

Trail of 0's in factorial.ipynb

kj2411-2411/Python

b1c6e28faca9f219261702bf9aef5e2f6613359a

[ "CC0-1.0" ]

null

20.509091

106

0.416223

[ [ [ "### Find and return number of trailing 0s in n factorial without calculating n factorial. Approach 1", "_____no_output_____" ] ], [ [ "N = int(input())\ncount = 0\ncount1 = 0\n\ndef trailingzeros(N):\n for i in range(N+1):\n if (i % 5 == 0):\n if (i >0):\n global count\n count = count+1\n \n j = i/5\n# print(j)\n if (j%5 ==0):\n global count1\n count1=count1 +1\n# print(\"new\",j)\n\n \n\n\n \ntrailingzeros(N)\nprint(count+count1)\n", "50\n12\n" ] ], [ [ "### Find and return number of trailing 0s in n factorial without calculating n factorial. Approach 1", "_____no_output_____" ] ], [ [ "def findTrailingZeros(n): \t\n\tcount = 0\t\n\ti=5\n\twhile (n/i>=1): \n\t\tcount += int(n/i) \n\t\ti *= 5\n\n\treturn int(count) \n\nn =int(input(\"\"))\nprint(findTrailingZeros(n)) \n ", "100\n24\n" ] ] ]

[ "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad63a756e2f98fbcc6c5bf8598fd3e6a718d18b

406,552

ipynb

Jupyter Notebook

notebooks/01_single/01_classifiers/01_binary/01_FeedForward.ipynb

t0kk35/d373c7

7780b97545e581244fb4fb74347bb1b052b9ec3f

[ "Apache-2.0" ]

1

2021-07-23T18:04:55.000Z

notebooks/01_single/01_classifiers/01_binary/01_FeedForward.ipynb

t0kk35/d373c7

7780b97545e581244fb4fb74347bb1b052b9ec3f

[ "Apache-2.0" ]

null

notebooks/01_single/01_classifiers/01_binary/01_FeedForward.ipynb

t0kk35/d373c7

7780b97545e581244fb4fb74347bb1b052b9ec3f

[ "Apache-2.0" ]

null

295.459302

114,664

0.898343

[ [ [ "# Binary Classifier on Single records", "_____no_output_____" ], [ "### Most basic example.\nThis notebook will show how to set-up learning features (i.e. fields we want to use for modeling) and read them from a CSV file. Then create a very simple feed-forward Neural Net to classify fraud vs. non-fraud, train the model and test it.\n\nThroughout these notebooks we will not explain how Neural Nets work, which loss functions exists, how NN's are optimized, what stochastic gradient decent is etc... There are some excellent resources online which do this in great detail.\n\n#### The math\nOne thing we will quickly do is recap what a FeedForward (aka Linear Layer) is, just to build some intuition and to contrast this with other layers which will be explained later.\n\nLinear Layers are often presented as shown below on the left, basically as weights that connect one layer to the next. Each node in the second layer is a weighted sum of the input values to which a bias term is added. So for nodes $h_j$ for h=1->3; $h_j= (\\sum_{i=1}^4 x_i*w_{ij})+\\beta_{j}$. \n\nIf we look more in detail what happens on the right hand size we see that for a single node we effectively take the weighted sum, add a bias and then perform an activation function. The activation is needed so the model can learn non-linearity. If all NN's did was stacking linear operations, the end-result would be a linear combination of the input, ideally we would want models to learn more complex relations, non-linear activations enable that. In most DeepLearning cases the `ReLU` function is used for activation. \n\nTaking the weighted sum on a large scale for multiple nodes in one go, is nothing more or less than taking matrix dot-product. Same for adding the bias, that is just a matrix element wise addition. Another way of looking at this is considering the input __I__ to be a vector of size (1,4) and the hidden layer __H__ a vector of size (1,3), if we set-up a weight matrix __W__ of size (4,3), and a bias vector __$\\beta$__ of size (1,3) then $H = act(I \\odot W + \\beta)$. The dot product of the input array with the weight matrix plus a bias vector which is then 'activated'\n\nThe hidden layer would typically be connected to a next layer, and a next and a next... until we get to an output layer. The formula for the output $O$ would thus be something like; $O = act(act(act(I \\odot W_1 + \\beta_1) \\odot W_2 + \\beta_2) \\odot W_3 + \\beta_3)$ if we had 3 hidden layers.", "_____no_output_____" ], [ "![01_FeedForward_Intro-2.png](attachment:01_FeedForward_Intro-2.png)", "_____no_output_____" ], [ "#### The Intuition\nWe can think of the __dot products and weights__ as taking bit of each input and combining that into a __hidden__ 'feature'. For instance if $X_1$ indicates 'age' category *elder* and $X_4$ indicates 'gender' *male*. If the inputs are 0 and 1 for yes/no, then using a positive weight $w_1$ and $w_4$ and 0 for the other weights, then we'd have a feature which 'activates' for an elder male customer.\n\nWe can think of the __bias__ as setting a minimal barrier or a reinforcement of the feature. For instance in above feature if we wanted our *elder male* hidden feature to only activate as it reaches .7 we could add a -.7 bias. If we wanted the feature to activate easier, we could add a positive bias.\n\nOur Neural Nets will learn which weights and biases work best in order to solve a specific task. In Feedforward NN's they are the *'learnable'* parameters", "_____no_output_____" ], [ "---\n#### Note on the data set \nThe data set used here is not particularly complex and/or big. It's not really all that challenging to find the fraud. In an ideal world we'd be using more complex data sets to show the real power of Deep Learning. There are a bunch of PCA'ed data sets available, but the PCA obfuscates some of the elements that are useful. \n*These examples are meant to show the possibilities, it's not so useful to interpret their performance on this data set*", "_____no_output_____" ], [ "## Imports", "_____no_output_____" ] ], [ [ "import torch\nimport numpy as np\nimport gc\n\nimport d373c7.features as ft\nimport d373c7.engines as en\nimport d373c7.pytorch as pt\nimport d373c7.pytorch.models as pm\nimport d373c7.plot as pl", "_____no_output_____" ] ], [ [ "## Set a random seed for Numpy and Torch\n> Will make sure we always sample in the same way. Makes it easier to compare results. At some point it should been removed to test the model stability.", "_____no_output_____" ] ], [ [ "# Numpy\nnp.random.seed(42)\n# Torch\ntorch.manual_seed(42)\ntorch.backends.cudnn.deterministic = True\ntorch.backends.cudnn.benchmark = False", "_____no_output_____" ] ], [ [ "## Define base feature and read the File\nThe base features are features found in the input file. \n\nBelow code snipped follows a structure we'll see a lot.\n- First we define features, these are fields we'll want to use. In this case we're defining 'Source' features, features that are in the data source, here the file. As parameters we provide the name as found in the first row of the file and a type.\n- Then we bundle them in a `TensorDefinition`, this is essentially a group of features. We give it a name and list of features as input.\n- Lastly we set up and engine of type `EnginePandasNumpy` and call the `from_csv` method with the TensorDefinition and the `file` name. This `from_csv` method will read the file and return a Pandas DataFrame object. The `inference` parameter specifies that we are in training mode, so any stats the feature needs to use will be gathered.\n\nAll this will return a Pandas DataFrame with __594643 rows__ (the number of transactions) and __6 columns__ (the number of features we defined), it can be use to perform basic data analysis.\n", "_____no_output_____" ] ], [ [ "# Change this to read from another location\nfile = '../../../../data/bs140513_032310.csv'", "_____no_output_____" ], [ "age = ft.FeatureSource('age', ft.FEATURE_TYPE_CATEGORICAL)\ngender = ft.FeatureSource('gender', ft.FEATURE_TYPE_CATEGORICAL)\nmerchant = ft.FeatureSource('merchant', ft.FEATURE_TYPE_CATEGORICAL)\ncategory = ft.FeatureSource('category', ft.FEATURE_TYPE_CATEGORICAL)\namount = ft.FeatureSource('amount', ft.FEATURE_TYPE_FLOAT)\nfraud = ft.FeatureSource('fraud', ft.FEATURE_TYPE_INT_8)\n\nbase_features = ft.TensorDefinition(\n 'base', \n [\n age,\n gender,\n merchant,\n category,\n amount,\n fraud\n ])\n\nwith en.EnginePandasNumpy() as e:\n df = e.from_csv(base_features, file, inference=False)\n \ndf", "2021-09-15 15:11:54.144 d373c7.engines.common INFO Start Engine...\n2021-09-15 15:11:54.145 d373c7.engines.panda_numpy INFO Pandas Version : 1.1.4\n2021-09-15 15:11:54.145 d373c7.engines.panda_numpy INFO Numpy Version : 1.19.2\n2021-09-15 15:11:54.145 d373c7.engines.panda_numpy INFO Building Panda for : base from file ../../../../data/bs140513_032310.csv\n2021-09-15 15:11:54.351 d373c7.engines.panda_numpy INFO Building Panda for : <Source_Derive_Source> from DataFrame. Inference mode <False>\n2021-09-15 15:11:54.352 d373c7.engines.panda_numpy INFO Reshaping DataFrame to: Source_Derive_Source\n2021-09-15 15:11:54.354 d373c7.engines.panda_numpy INFO Done creating Source_Derive_Source. Shape=(594643, 6)\n2021-09-15 15:11:54.355 d373c7.engines.panda_numpy INFO Reshaping DataFrame to: base\n" ] ], [ [ "## Define some derived features\nAfter we've defined and read the source features, we can define some __derived__ features. Derived features apply a form of transformation to the source features, depending on the type of feature.\n\nIn this example 3 transformations are used;\n- `FeatureNormalizeScale` The amount is scaled between 0 and 1.\n- `FeatureOneHot` The categorical features are turned into one-hot encoded fields. \n- `FeatureLabelBinary` The Fraud field is marked as Label. This is not really a transformation, it's just so the model knows which label to use.\n\nWe apply above transformations because Neural Nets prefer data that is in Binary ranges 0->1 or normally distributed\n\nThis will create a total of 78 features we can use in the model. We create a second list with the label.", "_____no_output_____" ] ], [ [ "amount_scale = ft.FeatureNormalizeScale('amount_scale', ft.FEATURE_TYPE_FLOAT_32, amount)\nage_oh = ft.FeatureOneHot('age_one_hot', ft.FEATURE_TYPE_INT_8, age)\ngender_oh = ft.FeatureOneHot('gender_one_hot', ft.FEATURE_TYPE_INT_8, gender)\nmerchant_oh = ft.FeatureOneHot('merchant_one_hot', ft.FEATURE_TYPE_INT_8, merchant)\ncategory_oh = ft.FeatureOneHot('category_one_hot', ft.FEATURE_TYPE_INT_8, category)\nfraud_label = ft.FeatureLabelBinary('fraud_label', ft.FEATURE_TYPE_INT_8, fraud)\n\nfeatures = ft.TensorDefinition(\n 'features', \n [\n age_oh,\n gender_oh,\n merchant_oh,\n category_oh,\n amount_scale,\n ])\n\nlabel = ft.TensorDefinition('label', [fraud_label])\n\nmodel_features = ft.TensorDefinitionMulti([features, label])\n\nwith en.EnginePandasNumpy() as e:\n ft = e.from_csv(features, file, inference=False)\n lb = e.from_csv(label, file, inference=False)", "2021-09-15 15:11:57.210 d373c7.engines.common INFO Start Engine...\n2021-09-15 15:11:57.210 d373c7.engines.panda_numpy INFO Pandas Version : 1.1.4\n2021-09-15 15:11:57.211 d373c7.engines.panda_numpy INFO Numpy Version : 1.19.2\n2021-09-15 15:11:57.211 d373c7.engines.panda_numpy INFO Building Panda for : features from file ../../../../data/bs140513_032310.csv\n2021-09-15 15:11:57.397 d373c7.engines.panda_numpy INFO Building Panda for : <Source_Derive_Source> from DataFrame. Inference mode <False>\n2021-09-15 15:11:57.397 d373c7.engines.panda_numpy INFO Reshaping DataFrame to: Source_Derive_Source\n2021-09-15 15:11:57.398 d373c7.engines.panda_numpy INFO Done creating Source_Derive_Source. Shape=(594643, 5)\n2021-09-15 15:11:57.400 d373c7.engines.panda_numpy INFO Create amount_scale Normalize/Scale amount. Min. 0.00 Max. 8329.96\n2021-09-15 15:11:57.462 d373c7.engines.panda_numpy INFO Reshaping DataFrame to: features\n2021-09-15 15:11:57.482 d373c7.engines.panda_numpy INFO Building Panda for : label from file ../../../../data/bs140513_032310.csv\n2021-09-15 15:11:57.621 d373c7.engines.panda_numpy INFO Building Panda for : <Source_Derive_Source> from DataFrame. Inference mode <False>\n2021-09-15 15:11:57.622 d373c7.engines.panda_numpy INFO Reshaping DataFrame to: Source_Derive_Source\n2021-09-15 15:11:57.623 d373c7.engines.panda_numpy INFO Done creating Source_Derive_Source. Shape=(594643, 1)\n2021-09-15 15:11:57.625 d373c7.engines.panda_numpy INFO Reshaping DataFrame to: label\n" ], [ "ft", "_____no_output_____" ], [ "lb", "_____no_output_____" ] ], [ [ "## Convert to Numpy\n\nNow we convert the panda DataFrame to a list of Numpy arrays (which can be used for training). The `NumpyList` will have an entry for each of the Learning types. It will split out the *Binary*, *Continuous*, *Categorical* and *Label* Learning type __features__. Each Learning type will have a list entry in the `NumpyList` object\n\nThis step is needed so the models understand how to use the various features in the learning and testing processes. \n\nIn this case we have a first list with 77 binary (one-hot-endoded) features, the second is a list with 1 continuous feature (the amount) and the last list is the label (Fraud or Non-Fraud).", "_____no_output_____" ] ], [ [ "with en.EnginePandasNumpy() as e:\n ft_np = e.to_numpy_list(features, ft)\n lb_np = e.to_numpy_list(label, lb)\n\ndata_list = en.NumpyList(ft_np.lists + lb_np.lists)\n \nprint(data_list.shapes)\nprint(data_list.dtype_names)", "2021-09-15 15:12:07.269 d373c7.engines.common INFO Start Engine...\n2021-09-15 15:12:07.269 d373c7.engines.panda_numpy INFO Pandas Version : 1.1.4\n2021-09-15 15:12:07.270 d373c7.engines.panda_numpy INFO Numpy Version : 1.19.2\n2021-09-15 15:12:07.270 d373c7.engines.panda_numpy INFO Converting DataFrame to Numpy of type: int8\n2021-09-15 15:12:07.271 d373c7.engines.panda_numpy INFO Reshaping DataFrame to: Binary\n2021-09-15 15:12:07.289 d373c7.engines.panda_numpy INFO Converting DataFrame to Numpy of type: float32\n2021-09-15 15:12:07.289 d373c7.engines.panda_numpy INFO Reshaping DataFrame to: Continuous\n2021-09-15 15:12:07.291 d373c7.engines.panda_numpy INFO Converting DataFrame to Numpy of type: int8\n2021-09-15 15:12:07.291 d373c7.engines.panda_numpy INFO Reshaping DataFrame to: Label\n" ] ], [ [ "## Wrangle the data\nTime to split the data. For time series data it is very important to keep the order of the data. Below split will start from the end and work it's way to the front of the data. Doing so the training, validation and test data are nicely colocated in time. You almost *never* want to plain shuffle time based data.\n\n> 1. Split out a test-set of size `test_records`. This is used for model testing.\n> 2. Split out a validation-set of size `validation_records`. It will be used to monitor overfitting during training\n> 3. All the rest is considered training data.\n\n__Important__; please make sure the data is ordered in ascending fashion on a date(time) field. The split function does not order the data, it assumes the data is in the correct order.", "_____no_output_____" ], [ "![01_SplitTime.png](attachment:01_SplitTime.png)", "_____no_output_____" ] ], [ [ "test_records = 100000\nval_records = 30000 \n\ntrain_data, val_data, test_data = data_list.split_time(val_records, test_records) \n\nprint(f'Training Data shapes {train_data.shapes}')\nprint(f'Validation Data shapes {val_data.shapes}')\nprint(f'Test Data shapes {test_data.shapes}')\ndel ft, lb\ndel ft_np, lb_np\ndel data_list\n\ngc.collect()\nprint('Done')", "Training Data shapes [(464643, 77), (464643,), (464643,)]\nValidation Data shapes [(30000, 77), (30000,), (30000,)]\nTest Data shapes [(100000, 77), (100000,), (100000,)]\nDone\n" ] ], [ [ "## Set-up devices", "_____no_output_____" ] ], [ [ "device, cpu = pt.init_devices()", "2021-09-15 15:12:12.072 d373c7.pytorch.common INFO Torch Version : 1.9.0+cu111\n2021-09-15 15:12:12.144 d373c7.pytorch.common INFO GPU found. Using GPU <0>\n2021-09-15 15:12:12.144 d373c7.pytorch.common INFO Cuda Version : 11.1\n" ] ], [ [ "# Define Model\n\n\nThe training data set has to be balanced for Neural Nets. A balanced data set has a more or less equal amount of each class of the label. In our case fraud vs. non-fraud classes. Failing to balance the data can have dramatic results, Neural Nets are lazy, in extreme cases they might just plain always predict the majority class.\n\nFraud data-sets always have more non-fraud than fraud records. In this example the fraud class will be aggressively upsampled in the training phase by a custom `ClassSampler`. It oversamples the minority label until it matches the majority label in quantity. This may not be a good idea for a really large data sets.\n\n> 1. First set-up a NumpyListDataSet for both the training data-set and validation data-set. A NumpyListDataSet is a specialized `Pytorch Dataset` which keeps the data as numpy arrays in memory and converts on the fly to `Pytorch Tensors`\n> 2. Set-up a sampler for the training set only. The sampler will over-sample the '1'/fraud class. Note that this means the training and validation sets are balanced *differently*. This is important when interpreting the plots.\n> 3. Wrap the dataset in a Pytorch Dataloader. `Dataloaders` allow the training loop to iterate over `Datasets`\n> 4. Create a model. Here the most basic __GeneratedClassifier__ is used. __The GeneratedClassifier__ will create a model using the information it has about the features. *We are defining it to have 1 hidden layer of size 16*. ", "_____no_output_____" ] ], [ [ "# Setup Pytorch Datasets for the training and validation\nbatch_size = 128\ntrain_ds = pt.NumpyListDataSetMulti(model_features, train_data)\nval_ds = pt.NumpyListDataSetMulti(model_features, val_data)\ntrain_sampler = pt.ClassSamplerMulti(model_features, train_data).over_sampler()\n\n# Wrap them in a Pytorch Dataloader\ntrain_dl = train_ds.data_loader(cpu, batch_size, num_workers=2, sampler=train_sampler)\nval_dl = val_ds.data_loader(cpu, batch_size, num_workers=2)\n\n# Create a Model\nm = pm.GeneratedClassifier(model_features, linear_layers=[16])\nprint(m)", "GeneratedClassifier(\n Number of parameters : 1313. Loss : SingleLabelBCELoss, mean\n (streams): ModuleList(\n (0): TensorDefinitionHead(lcs=['Binary', 'Continuous'])\n )\n (tail): TailBinary(\n (layers): Sequential(\n (tail_lin_01): Linear(in_features=78, out_features=16, bias=True)\n (tail_act_01): ReLU()\n (tail_dropout_01): Dropout(p=0.1, inplace=False)\n (tail_batch_norm): BatchNorm1d(16, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)\n (tail_binary): Linear(in_features=16, out_features=1, bias=True)\n (tail_bin_act): Sigmoid()\n )\n )\n)\n" ] ], [ [ "The generated model consists of one stream -as there is one TensorDefinition containing modeling features-. That stream has a later of type __TensorDefinitionHead__ which will in this case just concatenate our *Binary* (77 one hot) and our Continuous (amount) features into a 78 shape tensor.\n\nWhich is consequently processed through the __Tail__. As we have a binary label, the tail is binary, it will process using the requested 16 size hidden layer and then output a final output layer of size 1. The output will be a score, ranging from 0 to 1 (because of the sigmoid), indicating the likelyhood of Fraud. The higher, the more certain the model is it bad.\n\nGrapically this network looks like below (Some of the layers have been omitted for simplicity)", "_____no_output_____" ], [ "![01_FeedForward.png](attachment:01_FeedForward.png)", "_____no_output_____" ], [ "# Start Training", "_____no_output_____" ], [ "### First find a decent Learning Rate. \n> Create a trainer and run the find_lr function and plot. This function iterates over the batches, gradually increasing the learning rate from a minimum to a maximum learning rate. It tends to show where we can find a good learning rate. In this case at around __3e-3__ we start a very steep decent. The model does not learn at lower than __1e-3__ learning rates. Beyond __1e-2__ it flattens out, stops learning at around __2e-2__ and explodes later. This exloding can be validate by running with a higher number of iteration and higher upper bound. A good learning rate is a location where the curve has a steep descent, but not too far down the curve. In this case around __5e-3__", "_____no_output_____" ] ], [ [ "t = pt.Trainer(m, device, train_dl, val_dl)\nr = t.find_lr(1e-4, 1e-1, 200)\npl.TrainPlot().plot_lr(r)", "2021-09-15 15:12:16.750 d373c7.pytorch.training INFO Saving model under ./temp_model.pt\nFinding LR in 200 steps: 100%|██████████| 200/200 [00:00<00:00, 488.25it/s]\n2021-09-15 15:12:19.506 d373c7.pytorch.training INFO Restoring model from ./temp_model.pt\n" ] ], [ [ "## Start Training and plot the results\nIn our examples we will use a one_cycle logic. This is a training logic which starts at a learning rate lower than the specified learning rate, over the course of training works its way up to the specified learning rate and decreases again towards the end of the learning cycle. Proposed by [Leslie N. Smith, A DISCIPLINED APPROACH TO NEURAL NETWORK HYPER-PARAMETERS](https://arxiv.org/pdf/1803.09820.pdf)\n\n> We train for __10 epochs__ and __learning rate 5e-3__. That means we run over the total training data set a total of 10 times/epochs where the model learns, after each epoch we use the trained model and perform a test run on the validation set.\n> The result graph plots the accuracy and loss evolution at each Epoch for both the training and the validation set. We see the model behaves fairly well during training. The loss goes up slightly in the middle of the training. This is the one_cycle logic which is reaching the max learning rate. ", "_____no_output_____" ] ], [ [ "t = pt.Trainer(m, device, train_dl, val_dl)\nh = t.train_one_cycle(10, 5e-3)\npl.TrainPlot().plot_history(h, fig_size=(10,10))", "Epoch 001/010: 100%|██████████| 3866/3866 [00:04<00:00, 784.66it/s, train_loss=0.126, train_acc=0.958, val_loss=0.035, val_acc=0.985]\nEpoch 002/010: 100%|██████████| 3866/3866 [00:04<00:00, 795.34it/s, train_loss=0.0673, train_acc=0.974, val_loss=0.0533, val_acc=0.975]\nEpoch 003/010: 100%|██████████| 3866/3866 [00:04<00:00, 827.62it/s, train_loss=0.0647, train_acc=0.975, val_loss=0.0475, val_acc=0.977]\nEpoch 004/010: 100%|██████████| 3866/3866 [00:05<00:00, 768.75it/s, train_loss=0.062, train_acc=0.976, val_loss=0.0402, val_acc=0.98]\nEpoch 005/010: 100%|██████████| 3866/3866 [00:04<00:00, 800.10it/s, train_loss=0.0617, train_acc=0.976, val_loss=0.0482, val_acc=0.976]\nEpoch 006/010: 100%|██████████| 3866/3866 [00:04<00:00, 783.87it/s, train_loss=0.06, train_acc=0.976, val_loss=0.0347, val_acc=0.984]\nEpoch 007/010: 100%|██████████| 3866/3866 [00:04<00:00, 790.58it/s, train_loss=0.0585, train_acc=0.977, val_loss=0.0386, val_acc=0.98]\nEpoch 008/010: 100%|██████████| 3866/3866 [00:04<00:00, 826.62it/s, train_loss=0.0574, train_acc=0.977, val_loss=0.0378, val_acc=0.982]\nEpoch 009/010: 100%|██████████| 3866/3866 [00:04<00:00, 808.95it/s, train_loss=0.0562, train_acc=0.977, val_loss=0.0323, val_acc=0.985]\nEpoch 010/010: 100%|██████████| 3866/3866 [00:04<00:00, 792.10it/s, train_loss=0.0553, train_acc=0.978, val_loss=0.0359, val_acc=0.983]\n" ] ], [ [ "## Test the model on the test data\n> Test the model on the test set, it is data that was not seen during training and allows us to validate model results. This model behaves fairly OK. It is really good at catching the fraud. It has a fairly low False Negative rate. (Lower left corner of the Confusion Matrix). But it also has a fairly large False Positive rate (Upper right corner of the Confusion Matrix).\nSome research would be needed but this is likely at least partially due to the oversampling. The model saw much more fraud during the training, so it might not be surprising it gets relatively good at predicting it.", "_____no_output_____" ] ], [ [ "test_ds = pt.NumpyListDataSetMulti(model_features, test_data)\ntest_dl = test_ds.data_loader(cpu, 128, num_workers=2)\n\nts = pt.Tester(m, device, test_dl)\npr = ts.test_plot()\ntp = pl.TestPlot()\ntp.print_classification_report(pr)\ntp.plot_confusion_matrix(pr, fig_size=(6,6))\ntp.plot_roc_curve(pr, fig_size=(6,6))\ntp.plot_precision_recall_curve(pr, fig_size=(6,6))", "Testing in 782 steps: 100%|██████████| 782/782 [00:00<00:00, 1046.42it/s]\n" ] ], [ [ "# Conslusion\nThis is a first example that showed how we can read a file, set-up some features (fields), test and train a really simple Feed-Forward Neural Net model.", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown", "markdown", "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ] ]

4ad642676f266c121558135b23886ea2d3966db3

10,966

ipynb

Jupyter Notebook

auto-labelling/strawberry autolabelling.ipynb

egorssed/digital_strawberry

02ede1135d10e8835b840ce89cac345a336b4a4b

[ "MIT" ]

null

auto-labelling/strawberry autolabelling.ipynb

egorssed/digital_strawberry

02ede1135d10e8835b840ce89cac345a336b4a4b

[ "MIT" ]

null

auto-labelling/strawberry autolabelling.ipynb

egorssed/digital_strawberry

02ede1135d10e8835b840ce89cac345a336b4a4b

[ "MIT" ]

1

2021-11-27T16:31:38.000Z

30.209366

129

0.534653

[ [ [ "### Before running:\nPlease set the file structure as follows: \n\n* parent folder\n\n * backgrounds - includes all background pictures\n\n * dataset - generated images and masks will be placed here \n\n * images\n\n * masks\n\n * strawberry autolabelling.ipynb\n\n * strawberry.png\n\n * strawberry2.png", "_____no_output_____" ] ], [ [ "from tqdm import tqdm\nimport os\nimport numpy as np\nimport matplotlib.pyplot as plt\nimport cv2\nimport pandas as pd", "_____no_output_____" ], [ "DATASET_PATH = \"backgrounds\" \nNEW_DATASET_PATH = \"dataset\"\n\nfiles = os.listdir(DATASET_PATH)\nlen(files), files[0]", "_____no_output_____" ], [ "MAX_SCALE = 1.5 # Max scale of strawberries during insertation", "_____no_output_____" ] ], [ [ "#### Load strawbery crops", "_____no_output_____" ] ], [ [ "image_to_insert_options = []\n\nfor file in [\"strawberry1.png\", \"strawberry2.png\"]:\n strawberry = cv2.imread(file, flags=cv2.IMREAD_UNCHANGED)\n strawberry = cv2.cvtColor(strawberry, cv2.COLOR_BGRA2RGBA)\n strawberry = cv2.rotate(strawberry, cv2.ROTATE_180)\n image_to_insert_options.append(strawberry)\n \nimage_to_insert_size = 100\nx_pad, y_pad = image_to_insert_size * MAX_SCALE, image_to_insert_size * MAX_SCALE # Insertation paddings", "_____no_output_____" ] ], [ [ "#### Modify colors in hsv space", "_____no_output_____" ] ], [ [ "def change_hsv(img, h_change=0, s_change=0, v_change=0):\n hsv = cv2.cvtColor(img, cv2.COLOR_BGR2HSV)\n h, s, v = cv2.split(hsv)\n \n h = cv2.add(h, h_change)\n h[h > 180] -= 180\n h[h < 0] += 180\n \n s = cv2.add(s,s_change)\n s[s > 255] = 255\n s[s < 0] = 0\n \n v = cv2.add(v,v_change)\n v[v > 255] = 255\n v[v < 0] = 0\n \n final_hsv = cv2.merge((h, s, v))\n img = cv2.cvtColor(final_hsv, cv2.COLOR_HSV2BGR)\n \n return img", "_____no_output_____" ] ], [ [ "#### Do augmentations on the image", "_____no_output_____" ] ], [ [ "def modify_image(image_to_insert):\n red2green = -57 # shift between colors in hue space\n \n # Rotate\n def rotate_image(image, angle):\n image_center = tuple(np.array(image_to_insert.shape[1::-1]) / 2)\n rot_mat = cv2.getRotationMatrix2D(image_center, angle, 1.0)\n \n return cv2.warpAffine(image_to_insert, rot_mat, image_to_insert.shape[1::-1], flags=cv2.INTER_LINEAR)\n\n image_to_insert = rotate_image(image_to_insert, angle=np.random.randint(-30, 30))\n\n # Flip\n if np.random.random() > 0.5:\n image_to_insert = cv2.flip(image_to_insert, 1)\n\n # Scale\n scale = np.random.random() + 0.5\n image_to_insert_resized = cv2.resize(image_to_insert, None, fx=scale, fy= scale)\n\n # Take the mask\n mask_strawberry = (image_to_insert_resized[:, :, 3] == 255) \n image_to_insert_resized = cv2.cvtColor(image_to_insert_resized, cv2.COLOR_RGBA2RGB)\n\n # Change color of strawberry\n r = np.random.random()\n if r > 0.66:\n # Keep red\n h_change = np.random.randint(-5, 5)\n s_change = np.random.randint(-25, -15)\n v_change = np.random.randint(-25, -15)\n elif r > 0.33:\n # Make half red\n h_change = np.random.randint(red2green + 30, red2green + 40)\n s_change = np.random.randint(-5, 5)\n v_change = np.random.randint(-5, 5)\n else:\n # Make green\n h_change = np.random.randint(red2green + 20, red2green + 25)\n s_change = np.random.randint(-150, -100)\n v_change = np.random.randint(30, 40)\n\n image_to_insert_resized = change_hsv(image_to_insert_resized, \n h_change=h_change, s_change=s_change, v_change=v_change)\n\n return image_to_insert_resized, mask_strawberry", "_____no_output_____" ], [ "backgrounds = os.listdir(os.path.join(DATASET_PATH))\nlen(backgrounds)", "_____no_output_____" ] ], [ [ "#### Generate new dataset", "_____no_output_____" ] ], [ [ "for filename in tqdm(backgrounds):\n \n # Prepare background\n background = cv2.imread(os.path.join(DATASET_PATH, filename))\n background = cv2.cvtColor(background, cv2.COLOR_BGR2RGB)\n\n max_dimension = max(background.shape)\n scale = 1000/max_dimension\n\n background = cv2.resize(background, None, fx=scale, fy=scale)\n\n # Find green regions to insert there strawberris\n background_hsv = cv2.cvtColor(background, cv2.COLOR_RGB2HSV)\n mask_green = cv2.inRange(background_hsv, (36, 25, 25), (86, 255,255)) # Green color range (36,25,25) ~ (86, 255,255)\n\n kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (15, 15)) # Clean up the mask\n mask_green_closed = cv2.morphologyEx(mask_green, cv2.MORPH_CLOSE, kernel)\n mask_green_clean = cv2.morphologyEx(mask_green_closed, cv2.MORPH_OPEN, kernel)\n\n # Get green pixels indices\n X, Y = np.where(mask_green_clean == 255)\n X, Y = zip(*[(x, y) for x, y in zip(X, Y) # Keep pixels within distance from borders\n if 0 < x < background.shape[0] - x_pad and 0 < y < background.shape[1] - y_pad])\n\n for iteration in range(3): # Use one background image to generate several (3) new images for dataset\n # Choose places to insert strawberries with random\n n_strawberries_to_insert = 5\n pixels_to_insert_ind = np.random.randint(0, len(X), n_strawberries_to_insert)\n\n # Insert strawberries\n new_image = background.copy()\n\n segmentation_masks = np.zeros(background.shape[0:2])\n bboxes = []\n\n for i, ind in enumerate(pixels_to_insert_ind): # Insert every strawberry\n image_to_insert = image_to_insert_options[np.random.randint(0, len(image_to_insert_options))]\n \n # Modify image\n image_to_insert_resized, mask_strawberry = modify_image(image_to_insert)\n \n # Insert crop\n x, y = X[ind], Y[ind]\n image_crop = new_image[x: x + image_to_insert_resized.shape[0], y: y + image_to_insert_resized.shape[1], :]\n image_crop[np.where(mask_strawberry)] = image_to_insert_resized[np.where(mask_strawberry)]\n\n # Create labels\n segmentation_masks[x: x + image_to_insert_resized.shape[0], \n y: y + image_to_insert_resized.shape[1]][mask_strawberry] = i + 1\n \n bboxes.append([x, y, image_to_insert_resized.shape[0], image_to_insert_resized.shape[1]])\n\n\n # Save new image and mask\n new_image = cv2.cvtColor(new_image, cv2.COLOR_RGB2BGR)\n cv2.imwrite(os.path.join(NEW_DATASET_PATH, \"images\", f\"{filename}_modified_{iteration}.png\"), new_image)\n\n with open(os.path.join(NEW_DATASET_PATH, \"masks\", f\"{filename}_modified_{iteration}.png\"), \"wb\") as f:\n np.save(f, segmentation_masks.astype(np.int8))\n#", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ] ]

4ad642b0477b82658c1906fc4fc50346d799e15c

434,717

ipynb

Jupyter Notebook

climate_analysis.ipynb

JoeKell/HonoluluClimateAnalysisAndApp

b7452a9aa9c5f0e0ab06f21f462c8e824c4b7b56

[ "ADSL" ]

null

climate_analysis.ipynb

JoeKell/HonoluluClimateAnalysisAndApp

b7452a9aa9c5f0e0ab06f21f462c8e824c4b7b56

[ "ADSL" ]

null

climate_analysis.ipynb

JoeKell/HonoluluClimateAnalysisAndApp

b7452a9aa9c5f0e0ab06f21f462c8e824c4b7b56

[ "ADSL" ]

null

516.904875

91,869

0.740666

[ [ [ "%matplotlib inline\nfrom matplotlib import style\nstyle.use('fivethirtyeight')\nimport matplotlib.pyplot as plt\nfrom scipy import stats as st", "_____no_output_____" ], [ "import numpy as np\nimport pandas as pd", "_____no_output_____" ], [ "import datetime as dt", "_____no_output_____" ] ], [ [ "# Reflect Tables into SQLAlchemy ORM", "_____no_output_____" ] ], [ [ "# Python SQL toolkit and Object Relational Mapper\nimport sqlalchemy\nfrom sqlalchemy.ext.automap import automap_base\nfrom sqlalchemy.orm import Session\nfrom sqlalchemy import create_engine, func", "_____no_output_____" ], [ "engine = create_engine(\"sqlite:///Resources/hawaii.sqlite\")\nconn=engine.connect()", "_____no_output_____" ], [ "# reflect an existing database into a new model\nBase = automap_base()\nBase.prepare(engine, reflect=True)\n# reflect the tables\nBase.classes.keys()", "_____no_output_____" ], [ "# Save references to each table\nMeasurement=Base.classes.measurement\nStation=Base.classes.station", "_____no_output_____" ], [ "# Create our session (link) from Python to the DB\nsession = Session(engine)", "_____no_output_____" ] ], [ [ "# Exploratory Climate Analysis", "_____no_output_____" ] ], [ [ "# Design a query to retrieve the last 12 months of precipitation data and plot the results\n# I've made the decision to limit my search to the station in Honolulu so I'm not plotting multiple values on the same date.\n\n# Calculate the date 1 year ago from the last data point in the database\nYearStart=dt.datetime.date(dt.datetime.fromisoformat(engine.execute(\"SELECT max(date) FROM measurement WHERE station in (SELECT station FROM station WHERE name like '%HONOLULU%')\").fetchall()[0][0]))- dt.timedelta(days=365)\nprint(f\"The date the year of data will start on is {YearStart}.\")\n\n# Perform a query to retrieve the date and precipitation scores\n# Since the start date is 2014-10-30, we know there isn't a leap day in our year of data.\nengine.execute(f\"SELECT date, prcp FROM measurement WHERE (date >= '{YearStart}') AND (station in (SELECT station FROM station WHERE name like '%HONOLULU%')) ORDER BY date LIMIT 10\").fetchall()\n\n# Save the query results as a Pandas DataFrame and set the index to the date column\ndf = pd.read_sql(f\"SELECT date, prcp FROM measurement WHERE (date >= '{YearStart}') AND (station in (SELECT station FROM station WHERE name like '%HONOLULU%')) ORDER BY date\", conn)\ndf.set_index('date', inplace=True)\n# Sort the dataframe by date\ndf.head()", "The date the year of data will start on is 2014-10-30.\n" ], [ "# Use Pandas Plotting with Matplotlib to plot the data\nx_axis=df.index\nbars=df.prcp\nplt.figure(figsize=(10,5))\nplt.bar(x_axis,bars,width=2)\nplt.xticks([x for i,x in enumerate(x_axis) if i%30==0],rotation = 45)\nplt.ylim(-.05,1.41)\nplt.xlabel(\"Date\")\nplt.ylabel(\"Precipitation (in)\")\nplt.title(\"Precipitation by Date in Honolulu\")\nplt.show()", "_____no_output_____" ], [ "# Use Pandas to calcualte the summary statistics for the precipitation data\ndf.describe()", "_____no_output_____" ], [ "# Design a query to show the number of stations available in this dataset.\nStation_Count=engine.execute(\"SELECT COUNT(distinct station) FROM measurement\").fetchall()[0][0]\nprint(f\"The number of unique stations available in the dataset is {Station_Count}.\")", "The number of unique stations available in the dataset is 9.\n" ], [ "# What are the most active stations? (i.e. what stations have the most rows)?\n# List the stations and the counts in descending order.\nengine.execute(\"SELECT station, COUNT(prcp) AS Records_Count, (SELECT name FROM station WHERE measurement.station = station.station) AS station_name FROM measurement GROUP BY station_name ORDER BY Records_Count DESC\").fetchall()\n", "_____no_output_____" ], [ "# Using the station id from the previous query, calculate the lowest temperature recorded, \n# highest temperature recorded, and average temperature of the most active station.\nStation_info=session.query(func.min(Measurement.tobs), func.max(Measurement.tobs), func.avg(Measurement.tobs)).filter(Measurement.station=='USC00519281').all()[0]\nprint(f\"The temp. info for station USC00519281 is the following:\\nLow: {Station_info[0]}\\nHigh: {Station_info[1]}\\nAvg: {round(Station_info[2],2)}\")\n", "The temp. info for station USC00519281 is the following:\nLow: 54.0\nHigh: 85.0\nAvg: 71.66\n" ], [ "# Choose the station with the highest number of temperature observations.\nengine.execute(\"SELECT station, COUNT(tobs) AS Records_Count, (SELECT name FROM station WHERE measurement.station = station.station) AS station_name FROM measurement GROUP BY station_name ORDER BY Records_Count DESC\").fetchall()\n", "_____no_output_____" ], [ "# Query the last 12 months of temperature observation data for this station and plot the results as a histogram\n# Calculate the date 1 year ago from the last data point in the database\nYearStart=dt.datetime.date(dt.datetime.fromisoformat(engine.execute(\"SELECT max(date) FROM measurement WHERE station = 'USC00519281'\").fetchall()[0][0]))- dt.timedelta(days=365)\nprint(f\"The date the year of data will start on is {YearStart}.\")\n\n# Save the query results as a Pandas DataFrame and set the index to the date column\nTemp_df = pd.read_sql(f\"SELECT date, tobs as temp FROM measurement WHERE (date >= '{YearStart}') AND (station = 'USC00519281') ORDER BY date\", conn)\nTemp_df.set_index('date', inplace=True)\n# Sort the dataframe by date\nTemp_df.head()", "The date the year of data will start on is 2016-08-18.\n" ], [ "#Plot the histogram\nTemp_df[\"temp\"].plot.hist(title=\"Temperature Distribution in Waihee\", color=\"red\",bins=12)\nplt.xlabel(\"Temp (F)\")\nplt.show()", "_____no_output_____" ] ], [ [ "## Bonus Challenge", "_____no_output_____" ], [ "## Hawaii is reputed to enjoy mild weather all year. Is there a meaningful difference between the temperature in June and December?", "_____no_output_____" ] ], [ [ "#Determine the average temp in June and December across all stations and all years of the data set.\nJune_Avg=engine.execute(\"SELECT avg(tobs), count(tobs), (SELECT SUM((tobs-(SELECT AVG(tobs) FROM measurement))*(tobs-(SELECT AVG(tobs) FROM measurement)) ) / (COUNT(tobs)-1)) AS Variance FROM measurement WHERE date LIKE '%-06-%' ORDER BY date\").fetchall()[0]\nDec_Avg=engine.execute(\"SELECT avg(tobs), count(tobs), (SELECT SUM((tobs-(SELECT AVG(tobs) FROM measurement))*(tobs-(SELECT AVG(tobs) FROM measurement)) ) / (COUNT(tobs)-1)) AS Variance FROM measurement WHERE date LIKE '%-12-%' ORDER BY date\").fetchall()[0]\nprint('(average , count, variance )')\n#To use the independent t-test to compare 2 means using an we need to see if the count in each sample is similar and if the variances are 'close' meaning one isn't more that twice the other. \nprint(June_Avg)\nprint(Dec_Avg)", "(average , count, variance )\n(74.94411764705882, 1700, 14.021092266714305)\n(71.04152933421226, 1517, 18.26358709666216)\n" ], [ "#We need to show that these samples are normally distributed from the histograms below, it is clear that they are.\npd.read_sql(\"SELECT tobs FROM measurement WHERE date LIKE '%-06-%' ORDER BY date\", conn).plot.hist(title=\"June Temperature Distribution\", color=\"red\", legend=None)\nplt.xlabel(\"Temp (F)\")\nplt.show()", "_____no_output_____" ], [ "pd.read_sql(\"SELECT tobs FROM measurement WHERE date LIKE '%-12-%' ORDER BY date\", conn).plot.hist(title=\"December Temperature Distribution\", color=\"red\", legend=None)\nplt.xlabel(\"Temp (F)\")\nplt.show()", "_____no_output_____" ], [ "June=np.ravel(engine.execute(\"SELECT tobs FROM measurement WHERE date LIKE '%-06-%' ORDER BY date\").fetchall())\nDecember=np.ravel(engine.execute(\"SELECT tobs FROM measurement WHERE date LIKE '%-12-%' ORDER BY date\").fetchall())\nprint(June,len(June))\nprint(December,len(December))", "[78. 74. 73. ... 75. 76. 75.] 1700\n[76. 73. 73. ... 72. 67. 65.] 1517\n" ], [ "#Use the t-test to determine whether the difference in the means, if any, is statistically significant. Will you use a paired t-test, or an unpaired t-test? Why? - answered above. Also we use an unpaired test becasue the samples are independent of eachother.\nst.ttest_ind(June,December)\n", "_____no_output_____" ] ], [ [ "According to the test the means are different and the p-value being so low indicates that with any other sample, this is extremely likely to be the outcome again. ", "_____no_output_____" ], [ "## Bonus part 2", "_____no_output_____" ] ], [ [ "# This function called `calc_temps` will accept start date and end date in the format '%Y-%m-%d' \n# and return the minimum, average, and maximum temperatures for that range of dates\ndef calc_temps(start_date, end_date):\n \"\"\"TMIN, TAVG, and TMAX for a list of dates.\n \n Args:\n start_date (string): A date string in the format %Y-%m-%d\n end_date (string): A date string in the format %Y-%m-%d\n \n Returns:\n TMIN, TAVE, and TMAX\n \"\"\"\n \n return session.query(func.min(Measurement.tobs), func.avg(Measurement.tobs), func.max(Measurement.tobs)).\\\n filter(Measurement.date >= start_date).filter(Measurement.date <= end_date).all()\n\n# function usage example\nprint(calc_temps('2012-02-28', '2012-03-05'))", "[(62.0, 69.57142857142857, 74.0)]\n" ], [ "# Use your previous function `calc_temps` to calculate the tmin, tavg, and tmax \n# for your trip using the previous year's data for those same dates.\ndef VacEst(VS,VE):\n LY_Vacation_Start=str(int(VS[:4])-1)+VS[4:]\n LY_Vacation_End=str(int(VE[:4])-1)+VE[4:]\n [(tmin,tavg,tmax)]=calc_temps(LY_Vacation_Start, LY_Vacation_End)\n return (tmin,tavg,tmax)\n\nVacation_Start='2016-04-03'\nVacation_End='2016-04-13'\n(tmin,tavg,tmax)=VacEst(Vacation_Start,Vacation_End)", "_____no_output_____" ], [ "# Plot the results from your previous query as a bar chart. \n# Use \"Trip Avg Temp\" as your Title\n# Use the average temperature for the y value\n# Use the peak-to-peak (tmax-tmin) value as the y error bar (yerr)\n\nplt.bar([\"\"],[tavg],width=.33, color='red',yerr=tmax-tmin)\nplt.ylabel(\"Average Temperature (F)\")\nplt.title(\"Trip Avg Temp\")\nplt.xlim(-.5,.5)\nplt.ylim(0,100)\nplt.show()", "_____no_output_____" ], [ "# Calculate the total amount of rainfall per weather station for your trip dates using the previous year's matching dates.\n# Sort this in descending order by precipitation amount and list the station, name, latitude, longitude, and elevation\nLY_Vacation_Start=str(int(Vacation_Start[:4])-1)+Vacation_Start[4:]\nLY_Vacation_End=str(int(Vacation_End[:4])-1)+Vacation_End[4:]\nengine.execute(f\"SELECT m.date, ROUND(SUM(m.prcp),2) as TotalRainfall, m.station, s.name, s.latitude, s.longitude, s.elevation FROM measurement m INNER JOIN station s ON s.station = m.station WHERE m.date <= '{LY_Vacation_End}' and m.date >= '{LY_Vacation_Start}' GROUP BY m.station ORDER BY TotalRainfall DESC\").fetchall()\n", "_____no_output_____" ], [ "# Create a query that will calculate the daily normals \n# (i.e. the averages for tmin, tmax, and tavg for all historic data matching a specific month and day)\n\ndef daily_normals(date):\n \"\"\"Daily Normals.\n \n Args:\n date (str): A date string in the format '%m-%d'\n \n Returns:\n A list of tuples containing the daily normals, tmin, tavg, and tmax\n \n \"\"\"\n \n sel = [func.min(Measurement.tobs), func.avg(Measurement.tobs), func.max(Measurement.tobs)]\n return session.query(*sel).filter(func.strftime(\"%m-%d\", Measurement.date) == date).all()\n \ndaily_normals(\"01-01\")", "_____no_output_____" ], [ "# calculate the daily normals for your trip\n# push each tuple of calculations into a list called `normals`\n\n# Set the start and end date of the trip\nVacation_Start='2016-04-03'\nVacation_End='2016-04-13'\n# Use the start and end date to create a range of dates\ndateRange=[]\ndateCounter=Vacation_Start\nwhile dateCounter<=Vacation_End:\n dateRange.append(dateCounter[5:])\n if int(dateCounter[-2:])<9:\n dateCounter=dateCounter[:9]+str(int(dateCounter[-1:])+1)\n else:\n dateCounter=dateCounter[:8]+str(int(dateCounter[-2:])+1)\n\n# Stip off the year and save a list of %m-%d strings\nprint(dateRange)\n# Loop through the list of %m-%d strings and calculate the normals for each date\nTripNormals=[]\nfor date in dateRange:\n TripNormals.append(daily_normals(date)[0])\n\nprint (TripNormals)", "['04-03', '04-04', '04-05', '04-06', '04-07', '04-08', '04-09', '04-10', '04-11', '04-12', '04-13']\n[(55.0, 72.2280701754386, 78.0), (58.0, 72.01639344262296, 78.0), (61.0, 72.28813559322033, 80.0), (67.0, 71.57142857142857, 77.0), (63.0, 71.10526315789474, 78.0), (66.0, 72.10526315789474, 77.0), (62.0, 71.6140350877193, 78.0), (64.0, 71.40350877192982, 79.0), (65.0, 72.49180327868852, 82.0), (65.0, 72.22413793103448, 80.0), (64.0, 71.52542372881356, 79.0)]\n" ], [ "# Load the previous query results into a Pandas DataFrame and add the `trip_dates` range as the `date` index\nTripMin=[x[0] for x in TripNormals]\nTripAvg=[round(x[1],2) for x in TripNormals]\nTripMax=[x[2] for x in TripNormals]\nTrip_df=pd.DataFrame({\n \"Date\": dateRange,\n \"Min\": TripMin,\n \"Avg\": TripAvg,\n \"Max\": TripMax\n})\nTrip_df.set_index('Date', inplace=True)\nTrip_df.head()", "_____no_output_____" ], [ "# Plot the daily normals as an area plot with `stacked=False`\nTrip_df.plot.area(stacked=False)\nplt.ylabel(\"Temperature (F)\")\nplt.title(\"Daily Normals for Hawaii Trip\")\nplt.show()\n#Although this is what was asked, using an area chart instead of a line chart here is a mistake. There is no value added by filling in the area.", "_____no_output_____" ] ] ]

[ "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad64942dde04b95d78a0f69b75189a8a43f4396

195,644

ipynb

Jupyter Notebook

Homework/Econ126_Winter2020_Homework_05.ipynb

t-hdd/econ126

17029937bd6c40e606d145f8d530728585c30a1d

[ "MIT" ]

null

Homework/Econ126_Winter2020_Homework_05.ipynb

t-hdd/econ126

17029937bd6c40e606d145f8d530728585c30a1d

[ "MIT" ]

null

Homework/Econ126_Winter2020_Homework_05.ipynb

t-hdd/econ126

17029937bd6c40e606d145f8d530728585c30a1d

[ "MIT" ]

null

299.149847

29,888

0.923765

[ [ [ "import numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nplt.style.use('classic')\n%matplotlib inline", "_____no_output_____" ] ], [ [ "# Homework 5\n\n**Instructions:** Complete the notebook below. Download the completed notebook in HTML format. Upload assignment using Canvas.\n\n**Due:** Feb. 20 at **12:30pm.**", "_____no_output_____" ], [ "## Exercise: AR(1) Process Simulation\n\nFor each of the AR(1) processes defined below:\n\n1. Compute $y$ for $201$ periods (i.e., $t=0,\\ldots, 200$).\n2. Construct a well-labeled plot of $y$ for each simulation.\n2. Print the mean and standard deviation of the stimulated process.\n\nYou may use the function from the Notebook for Lecture 10, write your own, or take a more brute force approach.", "_____no_output_____" ] ], [ [ "# Define a function for computing the first-order difference equation. CELL NOT PROVIDED\ndef ar1_sim(rho,sigma,y0,T,):\n '''Function for simulating an AR(1) process \n \n y[t] = rho*y[t-1] + epsilon[t]\n \n Args:\n rho (float): Autoregressive coefficient on y[t-1]\n sigma (float): Standard deviation of epsilon\n y0 (float): Initial value of the process\n T (int): Number of periods to simulate including initial period\n \n Returns:\n NumPy ndarray\n '''\n \n y = np.zeros(T)\n epsilon = np.random.normal(scale=sigma,size=T)\n y[0] = y0\n for t in range(T-1):\n y[t+1] = rho*y[t]+epsilon[t+1]\n \n return y\n\n\n# Create variable 'T' equal to number of simulation periods\nT = 201\n\n# Set NumPy RNG seed to 126 st that simulations look the same each time I run the Notebook\nnp.random.seed(126)", "_____no_output_____" ] ], [ [ "1. $y_{t} = 0.33 y_{t-1} + \\epsilon_t, \\; \\; \\; \\; \\epsilon \\sim \\mathcal{N}(0,1), \\; \\; \\; \\; y_0 = 0$", "_____no_output_____" ] ], [ [ "rho=0.33\nsigma = 1\ny0 = 0\n\ny = ar1_sim(rho,sigma,y0,T)\n\nplt.plot(y,lw=3)\nplt.title('AR(1): $\\\\rho='+str(rho)+'$ and $\\\\sigma='+str(sigma)+'$')\nplt.xlabel('$t$',fontsize=15)\nplt.ylabel('$y_t$',fontsize=15)\nplt.grid()", "_____no_output_____" ], [ "# print mean of y\nprint('mean of y: ',round(y.mean(),4))\n\n# Print standard deviation of y\nprint('standard deviation of y: ',round(y.std(),4))", "mean of y: -0.0743\nstandard deviation of y: 1.0322\n" ] ], [ [ "2. $y_{t} = y_{t-1} + \\epsilon_t, \\; \\; \\; \\; \\epsilon \\sim \\mathcal{N}(0,1), \\; \\; \\; \\; y_0 = 0$", "_____no_output_____" ] ], [ [ "rho=1\nsigma = 1\ny0 = 0\n\ny = ar1_sim(rho,sigma,y0,T)\n\nplt.plot(y,lw=3)\nplt.title('AR(1): $\\\\rho='+str(rho)+'$ and $\\\\sigma='+str(sigma)+'$')\nplt.xlabel('$t$',fontsize=15)\nplt.ylabel('$y_t$',fontsize=15)\nplt.grid()", "_____no_output_____" ], [ "# print mean of y\nprint('mean of y: ',round(y.mean(),4))\n\n# Print standard deviation of y\nprint('standard deviation of y:',round(y.std(),4))", "mean of y: -3.0308\nstandard deviation of y: 3.8289\n" ] ], [ [ "3. $y_{t} = -0.95 y_{t-1} + \\epsilon_t, \\; \\; \\; \\; \\epsilon \\sim \\mathcal{N}(0,1), \\; \\; \\; \\; y_0 = 0$", "_____no_output_____" ] ], [ [ "rho=-0.95\nsigma = 1\ny0 = 0\n\ny = ar1_sim(rho,sigma,y0,T)\n\nplt.plot(y,lw=3)\nplt.title('AR(1): $\\\\rho='+str(rho)+'$ and $\\\\sigma='+str(sigma)+'$')\nplt.xlabel('$t$',fontsize=15)\nplt.ylabel('$y_t$',fontsize=15)\nplt.grid()", "_____no_output_____" ], [ "# print mean of y\nprint('mean of y: ',round(y.mean(),4))\n\n# Print standard deviation of y\nprint('standard deviation of y: ',round(y.std(),4))", "mean of y: -0.0131\nstandard deviation of y: 3.1956\n" ] ], [ [ "4. $y_{t} = 0.75 y_{t-1} + \\epsilon_t , \\; \\; \\; \\; \\epsilon \\sim \\mathcal{N}(0,4), \\; \\; \\; \\; y_0 = 20$", "_____no_output_____" ] ], [ [ "rho=0.75\nsigma = 2\ny0 = 20\n\ny = ar1_sim(rho,sigma,y0,T)\n\nplt.plot(y,lw=3)\nplt.title('AR(1): $\\\\rho='+str(rho)+'$ and $\\\\sigma='+str(sigma)+'$')\nplt.xlabel('$t$',fontsize=15)\nplt.ylabel('$y_t$',fontsize=15)\nplt.grid()", "_____no_output_____" ], [ "# print mean of y\nprint('mean of y: ',round(y.mean(),4))\n\n# Print standard deviation of y\nprint('standard deviation of y:',round(y.std(),4))", "mean of y: 0.8346\nstandard deviation of y: 3.6313\n" ] ], [ [ "5. $y_{t} = 0.9 y_{t-1} + \\epsilon_t , \\; \\; \\; \\; \\epsilon \\sim \\mathcal{N}(0,3), \\; \\; \\; \\; y_0 = 0$", "_____no_output_____" ] ], [ [ "rho=0.9\nsigma = np.sqrt(3)\ny0 = 0\n\ny = ar1_sim(rho,sigma,y0,T)\n\nplt.plot(y,lw=3)\nplt.title('AR(1): $\\\\rho='+str(rho)+'$ and $\\\\sigma='+str(round(sigma,4))+'$')\nplt.xlabel('$t$',fontsize=15)\nplt.ylabel('$y_t$',fontsize=15)\nplt.grid()", "_____no_output_____" ], [ "# print mean of y\nprint('mean of y: ',round(y.mean(),4))\n\n# Print standard deviation of y\nprint('standard deviation of y: ',round(y.std(),4))", "mean of y: 2.8525\nstandard deviation of y: 4.5849\n" ] ], [ [ "## Exercise: Stochastic Growth\n\nWe've seen in class that total factory productivity (TFP) in the US fluctuates over the business cycle around a long-term growth trend. That is, total factor productivity is characterized by *stochastic growth*. In this problem, you will simulate a model of TFP with stochastic growth.\n\n### Background\n\nSuppose that the trend component of TFP $A^{trend}$ grows smoothly at the constant rate $g$:\n\n\\begin{align}\nA^{trend}_t & = (1+g)A_{t-1}^{trend}, \\tag{1}\n\\end{align}\n\nwhere $A_0$ is given. The log-deviations of actual TFP $A$ from trend TFP are determined according to an AR(1) process:\n\n\\begin{align}\n\\log\\left(A_t/A^{trend}_t\\right) & = \\rho \\log \\left(A_{t-1} / A^{trend}_{t-1}\\right) + \\epsilon_t, \\tag{2}\n\\end{align}\n\nwhere $A_0$ is given and $\\epsilon_t \\sim \\mathcal{N}(0,\\sigma^2)$. Solve quation (2) for $A_t$ to get an expression for $A_t$ in terms of $A_{t-1}$, $A^{trend}_{t}$, $A^{trend}_{t-1}$, and $\\epsilon_t$:\n\n\\begin{align}\nA_t & = A^{trend}_t \\displaystyle e^{\\rho \\log \\left(A_{t-1} / A^{trend}_{t-1}\\right) + \\epsilon_t} \\tag{3}\n\\end{align}\n\n### Simulation\n\nTo simulate this model:\n\n1. Specify values for $A_0$ and $A^{trend}_0$.\n2. For $t = 0, 1, \\ldots... T$, compute $A^{trend}_{t+1}$ using equation (1) and then compute $A_{t+1}$ using equation (3)\n\nUse the following prameter values for this simulation\n\n\n| $A^{trend}_0$ | $A_0$ | $g$ | $\\rho$ | $\\sigma$ | $T$ |\n|---------------|-------|------|--------|----------|-----|\n| 1 | 1 | 0.02 | 0.7 | 0.025 | 101 |\n\nIn the follwing cell, simulate $A^{trend}$ and $A$ using the parameter values provided.", "_____no_output_____" ] ], [ [ "# Simulate the stochastic growth model of TFP\nT = 101\ng = 0.02\nrho = 0.7\nsigma = 0.025\n\nepsilon = np.random.normal(scale=sigma,size=T)\na_trend = np.ones(T)*1\na = np.ones(T)*1\n\nfor t in range(T-1):\n a_trend[t+1] = (1+g)*a_trend[t]\n a[t+1] = a_trend[t+1]*np.exp(rho*np.log(a[t]/a_trend[t])+epsilon[t])", "_____no_output_____" ] ], [ [ "Contruct a well-labeled plot of the simulated values of $A$ and $A^{trend}$", "_____no_output_____" ] ], [ [ "# Construct a plot of simulated TFP with its trend with:\n# 1. Actual line: blue with lw=1, alpha=0.7, label = 'actual'\n# 2. Trend line: red with lw=3, alpha=0.25, label = 'trend'\nplt.plot(a_trend,'r',lw=3,alpha=0.25,label='Trend')\nplt.plot(a,label='Actual')\nplt.title('Simulated TFP with Stochastic Growth')\nplt.legend(loc='center left', bbox_to_anchor=(1, 0.5))\nplt.grid()", "_____no_output_____" ] ], [ [ "Contruct a well-labeled plot of the simulated values of $\\log \\left(A/A^{trend}\\right)$", "_____no_output_____" ] ], [ [ "# Construct a plot of simulated log deviation of TFP from trend\nplt.plot(np.log(a/a_trend),lw=3)\nplt.title('Simulated $\\log A_t/A^{trend}$')\nplt.grid()", "_____no_output_____" ] ], [ [ "Compute the mean and standard deviation of $\\log \\left(A/A^{trend}\\right)$", "_____no_output_____" ] ], [ [ "# print mean of log(A/A trend)\nprint('mean of log(A/A trend): ',round((np.log(a/a_trend)).mean(),4))\n\n# Print standard deviation of log(A/A trend)\nprint('standard deviation of log(A/A trend): ',round((np.log(a/a_trend)).std(),4))", "mean of log(A/A trend): -0.0192\nstandard deviation of log(A/A trend): 0.03\n" ] ], [ [ "## Exercise: Questions for Prescott (1983) Reading\n\n\nThe following questions are about Edward Prescott's 1986 article \"Theory Ahead of Business Cycle Measurement\" from the Fall 1986 issue of the Minneapolis Fed's *Quarterly Review* (link to article: [https://www.minneapolisfed.org/research/qr/qr1042.pdf](https://www.minneapolisfed.org/research/qr/qr1042.pdf)). The article is a summary of the research agenda of the author, his main collaborator, Finn Kydland, and others. That agenda entailed incorporating stochastic growth of TFP into the neoclassical growth model (essentially a Solow model but with the saving rate determined as a consequence a of utility maximization problem) as a way of modeling business cycle fluctuations. This line of research would later be called *real business cycle* theory.\n\nThe below questions are specifically about:\n* Pages 9-11 (including the table on page 12)\n* The final paragraph on page 21", "_____no_output_____" ], [ "**Question:** On page 10, Prescott writes that the \"models constructed within this theoretical framework are necessarily highly abstract. Consequently they are necessarily false.\" What does Kydland mean by this? In what sense are abstract models necessarily false? ", "_____no_output_____" ], [ "**Answer**\n\nAll theoretical models are approximations of reality and therefore will not relfect the complexity of the natural environment. But that doesn't mean that abstract models can't be useful. In fact, abstraction helps us understand complicated phenomena by emphasizing only the most relevant aspects of the issue. ", "_____no_output_____" ], [ "**Question:** On page 10, Prescott lists two reasons for why he doesn't like the term *business cycle*. What is the first reason and what does it mean?", "_____no_output_____" ], [ "**Answer**\n\nThe first reason that Prescott doesn't like the term *business cyle* is that it implies that short-term fluctuations are the result of economic forces that are *independent* of long-term growth. In Prescott's view, short-term fluctuations and long-term growth *are a consequence of the same process*: stochastic TFP growth. That is, TFP grows, but at a random and unpredictable rate.", "_____no_output_____" ], [ "**Question:** Table 1 on page 12 shows that household consumption measures in the US fluctuate less than investment measures. What is the intuition for why consumption varies less than investment?", "_____no_output_____" ], [ "**Answer**\n\nPeople like to smooth their consumption over time as much as they can. When houseohlds have higher-than-average income, they save most for the future and so investment rises more than consumption when incomes are high. And when household incomes are lower-than-average, households primarily reduce saving and investment falls more than consumption when incomes are high.", "_____no_output_____" ], [ "**Question:** In the last paragraph of page 21, to what does Prescott attribute economic fluctuations? Can you think of any reasons why this statement was, and still is, controversial?", "_____no_output_____" ], [ "**Answer**\n\n\"Economic fluctuations are the optimal responses to uncertainty in the rate of technological change.\" The statement is controversial for several reasons. \n* First, the argument attributes *all* economic fluctuations to TFP variation and disregards other plausible candidates like, for example, exogenous variation in monetary policy. \n* Second, Prescott's claim that fluctuations are *optimal* implies that the economy is *always* doing the best it can with what it has: When unemployment is high, it's because TFP fell relative to trend and households decided to take more leisure. \n* Third, the statement implies that, even in principle, policymakers cannot improve the welfare of the public by stabilizing the business cycle. ", "_____no_output_____" ] ] ]

[ "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown" ] ]

4ad64e1ae8b26f14303b9246d51f089b327c9213

315,014

ipynb

Jupyter Notebook

notebooks/.ipynb_checkpoints/read_avhr_temp_data-checkpoint.ipynb

etrieschman/ocean-carbon-sampling

dee5882566c02448bf7e91e30b084095b1d05a08

[ "MIT" ]

null

notebooks/.ipynb_checkpoints/read_avhr_temp_data-checkpoint.ipynb

etrieschman/ocean-carbon-sampling

dee5882566c02448bf7e91e30b084095b1d05a08

[ "MIT" ]

null

notebooks/.ipynb_checkpoints/read_avhr_temp_data-checkpoint.ipynb

etrieschman/ocean-carbon-sampling

dee5882566c02448bf7e91e30b084095b1d05a08

[ "MIT" ]

null

816.098446

282,792

0.95598

[ [ [ "from pathlib import Path\nfrom tqdm import tqdm\nimport pandas as pd\nimport numpy as np\nimport matplotlib as matplotlib\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nimport os\nimport datetime as dt\nfrom shapely import wkt\nfrom shapely.geometry import Point, Polygon\nimport geopandas as gpd\n# import rtree, pygeos, fiona\n# import netCDF4\nimport xarray as xr\n# import dask", "/Users/etriesch/.pyenv/versions/venv.ocean-carbon-sampling/lib/python3.9/site-packages/geopandas/_compat.py:111: UserWarning: The Shapely GEOS version (3.10.2-CAPI-1.16.0) is incompatible with the GEOS version PyGEOS was compiled with (3.10.1-CAPI-1.16.0). Conversions between both will be slow.\n warnings.warn(\n" ], [ "repo_path = Path('/Users/etriesch/dev/ocean-carbon-sampling/')\nwrite_path = repo_path / 'data/clean/'\ncrs = 'epsg:4326'", "_____no_output_____" ], [ "# Northeast Pacific mask\np_min_lat, p_max_lat = 29, 48\np_min_lon, p_max_lon = 360-140, 360-116 \n#Northwest Atlantic mask\na_min_lat, a_max_lat = 20, 48\na_min_lon, a_max_lon = 360-97, 360-60", "_____no_output_____" ] ], [ [ "# Read in data", "_____no_output_____" ] ], [ [ "# get filenames\ntemp_path = repo_path / 'data/raw/oisst/'\nt_files = [f for f in os.listdir(temp_path) \n if f.endswith('.nc')]\nt_files.sort()\nprint('files to read:', len(t_files))", "files to read: 120\n" ], [ "# read in data\nt_pac = pd.DataFrame()\nt_atl = pd.DataFrame()\nfor f in t_files:\n t = xr.open_dataset(temp_path + f)\n # create pacific and atlantic masks\n p_mask_lon = (t.lon >= p_min_lon) & (t.lon <= p_max_lon)\n p_mask_lat = (t.lat >= p_min_lat) & (t.lat <= p_max_lat)\n a_mask_lon = (t.lon >= a_min_lon) & (t.lon <= a_max_lon)\n a_mask_lat = (t.lat >= a_min_lat) & (t.lat <= a_max_lat)\n pac = t.where(p_mask_lon & p_mask_lat, drop=True)\n atl = t.where(a_mask_lon & a_mask_lat, drop=True)\n # convert to dataset and append\n pac = pac.sst.to_dataframe().reset_index()\n atl = atl.sst.to_dataframe().reset_index()\n t_pac = pd.concat([pac, t_pac])\n t_atl = pd.concat([atl, t_atl])", "_____no_output_____" ] ], [ [ "# Clean and subset", "_____no_output_____" ] ], [ [ "# make single df for cleaning\nt_pac['pacific'] = True\nt_atl['pacific'] = False\nt_raw = pd.concat([t_pac, t_atl])", "_____no_output_____" ], [ "t_raw['date'] = pd.to_datetime(t_raw.time).dt.date\nt_raw['year'] = pd.to_datetime(t_raw.time).dt.year", "_____no_output_____" ], [ "# clean temperature values\n# set outlier as 5 times the 75th percentile\npctl, outl = 0.75, 5\npctl_colname = 'pctl'+str(int(pctl*100))\nt_raw[pctl_colname] = t_raw.groupby(['year', 'pacific'])['sst'].transform(lambda x: x.quantile(pctl))\nprint('Outliers dropped (Pacific):', (t_raw.loc[t_raw.pacific].sst > outl * t_raw.loc[t_raw.pacific, pctl_colname]).sum() / t_raw.loc[t_raw.pacific].shape[0])\nprint('Outliers dropped (Atlantic):', (t_raw.loc[~t_raw.pacific].sst > outl * t_raw.loc[~t_raw.pacific, pctl_colname]).sum() / t_raw.loc[~t_raw.pacific].shape[0])\nt = t_raw.loc[t_raw.sst <= outl * t_raw[pctl_colname]].reset_index()", "Outliers dropped (Pacific): 0.0\nOutliers dropped (Atlantic): 0.0\n" ] ], [ [ "# Create analytical columns\n* Max SST\n* Min SST\n* Average annual SST\n* Average standard deviation S?ST", "_____no_output_____" ] ], [ [ "pd.set_option('display.max_rows', 200)", "_____no_output_____" ], [ "# get mean, max, std chlorophyll per year at each location\nt_s = t.groupby(['pacific', 'lat', 'lon', 'year'])['sst'].agg(['std', 'mean', 'max', 'min']).reset_index()\n# average across years\nt_s = t_s.groupby(['pacific', 'lat', 'lon']).agg('mean').reset_index()", "_____no_output_____" ] ], [ [ "# Write data to computer", "_____no_output_____" ] ], [ [ "\nfilename = 'oisst.csv'\n(write_path / filename).parent.mkdir(parents=True, exist_ok=True)", "_____no_output_____" ], [ "t_s.to_csv(filepath, index=False)", "_____no_output_____" ] ], [ [ "# Visualize", "_____no_output_____" ], [ "## temperature over time", "_____no_output_____" ] ], [ [ "# Visualize chlorophll at lat/long over time\ni = np.random.randint(0,t_s.shape[0])\nsub = t.loc[(t.lat == t_s.lat.iloc[i]) & (t.lon == t_s.lon.iloc[i])].sort_values('date')\n\nfig, ax = plt.subplots(figsize=(15,5))\nplt.plot(sub['date'], sub['sst'])\nplt.hlines(y=sub.sst.mean(), xmin=sub.date.min(), xmax=sub.date.max(), color='red')\n# print info and chart\nprint('index:', i)\nprint('location:', round(t_s.lat.iloc[i], 4), round(t_s.lon.iloc[i], 4))\nplt.show()", "index: 14573\nlocation: 44.375 233.875\n" ] ], [ [ "## Map of temperature property", "_____no_output_____" ] ], [ [ "world = gpd.read_file(gpd.datasets.get_path('naturalearth_lowres'))\n# world_snip = gpd.clip(world, gp)", "_____no_output_____" ], [ "# Visualize\n# temperature - pacific mean\ngeometry = [Point(xy) for xy in zip(t_s['lon'], t_s['lat'])]\ngp = gpd.GeoDataFrame(t_s, crs=crs, geometry=geometry)\n\nfig,ax = plt.subplots(figsize=(15,10))\ngp.plot(ax=ax, column='max', markersize=1, legend=True)\nplt.title('Mean')\nplt.show()", "_____no_output_____" ] ] ]

[ "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ] ]

4ad65f562fdc075c42a011e81497711ff03dfaa9

273,817

ipynb

Jupyter Notebook

Code_cnn-rnn.ipynb

saikiabedanta/Kaggle-Beyond-Analysis

6dc2bbec297dc1dfc6d293b272b3b4194b7ce26a

[ "MIT" ]

null

Code_cnn-rnn.ipynb

saikiabedanta/Kaggle-Beyond-Analysis

6dc2bbec297dc1dfc6d293b272b3b4194b7ce26a

[ "MIT" ]

null

Code_cnn-rnn.ipynb

saikiabedanta/Kaggle-Beyond-Analysis

6dc2bbec297dc1dfc6d293b272b3b4194b7ce26a

[ "MIT" ]

null

95.40662

119,698

0.718703

[ [ [ "", "_____no_output_____" ], [ "from google.colab import drive\ndrive.mount('/content/drive')", "Mounted at /content/drive\n" ], [ "import numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nimport tensorflow as tf\nimport tensorflow.keras.layers as tfl\nfrom tensorflow import keras", "_____no_output_____" ], [ "device_name = tf.test.gpu_device_name()\nif device_name != '/device:GPU:0':\n raise SystemError('GPU device not found')\nprint('Found GPU at: {}'.format(device_name))", "Found GPU at: /device:GPU:0\n" ], [ "!nvidia-smi", "Wed Sep 1 10:45:13 2021 \n+-----------------------------------------------------------------------------+\n| NVIDIA-SMI 470.57.02 Driver Version: 460.32.03 CUDA Version: 11.2 |\n|-------------------------------+----------------------+----------------------+\n| GPU Name Persistence-M| Bus-Id Disp.A | Volatile Uncorr. ECC |\n| Fan Temp Perf Pwr:Usage/Cap| Memory-Usage | GPU-Util Compute M. |\n| | | MIG M. |\n|===============================+======================+======================|\n| 0 Tesla K80 Off | 00000000:00:04.0 Off | 0 |\n| N/A 37C P0 57W / 149W | 121MiB / 11441MiB | 0% Default |\n| | | N/A |\n+-------------------------------+----------------------+----------------------+\n \n+-----------------------------------------------------------------------------+\n| Processes: |\n| GPU GI CI PID Type Process name GPU Memory |\n| ID ID Usage |\n|=============================================================================|\n| No running processes found |\n+-----------------------------------------------------------------------------+\n" ], [ "from keras import layers\nfrom keras import Input\nfrom keras import optimizers\nfrom keras.models import Model\nfrom keras.layers import Dropout, Flatten, Conv1D,LSTM,Bidirectional\nimport random", "_____no_output_____" ], [ "from sklearn.preprocessing import StandardScaler\nfrom platform import python_version\nfrom keras.constraints import maxnorm", "_____no_output_____" ], [ "data = pd.read_csv(\"/content/drive/MyDrive/techniche/train.csv\")\ntest = pd.read_csv(\"/content/drive/MyDrive/techniche/test.csv\")", "_____no_output_____" ], [ "def one_hot_encoder(df, features):\n dummies = pd.get_dummies(df[features], drop_first=True)\n res = pd.concat([df, dummies], axis=1)\n res = res.drop(features, axis=1)\n return(res)", "_____no_output_____" ], [ "categorical_cols = list(data.select_dtypes(include=[\"object\"]).columns)\ndata_onehot = one_hot_encoder(data, categorical_cols)\ntest_onehot = one_hot_encoder(test, categorical_cols)", "_____no_output_____" ], [ "cols_train = list(data_onehot.columns[:25]) + ['CATEGORY_1_D', 'CATEGORY_1_E'] + list(data_onehot.columns[25:])\ncols_train\ndata_onehot['CATEGORY_1_D']=0\ndata_onehot['CATEGORY_1_E']=0\ndata_onehot = data_onehot[cols_train]\ndata_onehot.head()", "_____no_output_____" ], [ "Y_train = data_onehot.groupby('UNIQUE_IDENTIFIER').mean()[['Y1', 'Y2']]\nY_train.head()", "_____no_output_____" ], [ "X_train = data_onehot.drop([\"Y1\", \"Y2\"], axis=1)\nX_train.head()", "_____no_output_____" ], [ "def input_generator(df_input):\n \n customer_stamp = df_input['UNIQUE_IDENTIFIER'].value_counts()\n customer_id = df_input['UNIQUE_IDENTIFIER'].unique()\n n_customer_stamp = customer_stamp[customer_id].values\n cumsum_customer_stamp = np.cumsum(n_customer_stamp, axis=0)\n time_stamp = np.hstack((0, cumsum_customer_stamp))\n row_size = np.max(n_customer_stamp)\n input_array = np.ones((customer_id.shape[0], row_size, df_input.shape[1]))*(-1)\n \n for i in range(time_stamp.shape[0]-1):\n temp_array = df_input.iloc[time_stamp[i]:time_stamp[i+1], :].values\n input_array[i, :, :] = np.pad(temp_array, ((0, row_size-temp_array.shape[0]), (0, 0)), mode='constant', constant_values=(-1, 0))\n \n return input_array", "_____no_output_____" ], [ "scaler2 = StandardScaler()\nX2_train_scaled = pd.DataFrame(scaler2.fit_transform(X_train), columns=X_train.columns)\nX2_test_scaled = pd.DataFrame(scaler2.transform(test_onehot), columns=test_onehot.columns)", "_____no_output_____" ], [ "X2_train = input_generator(X2_train_scaled)", "_____no_output_____" ], [ "X2_test = input_generator(X2_test_scaled)", "_____no_output_____" ], [ "X2_train = X2_train[:, :, 1:]\nX2_test = X2_test[:, :, 1:]", "_____no_output_____" ], [ "print(X2_train.shape)\nprint(X2_test.shape)", "(96298, 30, 39)\n(65242, 30, 39)\n" ] ], [ [ "# DL Model", "_____no_output_____" ], [ "CONV1D + BI-LSTM\n", "_____no_output_____" ] ], [ [ "\ninput_ = tfl.Input(shape=(X2_train.shape[1], X2_train.shape[2]))\nprint(input_.shape)\n\n#1st layer\nx = tfl.Conv1D(filters=16, kernel_size=3,padding = 'same')(input_)\nx = tfl.ReLU()(x)\n#x = tfl.BatchNormalization()(x)\nx = tfl.MaxPool1D(1)(x)\n\n#2nd layer\nx = tfl.Conv1D(filters=32, kernel_size=3, padding = 'same')(x)\nx = tfl.ReLU()(x)\n#x = tfl.BatchNormalization()(x)\nx = tfl.MaxPool1D(1)(x)\n\n#3nd layer\n#x = tfl.Conv1D(filters=64, kernel_size=3, padding = 'same')(x)\n#x = tfl.ReLU()(x)\n#x = tfl.BatchNormalization()(x)\n#x = tfl.MaxPool1D(1)(x)\n\nprint(x.shape)\n\nx=tfl.Bidirectional(tfl.LSTM(64, input_shape=(x.shape[1], x.shape[2])))(x)\n\nprint(x.shape)\n\n#3rd layer\n#x = tfl.Flatten()(x)\n#print(x.shape)\n#x = tfl.Dense(450,activation='selu',kernel_initializer='he_uniform',kernel_constraint=maxnorm(5))(x)\n#x = tfl.Dropout(0.7)(x)\n#x = tfl.ELU(alpha=1)(x)\nx = tfl.Dense(200,activation='selu',kernel_constraint=maxnorm(5))(x)\nx = tfl.Dropout(0.8)(x)\n#x = tfl.ELU(alpha=1)(x)\nx = tfl.Dense(20,activation='selu',kernel_constraint=maxnorm(5))(x)\nx = tfl.Dropout(0.5)(x)\n#x = tfl.ELU(alpha=1)(x)\noutput = tfl.Dense(1)(x)\n\nmodel = tf.keras.Model(inputs=input_, outputs=output)\n", "(None, 30, 39)\n(None, 30, 32)\n(None, 128)\n" ] ], [ [ "CONV2D + BI-LSTM", "_____no_output_____" ] ], [ [ "'''\ninput_ = tfl.Input(shape=(X2_train.shape[1], X2_train.shape[2]))\nprint(input_.shape)\n\nx=tf.keras.layers.Reshape((X2_train.shape[1], X2_train.shape[2],1), input_shape=(X2_train.shape[1], X2_train.shape[2]))(input_)\nprint(x.shape)\n#1st layer\nx = tfl.Conv2D(filters=16, kernel_size=3,padding = 'same')(x)\nx = tfl.ReLU()(x)\n#x = tfl.BatchNormalization()(x)\n#x = tfl.MaxPool2D(2)(x)\nprint(x.shape)\n\n#2nd layer\nx = tfl.Conv2D(filters=32, kernel_size=3, padding = 'same')(x)\nx = tfl.ReLU()(x)\n#x = tfl.BatchNormalization()(x)\n#x = tfl.MaxPool2D(2)(x)\nprint(x.shape)\nx = tfl.Reshape([x.shape[1],x.shape[2]*x.shape[3]])(x)\nprint(x.shape)\nx=tfl.Bidirectional(tfl.LSTM(256, return_sequences=True), input_shape=(x.shape[1], x.shape[2]))(x)\n\nprint(x.shape)\n#3rd layer\nx = tfl.Flatten()(x)\nprint(x.shape)\nx = tfl.Dense(1500,activation='selu',kernel_initializer='he_uniform',kernel_constraint=maxnorm(5))(x)\nx = tfl.Dropout(0.5)(x)\n#x = tfl.ELU(alpha=1)(x)\nx = tfl.Dense(800,activation='selu',kernel_constraint=maxnorm(5))(x)\nx = tfl.Dropout(0.5)(x)\n#x = tfl.ELU(alpha=1)(x)\nprint(x.shape)\nx = tfl.Dense(400,activation='selu',kernel_constraint=maxnorm(5))(x)\nx = tfl.Dropout(0.5)(x)\nx = tfl.Dense(100,activation='selu',kernel_constraint=maxnorm(5))(x)\nx = tfl.Dropout(0.5)(x)\nx = tfl.Dense(50,activation='selu',kernel_constraint=maxnorm(5))(x)\nx = tfl.Dropout(0.5)(x)\nprint(x.shape)\n#x = tfl.ELU(alpha=1)(x)\noutput = tfl.Dense(1)(x)\n\nmodel = tf.keras.Model(inputs=input_, outputs=output)\n'''", "(None, 30, 39)\n(None, 30, 39, 1)\n(None, 30, 39, 16)\n(None, 30, 39, 32)\n(None, 30, 1248)\n(None, 30, 512)\n(None, 15360)\n(None, 800)\n(None, 50)\n" ] ], [ [ "201 MSE AND REAL SCORE OF 72", "_____no_output_____" ] ], [ [ "# After 40 epochs this was over fitting\n'''\ninput_ = tfl.Input(shape=(X2_train.shape[1], X2_train.shape[2]))\nprint(input_.shape)\n\nx=tf.keras.layers.Reshape((X2_train.shape[1], X2_train.shape[2],1), input_shape=(X2_train.shape[1], X2_train.shape[2]))(input_)\nprint(x.shape)\n#1st layer\nx = tfl.Conv2D(filters=8, kernel_size=3,padding = 'same')(x)\nx = tfl.ReLU()(x)\n#x = tfl.BatchNormalization()(x)\n#x = tfl.MaxPool2D(2)(x)\nprint(x.shape)\n\n#2nd layer\nx = tfl.Conv2D(filters=16, kernel_size=3, padding = 'same')(x)\nx = tfl.ReLU()(x)\n#x = tfl.BatchNormalization()(x)\n#x = tfl.MaxPool2D(2)(x)\nprint(x.shape)\nx = tfl.Reshape([x.shape[1],x.shape[2]*x.shape[3]])(x)\nprint(x.shape)\nx=tfl.Bidirectional(tfl.LSTM(64, return_sequences=True), input_shape=(x.shape[1], x.shape[2]))(x)\n\nprint(x.shape)\n#3rd layer\nx = tfl.Flatten()(x)\nprint(x.shape)\nx = tfl.Dense(1000,activation='selu',kernel_initializer='he_uniform',kernel_constraint=maxnorm(5))(x)\nx = tfl.Dropout(0.5)(x)\n#x = tfl.ELU(alpha=1)(x)\nx = tfl.Dense(500,activation='selu',kernel_constraint=maxnorm(5))(x)\nx = tfl.Dropout(0.5)(x)\n#x = tfl.ELU(alpha=1)(x)\nprint(x.shape)\nx = tfl.Dense(200,activation='selu',kernel_constraint=maxnorm(5))(x)\nx = tfl.Dropout(0.5)(x)\nx = tfl.Dense(50,activation='selu',kernel_constraint=maxnorm(5))(x)\nx = tfl.Dropout(0.5)(x)\nprint(x.shape)\n#x = tfl.ELU(alpha=1)(x)\noutput = tfl.Dense(1)(x)\n\nmodel = tf.keras.Model(inputs=input_, outputs=output)\n'''", "_____no_output_____" ] ], [ [ "1D-CONVNET + BI-LSTM , gave MSE=193", "_____no_output_____" ] ], [ [ "'''\ndef conv1D_full(inputs):\n \n\n inputs=tf.expand_dims(inputs,axis=-1)\n #print(inputs.shape)\n x = Conv1D(filters=8, kernel_size=3, activation='selu', padding='same')(inputs)\n x = Conv1D(filters=16, kernel_size=3, activation='selu', padding='same')(x)\n #print(x.shape)\n #x = layers.MaxPooling1D(2)(x)\n # x = Conv1D(filters=32, kernel_size=3, activation='selu', padding='same')(x)\n # x = Conv1D(filters=64, kernel_size=3, activation='selu', padding='same')(x)\n # x = layers.MaxPooling1D(2)(x)\n #x = layers.Flatten()(x)\n #x = layers.Dense(200, activation='selu')(x)\n #print(x.shape)\n \n return x\n'''", "_____no_output_____" ], [ "'''\ndef multiple_cnn1D(nb):\n \n inputs = Input(shape=(30, nb))\n\n outputs = conv1D_full(inputs[:,:,0])\n \n for i in range(1,nb):\n x_i = conv1D_full(inputs[:,:,i])\n outputs = tf.concat([ outputs , x_i ] , 2 )\n #print(outputs.shape)\n \n x=Bidirectional(LSTM(64, return_sequences=True), input_shape=(outputs.shape[1],outputs.shape[2]))(outputs)\n #print(x.shape)\n x = Dropout(0.5)(x)\n x = layers.Flatten()(x)\n #print(x.shape)\n x = layers.Dense(1000, activation='selu',kernel_initializer='he_uniform',kernel_constraint=maxnorm(5))(x)\n x = Dropout(0.5)(x)\n x = layers.Dense(500, activation='selu',kernel_constraint=maxnorm(5))(x)\n x = Dropout(0.5)(x)\n x = layers.Dense(100,activation='selu',kernel_constraint=maxnorm(5))(x)\n x = Dropout(0.5)(x)\n x = layers.Dense(20,activation='selu',kernel_constraint=maxnorm(5))(x)\n x = Dropout(0.5)(x)\n answer = layers.Dense(1)(x)\n \n #print(answer.shape)\n \n model = Model(inputs, answer)\n opt = keras.optimizers.Adam(learning_rate=0.0001)\n model.compile(loss='mse', optimizer=opt, metrics=[tf.keras.metrics.MeanSquaredError()])\n print(model.summary())\n \n #dot_img_file = 'model_LSTM.png'\n #tf.keras.utils.plot_model(model, to_file=dot_img_file, show_shapes=True, rankdir='LR')\n \n return model\n'''", "_____no_output_____" ], [ "#model = multiple_cnn1D(39)", "_____no_output_____" ], [ "tf.keras.utils.plot_model(model, show_shapes=True)", "_____no_output_____" ], [ "model.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.001), loss='mse',metrics=[tf.keras.metrics.MeanSquaredError()])", "_____no_output_____" ], [ "#over_fitting after 30 epochs\nmodel.fit(x=X2_train, y=Y_train['Y2'].values, validation_split=0.1, epochs=25, batch_size=128)", "Epoch 1/25\n678/678 [==============================] - 48s 23ms/step - loss: 21753.3340 - mean_squared_error: 21753.3340 - val_loss: 16384.7363 - val_mean_squared_error: 16384.7363\nEpoch 2/25\n678/678 [==============================] - 15s 22ms/step - loss: 18758.3809 - mean_squared_error: 18758.3809 - val_loss: 16188.6855 - val_mean_squared_error: 16188.6855\nEpoch 3/25\n678/678 [==============================] - 15s 22ms/step - loss: 18467.6816 - mean_squared_error: 18467.6816 - val_loss: 15212.6357 - val_mean_squared_error: 15212.6357\nEpoch 4/25\n678/678 [==============================] - 15s 22ms/step - loss: 18085.6855 - mean_squared_error: 18085.6855 - val_loss: 15083.2480 - val_mean_squared_error: 15083.2480\nEpoch 5/25\n678/678 [==============================] - 15s 21ms/step - loss: 17878.2480 - mean_squared_error: 17878.2480 - val_loss: 15064.9453 - val_mean_squared_error: 15064.9453\nEpoch 6/25\n678/678 [==============================] - 14s 21ms/step - loss: 17864.5918 - mean_squared_error: 17864.5918 - val_loss: 15130.7295 - val_mean_squared_error: 15130.7295\nEpoch 7/25\n678/678 [==============================] - 15s 21ms/step - loss: 17619.2344 - mean_squared_error: 17619.2344 - val_loss: 14871.6777 - val_mean_squared_error: 14871.6777\nEpoch 8/25\n678/678 [==============================] - 15s 21ms/step - loss: 17501.4590 - mean_squared_error: 17501.4590 - val_loss: 14818.6436 - val_mean_squared_error: 14818.6436\nEpoch 9/25\n678/678 [==============================] - 14s 21ms/step - loss: 17241.2383 - mean_squared_error: 17241.2383 - val_loss: 15007.1445 - val_mean_squared_error: 15007.1445\nEpoch 10/25\n678/678 [==============================] - 14s 21ms/step - loss: 17172.0195 - mean_squared_error: 17172.0195 - val_loss: 14918.8896 - val_mean_squared_error: 14918.8896\nEpoch 11/25\n678/678 [==============================] - 15s 21ms/step - loss: 17065.1875 - mean_squared_error: 17065.1875 - val_loss: 15602.3496 - val_mean_squared_error: 15602.3496\nEpoch 12/25\n678/678 [==============================] - 15s 22ms/step - loss: 17015.2246 - mean_squared_error: 17015.2246 - val_loss: 14856.5234 - val_mean_squared_error: 14856.5234\nEpoch 13/25\n678/678 [==============================] - 14s 21ms/step - loss: 16849.3086 - mean_squared_error: 16849.3086 - val_loss: 14937.0088 - val_mean_squared_error: 14937.0088\nEpoch 14/25\n678/678 [==============================] - 15s 22ms/step - loss: 16842.2969 - mean_squared_error: 16842.2969 - val_loss: 14931.9082 - val_mean_squared_error: 14931.9082\nEpoch 15/25\n678/678 [==============================] - 15s 21ms/step - loss: 16735.1504 - mean_squared_error: 16735.1504 - val_loss: 15625.1816 - val_mean_squared_error: 15625.1816\nEpoch 16/25\n678/678 [==============================] - 15s 21ms/step - loss: 16541.0898 - mean_squared_error: 16541.0898 - val_loss: 14819.2451 - val_mean_squared_error: 14819.2451\nEpoch 17/25\n678/678 [==============================] - 14s 21ms/step - loss: 16413.6855 - mean_squared_error: 16413.6855 - val_loss: 15152.4258 - val_mean_squared_error: 15152.4258\nEpoch 18/25\n678/678 [==============================] - 14s 21ms/step - loss: 16619.7305 - mean_squared_error: 16619.7305 - val_loss: 15063.0830 - val_mean_squared_error: 15063.0830\nEpoch 19/25\n678/678 [==============================] - 14s 21ms/step - loss: 16303.7861 - mean_squared_error: 16303.7861 - val_loss: 15468.4316 - val_mean_squared_error: 15468.4316\nEpoch 20/25\n678/678 [==============================] - 15s 21ms/step - loss: 16297.4688 - mean_squared_error: 16297.4688 - val_loss: 15144.3145 - val_mean_squared_error: 15144.3145\nEpoch 21/25\n678/678 [==============================] - 14s 21ms/step - loss: 16166.4014 - mean_squared_error: 16166.4014 - val_loss: 15034.6465 - val_mean_squared_error: 15034.6465\nEpoch 22/25\n678/678 [==============================] - 14s 21ms/step - loss: 16130.2832 - mean_squared_error: 16130.2832 - val_loss: 15181.7490 - val_mean_squared_error: 15181.7490\nEpoch 23/25\n678/678 [==============================] - 14s 21ms/step - loss: 16185.5664 - mean_squared_error: 16185.5664 - val_loss: 15025.8477 - val_mean_squared_error: 15025.8477\nEpoch 24/25\n678/678 [==============================] - 14s 21ms/step - loss: 16005.9551 - mean_squared_error: 16005.9551 - val_loss: 14993.1797 - val_mean_squared_error: 14993.1797\nEpoch 25/25\n678/678 [==============================] - 14s 21ms/step - loss: 15820.0791 - mean_squared_error: 15820.0791 - val_loss: 15127.5879 - val_mean_squared_error: 15127.5879\n" ], [ "pd.DataFrame(model.history.history).plot()", "_____no_output_____" ], [ "predict_y2 = model.predict(X2_test)", "_____no_output_____" ], [ "predict_y2", "_____no_output_____" ], [ "predict_y2.shape", "_____no_output_____" ] ], [ [ "# Simple Model", "_____no_output_____" ] ], [ [ "X1_train = X_train.groupby('UNIQUE_IDENTIFIER').mean()\nX1_test = test_onehot.groupby('UNIQUE_IDENTIFIER').mean()", "_____no_output_____" ], [ "X1_train.head()", "_____no_output_____" ], [ "scaler1 = StandardScaler()\nX1_scaled = scaler1.fit_transform(X1_train)\nX1_test_scaled = scaler1.transform(X1_test)", "_____no_output_____" ], [ "input1 = tfl.Input(shape=(X1_scaled.shape[1]))\n#1st layer\nx = tfl.Dense(256, kernel_initializer=\"glorot_normal\", kernel_regularizer=\"l2\")(input1)\nx = tfl.BatchNormalization()(x)\nx = tfl.ELU(alpha=1)(x)\nx = tfl.Dropout(0.5)(x)\n\n# 2nd layer\nx = tfl.Dense(128, kernel_initializer=\"glorot_normal\", kernel_regularizer=\"l2\")(x)\nx = tfl.BatchNormalization()(x)\nx = tfl.ELU(alpha=1)(x)\nx = tfl.Dropout(0.5)(x)\n\n# # 3rd layer\n# x = tfl.Dense(64, kernel_initializer=\"glorot_normal\", kernel_regularizer=\"l2\")(x)\n# x = tfl.BatchNormalization()(x)\n# x = tfl.ELU(alpha=1)(x)\n\n# #4th layer\n# x = tfl.Dense(32, kernel_initializer=\"glorot_normal\", kernel_regularizer=\"l2\")(x)\n# x = tfl.BatchNormalization()(x)\n# x = tfl.ELU(alpha=1)(x)\n\n# 5th layer\noutput1 = tfl.Dense(1, kernel_initializer=\"glorot_normal\", kernel_regularizer=\"l2\")(x)\n\nmodel1 = tf.keras.Model(inputs=input1, outputs=output1)", "_____no_output_____" ], [ "model1.compile(optimizer=tf.keras.optimizers.Adam(learning_rate=0.0001), loss=\"mse\")", "_____no_output_____" ], [ "model1.fit(x=X1_scaled, y=Y_train['Y1'].values, epochs=100, batch_size=512, validation_split=0.1)", "Epoch 1/100\n170/170 [==============================] - 2s 6ms/step - loss: 94.0625 - val_loss: 101.9454\nEpoch 2/100\n170/170 [==============================] - 1s 5ms/step - loss: 83.4638 - val_loss: 92.3889\nEpoch 3/100\n170/170 [==============================] - 1s 5ms/step - loss: 76.6696 - val_loss: 86.3388\nEpoch 4/100\n170/170 [==============================] - 1s 5ms/step - loss: 72.8523 - val_loss: 81.6360\nEpoch 5/100\n170/170 [==============================] - 1s 5ms/step - loss: 68.7738 - val_loss: 77.6942\nEpoch 6/100\n170/170 [==============================] - 1s 5ms/step - loss: 65.3674 - val_loss: 74.2461\nEpoch 7/100\n170/170 [==============================] - 1s 5ms/step - loss: 62.3842 - val_loss: 71.5178\nEpoch 8/100\n170/170 [==============================] - 1s 5ms/step - loss: 60.6828 - val_loss: 68.8069\nEpoch 9/100\n170/170 [==============================] - 1s 5ms/step - loss: 58.4525 - val_loss: 66.6117\nEpoch 10/100\n170/170 [==============================] - 1s 5ms/step - loss: 57.1322 - val_loss: 64.9800\nEpoch 11/100\n170/170 [==============================] - 1s 5ms/step - loss: 56.0076 - val_loss: 63.2920\nEpoch 12/100\n170/170 [==============================] - 1s 5ms/step - loss: 55.1823 - val_loss: 62.3347\nEpoch 13/100\n170/170 [==============================] - 1s 5ms/step - loss: 54.7657 - val_loss: 61.2984\nEpoch 14/100\n170/170 [==============================] - 1s 5ms/step - loss: 54.5430 - val_loss: 60.6496\nEpoch 15/100\n170/170 [==============================] - 1s 5ms/step - loss: 53.6069 - val_loss: 59.9471\nEpoch 16/100\n170/170 [==============================] - 1s 6ms/step - loss: 52.4687 - val_loss: 59.2707\nEpoch 17/100\n170/170 [==============================] - 1s 5ms/step - loss: 52.5070 - val_loss: 58.9770\nEpoch 18/100\n170/170 [==============================] - 1s 5ms/step - loss: 52.7909 - val_loss: 58.6403\nEpoch 19/100\n170/170 [==============================] - 1s 5ms/step - loss: 53.2483 - val_loss: 58.3310\nEpoch 20/100\n170/170 [==============================] - 1s 5ms/step - loss: 51.6651 - val_loss: 58.1917\nEpoch 21/100\n170/170 [==============================] - 1s 5ms/step - loss: 52.4711 - val_loss: 57.8703\nEpoch 22/100\n170/170 [==============================] - 1s 5ms/step - loss: 52.3295 - val_loss: 57.5261\nEpoch 23/100\n170/170 [==============================] - 1s 6ms/step - loss: 52.0615 - val_loss: 57.4424\nEpoch 24/100\n170/170 [==============================] - 1s 5ms/step - loss: 52.1578 - val_loss: 57.4241\nEpoch 25/100\n170/170 [==============================] - 1s 6ms/step - loss: 51.0860 - val_loss: 57.0847\nEpoch 26/100\n170/170 [==============================] - 1s 6ms/step - loss: 52.0336 - val_loss: 56.9726\nEpoch 27/100\n170/170 [==============================] - 1s 5ms/step - loss: 51.1071 - val_loss: 56.8809\nEpoch 28/100\n170/170 [==============================] - 1s 5ms/step - loss: 50.9523 - val_loss: 56.8582\nEpoch 29/100\n170/170 [==============================] - 1s 5ms/step - loss: 50.0660 - val_loss: 56.6695\nEpoch 30/100\n170/170 [==============================] - 1s 5ms/step - loss: 51.8585 - val_loss: 56.6303\nEpoch 31/100\n170/170 [==============================] - 1s 5ms/step - loss: 50.8192 - val_loss: 56.4105\nEpoch 32/100\n170/170 [==============================] - 1s 5ms/step - loss: 50.5959 - val_loss: 56.3727\nEpoch 33/100\n170/170 [==============================] - 1s 5ms/step - loss: 50.9725 - val_loss: 56.1655\nEpoch 34/100\n170/170 [==============================] - 1s 5ms/step - loss: 50.6479 - val_loss: 55.9169\nEpoch 35/100\n170/170 [==============================] - 1s 5ms/step - loss: 50.2262 - val_loss: 56.0999\nEpoch 36/100\n170/170 [==============================] - 1s 5ms/step - loss: 50.6265 - val_loss: 56.1445\nEpoch 37/100\n170/170 [==============================] - 1s 5ms/step - loss: 50.1574 - val_loss: 56.0224\nEpoch 38/100\n170/170 [==============================] - 1s 5ms/step - loss: 50.8434 - val_loss: 55.7468\nEpoch 39/100\n170/170 [==============================] - 1s 6ms/step - loss: 49.3496 - val_loss: 55.7265\nEpoch 40/100\n170/170 [==============================] - 1s 5ms/step - loss: 50.4982 - val_loss: 55.7247\nEpoch 41/100\n170/170 [==============================] - 1s 5ms/step - loss: 50.1722 - val_loss: 55.5869\nEpoch 42/100\n170/170 [==============================] - 1s 5ms/step - loss: 50.4376 - val_loss: 55.3469\nEpoch 43/100\n170/170 [==============================] - 1s 5ms/step - loss: 49.3263 - val_loss: 55.6514\nEpoch 44/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.9222 - val_loss: 55.4182\nEpoch 45/100\n170/170 [==============================] - 1s 5ms/step - loss: 49.2236 - val_loss: 55.5216\nEpoch 46/100\n170/170 [==============================] - 1s 5ms/step - loss: 49.3543 - val_loss: 55.4481\nEpoch 47/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.9134 - val_loss: 55.1329\nEpoch 48/100\n170/170 [==============================] - 1s 6ms/step - loss: 49.7145 - val_loss: 55.1939\nEpoch 49/100\n170/170 [==============================] - 1s 5ms/step - loss: 49.3065 - val_loss: 55.0590\nEpoch 50/100\n170/170 [==============================] - 1s 5ms/step - loss: 49.6363 - val_loss: 55.0260\nEpoch 51/100\n170/170 [==============================] - 1s 5ms/step - loss: 49.0960 - val_loss: 55.1413\nEpoch 52/100\n170/170 [==============================] - 1s 5ms/step - loss: 49.1435 - val_loss: 55.0633\nEpoch 53/100\n170/170 [==============================] - 1s 5ms/step - loss: 49.3744 - val_loss: 55.0692\nEpoch 54/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.7209 - val_loss: 54.9497\nEpoch 55/100\n170/170 [==============================] - 1s 5ms/step - loss: 49.5476 - val_loss: 55.1135\nEpoch 56/100\n170/170 [==============================] - 1s 5ms/step - loss: 49.4020 - val_loss: 54.9188\nEpoch 57/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.2114 - val_loss: 54.9081\nEpoch 58/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.5659 - val_loss: 54.7956\nEpoch 59/100\n170/170 [==============================] - 1s 6ms/step - loss: 48.0586 - val_loss: 54.9634\nEpoch 60/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.8762 - val_loss: 54.8859\nEpoch 61/100\n170/170 [==============================] - 1s 6ms/step - loss: 48.7398 - val_loss: 54.7957\nEpoch 62/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.3987 - val_loss: 54.9234\nEpoch 63/100\n170/170 [==============================] - 1s 6ms/step - loss: 47.8236 - val_loss: 54.6880\nEpoch 64/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.8389 - val_loss: 54.9171\nEpoch 65/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.5694 - val_loss: 54.9113\nEpoch 66/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.6297 - val_loss: 55.0087\nEpoch 67/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.8245 - val_loss: 54.7958\nEpoch 68/100\n170/170 [==============================] - 1s 5ms/step - loss: 47.5334 - val_loss: 54.6693\nEpoch 69/100\n170/170 [==============================] - 1s 5ms/step - loss: 47.0342 - val_loss: 54.7005\nEpoch 70/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.2733 - val_loss: 54.5422\nEpoch 71/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.1769 - val_loss: 54.5524\nEpoch 72/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.7116 - val_loss: 54.6058\nEpoch 73/100\n170/170 [==============================] - 1s 5ms/step - loss: 47.9591 - val_loss: 54.4939\nEpoch 74/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.8330 - val_loss: 54.8089\nEpoch 75/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.7857 - val_loss: 54.6737\nEpoch 76/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.4693 - val_loss: 54.7285\nEpoch 77/100\n170/170 [==============================] - 1s 5ms/step - loss: 47.3919 - val_loss: 54.7417\nEpoch 78/100\n170/170 [==============================] - 1s 5ms/step - loss: 47.6757 - val_loss: 54.6092\nEpoch 79/100\n170/170 [==============================] - 1s 6ms/step - loss: 48.1317 - val_loss: 54.4587\nEpoch 80/100\n170/170 [==============================] - 1s 5ms/step - loss: 47.6940 - val_loss: 54.7272\nEpoch 81/100\n170/170 [==============================] - 1s 5ms/step - loss: 47.6293 - val_loss: 54.5535\nEpoch 82/100\n170/170 [==============================] - 1s 6ms/step - loss: 47.6296 - val_loss: 54.7563\nEpoch 83/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.0261 - val_loss: 54.6735\nEpoch 84/100\n170/170 [==============================] - 1s 5ms/step - loss: 46.6133 - val_loss: 54.5821\nEpoch 85/100\n170/170 [==============================] - 1s 6ms/step - loss: 48.4482 - val_loss: 54.4499\nEpoch 86/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.0261 - val_loss: 54.5914\nEpoch 87/100\n170/170 [==============================] - 1s 5ms/step - loss: 47.3326 - val_loss: 54.4471\nEpoch 88/100\n170/170 [==============================] - 1s 5ms/step - loss: 47.7484 - val_loss: 54.6473\nEpoch 89/100\n170/170 [==============================] - 1s 5ms/step - loss: 47.0888 - val_loss: 54.5780\nEpoch 90/100\n170/170 [==============================] - 1s 6ms/step - loss: 47.4789 - val_loss: 54.5326\nEpoch 91/100\n170/170 [==============================] - 1s 6ms/step - loss: 47.7369 - val_loss: 54.6960\nEpoch 92/100\n170/170 [==============================] - 1s 5ms/step - loss: 48.3427 - val_loss: 54.4589\nEpoch 93/100\n170/170 [==============================] - 1s 6ms/step - loss: 47.7092 - val_loss: 54.5723\nEpoch 94/100\n170/170 [==============================] - 1s 6ms/step - loss: 48.1318 - val_loss: 54.5318\nEpoch 95/100\n170/170 [==============================] - 1s 5ms/step - loss: 47.6784 - val_loss: 54.3874\nEpoch 96/100\n170/170 [==============================] - 1s 6ms/step - loss: 46.8542 - val_loss: 54.3588\nEpoch 97/100\n170/170 [==============================] - 1s 6ms/step - loss: 47.6580 - val_loss: 54.7165\nEpoch 98/100\n170/170 [==============================] - 1s 5ms/step - loss: 46.8936 - val_loss: 54.4203\nEpoch 99/100\n170/170 [==============================] - 1s 6ms/step - loss: 47.3634 - val_loss: 54.5099\nEpoch 100/100\n170/170 [==============================] - 1s 5ms/step - loss: 46.8520 - val_loss: 54.2735\n" ], [ "pd.DataFrame(model1.history.history).plot()", "_____no_output_____" ], [ "predict_y1 = model1.predict(X1_test_scaled)", "_____no_output_____" ], [ "predict_y1", "_____no_output_____" ], [ "customer_index = test_onehot['UNIQUE_IDENTIFIER'].unique()", "_____no_output_____" ], [ "customer_index = np.expand_dims(customer_index, axis=1)", "_____no_output_____" ], [ "customer_index", "_____no_output_____" ], [ "predictions = pd.DataFrame(np.hstack((customer_index, predict_y1, predict_y2)), columns=['UNIQUE_IDENTIFIER', 'Y1', 'Y2'])", "_____no_output_____" ], [ "predictions", "_____no_output_____" ], [ "predictions['UNIQUE_IDENTIFIER'] = predictions['UNIQUE_IDENTIFIER'].astype(\"int64\")", "_____no_output_____" ], [ "predictions", "_____no_output_____" ], [ "predict=pd.DataFrame(model.predict(X2_train))\npredict1=pd.DataFrame(model1.predict(X1_scaled))\npred=pd.concat([predict, predict1], axis=1)\nexact=pd.DataFrame(pred)\n\nfrom sklearn.metrics import mean_squared_error\nm1 = np.sqrt(mean_squared_error(Y_train.iloc[:,1],exact.iloc[:,0]))\nm2 = np.sqrt(mean_squared_error(Y_train.iloc[:,0],exact.iloc[:,1]))\nprint((m1+m2)/2)", "61.68752353976654\n" ], [ "predictions.to_csv(\"1D_Bi_LSTM_61_modified.csv\", index=False)", "_____no_output_____" ] ] ]

[ "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad664a44045a3a3a79dd76800d6c30131831e25

303,685

ipynb

Jupyter Notebook

ml/ML_in_business_hw3.ipynb

QuackingKnox/geekbrains

1f95d970429729513667981df7f6edc805a6509e

[ "MIT" ]

null

ml/ML_in_business_hw3.ipynb

QuackingKnox/geekbrains

1f95d970429729513667981df7f6edc805a6509e

[ "MIT" ]

null

ml/ML_in_business_hw3.ipynb

QuackingKnox/geekbrains

1f95d970429729513667981df7f6edc805a6509e

[ "MIT" ]

null

233.245008

49,574

0.890854

[ [ [ "### Домашняя работа №3", "_____no_output_____" ] ], [ [ "import numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nimport itertools\n\n\nfrom sklearn.feature_extraction.text import TfidfVectorizer\nfrom sklearn.linear_model import LogisticRegression\nfrom sklearn.model_selection import cross_val_score, train_test_split\nfrom sklearn.metrics import precision_recall_curve, roc_curve, roc_auc_score, confusion_matrix, log_loss\nfrom sklearn.pipeline import Pipeline, make_pipeline, FeatureUnion\nfrom sklearn.base import BaseEstimator, TransformerMixin\nfrom sklearn.preprocessing import StandardScaler\nfrom sklearn.tree import DecisionTreeClassifier\nfrom sklearn.ensemble import RandomForestClassifier, AdaBoostClassifier\n\n\nfrom scipy.sparse import hstack", "_____no_output_____" ], [ "def plot_confusion_matrix(cm, classes,\n normalize=False,\n title='Confusion matrix',\n cmap=plt.cm.Blues):\n \"\"\"\n This function prints and plots the confusion matrix.\n Normalization can be applied by setting `normalize=True`.\n \"\"\"\n plt.imshow(cm, interpolation='nearest', cmap=cmap)\n plt.title(title)\n plt.colorbar()\n tick_marks = np.arange(len(classes))\n plt.xticks(tick_marks, classes, rotation=45)\n plt.yticks(tick_marks, classes)\n\n if normalize:\n cm = cm.astype('float') / cm.sum(axis=1)[:, np.newaxis]\n print(\"Normalized confusion matrix\")\n else:\n print('Confusion matrix, without normalization')\n\n print(cm)\n\n thresh = cm.max() / 2.\n for i, j in itertools.product(range(cm.shape[0]), range(cm.shape[1])):\n plt.text(j, i, cm[i, j],\n horizontalalignment=\"center\",\n color=\"white\" if cm[i, j] > thresh else \"black\")\n\n plt.tight_layout()\n plt.ylabel('True label')\n plt.xlabel('Predicted label')", "_____no_output_____" ], [ "class ColumnSelector(BaseEstimator, TransformerMixin):\n \"\"\"\n Transformer to select a single column from the data frame to perform additional transformations on\n \"\"\"\n def __init__(self, key):\n self.key = key\n\n def fit(self, X, y=None):\n return self\n\n def transform(self, X):\n return X[self.key]\n \nclass NumberSelector(BaseEstimator, TransformerMixin):\n \"\"\"\n Transformer to select a single column from the data frame to perform additional transformations on\n Use on numeric columns in the data\n \"\"\"\n def __init__(self, key):\n self.key = key\n\n def fit(self, X, y=None):\n return self\n\n def transform(self, X):\n return X[[self.key]]\n \nclass OHEEncoder(BaseEstimator, TransformerMixin):\n def __init__(self, key):\n self.key = key\n self.columns = []\n\n def fit(self, X, y=None):\n self.columns = [col for col in pd.get_dummies(X, prefix=self.key).columns]\n return self\n\n def transform(self, X):\n X = pd.get_dummies(X, prefix=self.key)\n test_columns = [col for col in X.columns]\n for col_ in test_columns:\n if col_ not in self.columns:\n X[col_] = 0\n return X[self.columns]", "_____no_output_____" ] ], [ [ "В рамках конкурса вам нужно предсказать наличие сердечно-сосудистых заболеваний по результатам классического врачебного осмотра. Датасет сформирован из 100.000 реальных клинических анализов, и в нём используются признаки, которые можно разбить на 3 группы:\n\n \n\nОбъективные признаки:\n\n - Возраст\n - Рост\n - Вес\n - Пол\n \n\nРезультаты измерения:\n\n - Артериальное давление верхнее и нижнее\n - Холестерин\n - Глюкоза\n \n\nСубъективные признаки:\n\n - Курение\n - Употребление Алкоголя\n - Физическая активность\n \n\nВозраст дан в днях. Значения показателей холестерина и глюкозы представлены одним из трех классов: норма, выше нормы, значительно выше нормы. Значения субъективных признаков — бинарны.\n\nВсе показатели даны на момент осмотра.", "_____no_output_____" ] ], [ [ "from google.colab import drive\ndrive.mount('/gdrive')", "Drive already mounted at /gdrive; to attempt to forcibly remount, call drive.mount(\"/gdrive\", force_remount=True).\n" ], [ "df = pd.read_csv('train_case2.csv', ';')\ndf.head()", "_____no_output_____" ], [ "#разделим данные на train/test\nX_train, X_test, y_train, y_test = train_test_split(df.drop('cardio', 1), \n df['cardio'], random_state=0)", "_____no_output_____" ] ], [ [ "К полям:\n- gender, cholesterol применим OHE-кодирование\n- age, height, weight, ap_hi, ap_lo - standardScaler\n- gluc, smoke, alco, active - оставим пока как есть", "_____no_output_____" ] ], [ [ "continuos_cols = ['age', 'height', 'weight', 'ap_hi', 'ap_lo']\ncat_cols = ['gender', 'cholesterol']\nbase_cols = ['gluc', 'smoke', 'alco', 'active']\n\ncontinuos_transformers = []\ncat_transformers = []\nbase_transformers = []\n\nfor cont_col in continuos_cols:\n transfomer = Pipeline([\n ('selector', NumberSelector(key=cont_col)),\n ('standard', StandardScaler())\n ])\n continuos_transformers.append((cont_col, transfomer))\n \nfor cat_col in cat_cols:\n cat_transformer = Pipeline([\n ('selector', ColumnSelector(key=cat_col)),\n ('ohe', OHEEncoder(key=cat_col))\n ])\n cat_transformers.append((cat_col, cat_transformer))\n \nfor base_col in base_cols:\n base_transformer = Pipeline([\n ('selector', NumberSelector(key=base_col))\n ])\n base_transformers.append((base_col, base_transformer))", "_____no_output_____" ] ], [ [ "Теперь объединим все наши трансформеры с помощью FeatureUnion", "_____no_output_____" ] ], [ [ "feats = FeatureUnion(continuos_transformers+cat_transformers+base_transformers)\nfeature_processing = Pipeline([('feats', feats)])\n\nfeature_processing.fit_transform(X_train)", "_____no_output_____" ] ], [ [ "## LogisticRegression", "_____no_output_____" ] ], [ [ "classifier = Pipeline([\n ('features',feats),\n ('classifier', LogisticRegression(random_state = 42)),\n])\n\n\n#запустим кросс-валидацию\ncv_scores = cross_val_score(classifier, X_train, y_train, cv=7, scoring='roc_auc')\ncv_score = np.mean(cv_scores)\ncv_score_std = np.std(cv_scores)\nprint('CV score is {}+-{}'.format(cv_score, cv_score_std))\n\n#обучим пайплайн на всем тренировочном датасете\nclassifier.fit(X_train, y_train)\ny_score = classifier.predict_proba(X_test)[:, 1]", "CV score is 0.7864573689384385+-0.004422021036885763\n" ] ], [ [ "Посчитаем precision/recall/f_score", "_____no_output_____" ] ], [ [ "b=1\nprecision, recall, thresholds = precision_recall_curve(y_test.values, y_score)\nfscore = (1+b**2)*(precision * recall) / (b**2*precision + recall)\n# locate the index of the largest f score\nix = np.argmax(fscore)\nprint('Best Threshold=%f, F-Score=%.3f, Precision=%.3f, Recall=%.3f' % (thresholds[ix], \n fscore[ix],\n precision[ix],\n recall[ix]))", "Best Threshold=0.386937, F-Score=0.730, Precision=0.647, Recall=0.838\n" ] ], [ [ "Нарисуем roc auc кривую (кстати, наверное неплохо бы ее вынести в отдельную функцию)", "_____no_output_____" ] ], [ [ "sns.set(font_scale=1.5)\nsns.set_color_codes(\"muted\")\n\nplt.figure(figsize=(10, 8))\nfpr, tpr, thresholds_ = roc_curve(y_test, y_score, pos_label=1)\nlw = 2\nplt.plot(fpr, tpr, lw=lw, label='ROC curve ')\nplt.plot([0, 1], [0, 1])\nplt.xlim([0.0, 1.0])\nplt.ylim([0.0, 1.05])\nplt.xlabel('False Positive Rate')\nplt.ylabel('True Positive Rate')\nplt.title('ROC curve')\nplt.savefig(\"ROC.png\")\nplt.show()", "_____no_output_____" ], [ "roc_auc_score_lr = roc_auc_score(y_true=y_test, y_score=classifier.predict_proba(X_test)[:,1])\nlog_loss_lr = log_loss(y_true=y_test, y_pred=classifier.predict_proba(X_test)[:,1])\nprint(\"roc auc score: {}\".format(roc_auc_score_lr))\nprint(\"log loss score: {}\".format(log_loss_lr))", "roc auc score: 0.7840347790421852\nlog loss score: 0.5779604008230668\n" ] ], [ [ "Посомтрим на матрицу ошибок", "_____no_output_____" ] ], [ [ "#мы уже нашли ранее \"оптимальный\" порог, когда максимизировали f_score\nfont = {'size' : 15}\n\nplt.rc('font', **font)\n\ncnf_matrix = confusion_matrix(y_test, y_score>thresholds[ix])\nplt.figure(figsize=(10, 8))\nplot_confusion_matrix(cnf_matrix, classes=['cardio_0', 'cardio_1'],\n title='Confusion matrix')\nplt.savefig(\"conf_matrix.png\")\nplt.show()", "Confusion matrix, without normalization\n[[4861 3959]\n [1411 7269]]\n" ] ], [ [ "Посчитаем FPR, TPR", "_____no_output_____" ] ], [ [ "TN = cnf_matrix[0][0]\nFN = cnf_matrix[1][0]\nTP = cnf_matrix[1][1]\nFP = cnf_matrix[0][1]\n\nTPR = TP/(TP+FN)\nFPR = FP/(FP+TN)\nTNR = TN/(FP+TN)\nTPR, FPR, TNR", "_____no_output_____" ] ], [ [ "False Positive Rate довольно высокий ~ 0.45.\n\nЭто означает, что 45 процентов всех пациентов получат метку 1 при том, что они на самом деле здоровы", "_____no_output_____" ], [ "### Random Forest", "_____no_output_____" ] ], [ [ "classifier_tree = Pipeline([\n ('features',feats),\n ('classifier', RandomForestClassifier(random_state = 42)),\n])\n\n\n#запустим кросс-валидацию\ncv_scores_tree = cross_val_score(classifier_tree, X_train, y_train, cv=7, scoring='roc_auc')\ncv_score_tree = np.mean(cv_scores_tree)\ncv_score_std_tree = np.std(cv_scores_tree)\nprint('CV score is {}+-{}'.format(cv_score_tree, cv_score_std_tree))\n\n#обучим пайплайн на всем тренировочном датасете\nclassifier_tree.fit(X_train, y_train)\ny_score_tree = classifier_tree.predict_proba(X_test)[:, 1]", "CV score is 0.7743796633622809+-0.003611330100912611\n" ], [ "b=2\nprecision_tree, recall_tree, thresholds_tree = precision_recall_curve(y_test.values, y_score_tree)\nfscore_tree = (1+b**2)*(precision_tree * recall_tree) / (b**2*precision_tree + recall_tree)\n# locate the index of the largest f score\nix_tree = np.argmax(fscore_tree)\nprint('Best Threshold=%f, F-Score=%.3f, Precision=%.3f, Recall=%.3f' % (thresholds_tree[ix_tree], \n fscore_tree[ix_tree],\n precision_tree[ix_tree],\n recall_tree[ix_tree]))", "Best Threshold=0.076667, F-Score=0.833, Precision=0.521, Recall=0.981\n" ], [ "sns.set(font_scale=1.5)\nsns.set_color_codes(\"muted\")\n\nplt.figure(figsize=(10, 8))\nfpr, tpr, thresholds_tree = roc_curve(y_test, y_score_tree, pos_label=1)\nlw = 2\nplt.plot(fpr, tpr, lw=lw, label='ROC curve ')\nplt.plot([0, 1], [0, 1])\nplt.xlim([0.0, 1.0])\nplt.ylim([0.0, 1.05])\nplt.xlabel('False Positive Rate')\nplt.ylabel('True Positive Rate')\nplt.title('ROC curve')\nplt.savefig(\"ROC.png\")\nplt.show()", "_____no_output_____" ], [ "roc_auc_score_tree = roc_auc_score(y_true=y_test, y_score=classifier_tree.predict_proba(X_test)[:,1])\nlog_loss_tree = log_loss(y_true=y_test, y_pred=classifier_tree.predict_proba(X_test)[:,1])\nprint(\"roc auc score: {}\".format(roc_auc_score_tree))\nprint(\"log loss score: {}\".format(log_loss_tree))", "roc auc score: 0.7710366181802983\nlog loss score: 0.5992984853728378\n" ], [ "font = {'size' : 15}\n\nplt.rc('font', **font)\n\ncnf_matrix_tree = confusion_matrix(y_test, y_score>thresholds_tree[ix_tree])\nplt.figure(figsize=(10, 8))\nplot_confusion_matrix(cnf_matrix_tree, classes=['cardio_0', 'cardio_1'],\n title='Confusion matrix')\nplt.savefig(\"conf_matrix_tree.png\")\nplt.show()", "Confusion matrix, without normalization\n[[8672 148]\n [7747 933]]\n" ], [ "TN_tree = cnf_matrix_tree[0][0]\nFN_tree = cnf_matrix_tree[1][0]\nTP_tree = cnf_matrix_tree[1][1]\nFP_tree = cnf_matrix_tree[0][1]\n\nTPR_tree = TP_tree/(TP_tree+FN_tree)\nFPR_tree = FP_tree/(FP_tree+TN_tree)\nTNR_tree = TN_tree/(FP_tree+TN_tree)\nTPR_tree, FPR_tree, TNR_tree", "_____no_output_____" ], [ "classifier_ada = Pipeline([\n ('features',feats),\n ('classifier', AdaBoostClassifier(random_state = 42)),\n])\n\n\n#запустим кросс-валидацию\ncv_scores_ada = cross_val_score(classifier_ada, X_train, y_train, cv=7, scoring='roc_auc')\ncv_score_ada = np.mean(cv_scores_ada)\ncv_score_std_ada = np.std(cv_scores_ada)\nprint('CV score is {}+-{}'.format(cv_score_ada, cv_score_std_ada))\n\n#обучим пайплайн на всем тренировочном датасете\nclassifier_ada.fit(X_train, y_train)\ny_score_ada = classifier_ada.predict_proba(X_test)[:, 1]", "CV score is 0.7947234142489508+-0.003603068933381472\n" ], [ "b=1\nprecision_ada, recall_ada, thresholds_ada = precision_recall_curve(y_test.values, y_score_ada)\nfscore_ada = (1+b**2)*(precision_ada * recall_ada) / (b**2*precision_ada + recall_ada)\n# locate the index of the largest f score\nix_ada = np.argmax(fscore_ada)\nprint('Best Threshold=%f, F-Score=%.3f, Precision=%.3f, Recall=%.3f' % (thresholds_ada[ix_ada], \n fscore_ada[ix_ada],\n precision_ada[ix_ada],\n recall_ada[ix_ada]))", "Best Threshold=0.497430, F-Score=0.738, Precision=0.692, Recall=0.789\n" ], [ "sns.set(font_scale=1.5)\nsns.set_color_codes(\"muted\")\n\nplt.figure(figsize=(10, 8))\nfpr, tpr, thresholds_ada = roc_curve(y_test, y_score_ada, pos_label=1)\nlw = 2\nplt.plot(fpr, tpr, lw=lw, label='ROC curve ')\nplt.plot([0, 1], [0, 1])\nplt.xlim([0.0, 1.0])\nplt.ylim([0.0, 1.05])\nplt.xlabel('False Positive Rate')\nplt.ylabel('True Positive Rate')\nplt.title('ROC curve')\nplt.savefig(\"ROC.png\")\nplt.show()", "_____no_output_____" ], [ "roc_auc_score_ada = roc_auc_score(y_true=y_test, y_score=classifier_ada.predict_proba(X_test)[:,1])\nlog_loss_ada = log_loss(y_true=y_test, y_pred=classifier_ada.predict_proba(X_test)[:,1])\nprint(\"roc auc score: {}\".format(roc_auc_score_ada))\nprint(\"log loss score: {}\".format(log_loss_ada))", "roc auc score: 0.7945722371129712\nlog loss score: 0.6869766413646277\n" ], [ "font = {'size' : 15}\n\nplt.rc('font', **font)\n\ncnf_matrix_ada = confusion_matrix(y_test, y_score>thresholds_ada[ix_ada])\nplt.figure(figsize=(10, 8))\nplot_confusion_matrix(cnf_matrix_ada, classes=['cardio_0', 'cardio_1'],\n title='Confusion matrix')\nplt.savefig(\"conf_matrix_ada.png\")\nplt.show()", "Confusion matrix, without normalization\n[[6838 1982]\n [2875 5805]]\n" ], [ "TN_ada = cnf_matrix_ada[0][0]\nFN_ada = cnf_matrix_ada[1][0]\nTP_ada = cnf_matrix_ada[1][1]\nFP_ada = cnf_matrix_ada[0][1]\n\nTPR_ada = TP_ada/(TP_ada+FN_ada)\nFPR_ada = FP_ada/(FP_ada+TN_ada)\nTNR_ada = TN_ada/(FP_ada+TN_ada)\nTPR_ada, FPR_ada, TNR_ada", "_____no_output_____" ], [ "print(f'\\t\\tLogisticRegression \\n')\nprint('F-Score=%.3f, Precision=%.3f, Recall=%.3f, ROC AUC=%.3f' % (fscore[ix],\n precision[ix],\n recall[ix], roc_auc_score_lr))\nprint(f'TPR {TPR} FPR {FPR} TNR {TNR}')\nprint(f'\\n\\t\\tRandomForest \\n')\nprint('F-Score=%.3f, Precision=%.3f, Recall=%.3f, ROC AUC=%.3f' % (fscore_tree[ix_tree],\n precision_tree[ix_tree],\n recall_tree[ix_tree], roc_auc_score_tree))\nprint(f'TPR {TPR_tree} FPR {FPR_tree} TNR {TNR_tree}')\nprint(f'\\n\\t\\tAdaBoosting \\n')\nprint('F-Score=%.3f, Precision=%.3f, Recall=%.3f, ROC AUC=%.3f' % (fscore_ada[ix_ada],\n precision_ada[ix_ada],\n recall_ada[ix_ada], roc_auc_score_ada))\nprint(f'TPR {TPR_ada} FPR {FPR_ada} TNR {TNR_ada}')", "\t\tLogisticRegression \n\nF-Score=0.730, Precision=0.647, Recall=0.838, ROC AUC=0.784\nTPR 0.837442396313364 FPR 0.44886621315192743 TNR 0.5511337868480726\n\n\t\tRandomForest \n\nF-Score=0.833, Precision=0.521, Recall=0.981, ROC AUC=0.771\nTPR 0.10748847926267281 FPR 0.016780045351473923 TNR 0.9832199546485261\n\n\t\tAdaBoosting \n\nF-Score=0.738, Precision=0.692, Recall=0.789, ROC AUC=0.795\nTPR 0.668778801843318 FPR 0.22471655328798185 TNR 0.7752834467120181\n" ] ], [ [ "Делая вывод по данной работе, хочу сказать, что по моему мнению самая лучшая модель - модель обученная с помощью AdaBoosting. RandomForest выдает немного странные результаты, не понимаю почему идет такой сильный дисбаланс классов ответа.", "_____no_output_____" ], [ "При сильном дисбалансе классов roc_auc_curve подходит больше, потому что True Positive Rate и False Positive Rate устойчивы при несбалансированных данных.", "_____no_output_____" ] ], [ [ "", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code" ] ]

4ad665062cceb29eb22f2ca43a87b587385207ec

11,390

ipynb

Jupyter Notebook

Examples/Interpolation.ipynb

DrFriedrich/nmfce-2021-22

2ccee5a97b24bd5c1e80e531957240ffb7163897

[ "MIT" ]

null

Examples/Interpolation.ipynb

DrFriedrich/nmfce-2021-22

2ccee5a97b24bd5c1e80e531957240ffb7163897

[ "MIT" ]

null

Examples/Interpolation.ipynb

DrFriedrich/nmfce-2021-22

2ccee5a97b24bd5c1e80e531957240ffb7163897

[ "MIT" ]

null

31.464088

283

0.544425

[ [ [ "# Polynomial interpolation\n\n---\n\nPerform polynomial interpolation of air density from the data in the following table.\n$$\n\\begin{aligned}\n& \\text {Table with air density against temperature}\\\\\n&\\begin{array}{c|c}\n Temperature & Density \\\\\n ^\\circ\\,C & kg\\,m^{-3} \\\\\n \\hline\n 100 & 0.946 \\\\\n 150 & 0.835 \\\\\n 200 & 0.746 \\\\\n 250 & 0.675 \\\\\n 300 & 0.616 \\\\\n 400 & 0.525 \\\\\n 500 & 0.457\n\\end{array}\n\\end{aligned}\n$$", "_____no_output_____" ] ], [ [ "import numpy as np\nfrom IPython.display import display, Math\n\n# Input the data\n# Temperature\nT = [100, 150, 200, 250, 300, 400, 500] \n# Air density\nrho = [0.946, 0.835, 0.746, 0.675, 0.616, 0.525, 0.457]", "_____no_output_____" ], [ "# Plot the data\nimport matplotlib.pyplot as plt\n\nplt.plot(T, rho, 'g*')\nplt.xlabel(\"Temperature [$^\\circ\\,$C]\")\nplt.ylabel(\"Air density [kg$\\,$m$^{-3}$]\")\nplt.show()", "_____no_output_____" ] ], [ [ "## Part a)\n\nUse Lagrange interpolation to calculate the air density at $350^\\circ\\,$C from the measured data between $300^\\circ\\,$C and $500^\\circ\\,$C.", "_____no_output_____" ] ], [ [ "# Temperature at which the air density is sought\nT0 = 350\n\n# Form Lagrange multipliers\nL = []\nos = 4 # Offset to get to the correct position in the data\nfor i in range(3):\n tmp = 1\n for j in range(3):\n if i != j:\n tmp = tmp * (T0 - T[j+os])/(T[i+os] - T[j+os])\n L.append(tmp) \n\n\n# Calculate the air density at T0\nrho_T0 = 0\nfor i in range(3):\n rho_T0 = rho_T0 + L[i] * rho[i+os]\nrho_T0\ndisplay(Math(r\"\\text{{The air density at }} {} ^\\circ \\text{{C is }}{:0.4f} \\,kg\\, m^{{-3}}\".format(T0, rho_T0)))", "_____no_output_____" ] ], [ [ "*Remark:* You should try a few different values and add the value to the plot.", "_____no_output_____" ], [ "## Part b)\n\nCalculate the Newton interpolation coefficients for the support points $300^\\circ\\,$C and $400^\\circ\\,$C. Add a third support point and calculate the air density at $350^\\circ\\,$C.", "_____no_output_____" ] ], [ [ "# Temperature at which the air density is sought\nT0 = 350\n# Offset to get to the correct position in the data\nos = 4 \n\n# Calculate the Newton interpolation coefficients\na = []\na.append(rho[0+os])\na.append((rho[1+os] - rho[0+os]) / (T[1+os] - T[0+os]))", "_____no_output_____" ], [ "# Calculate the air density at T0\nrho_T0 = a[0] + a[1] * (T0 - T[0+os])\ndisplay(Math(r\"\\text{{The air density at }} {} ^\\circ \\text{{C is }}{:0.4f} \\,kg\\, m^{{-3}}\".format(T0, rho_T0)))", "_____no_output_____" ] ], [ [ "The air density at $T=350^\\circ\\,$C calculated by Newton interpolation with two support points is $\\rho=0.5705\\,$kg$\\,$m$^{-3}$.\n\nNow we add a third interpolation point.", "_____no_output_____" ] ], [ [ "# Add a third interpolation points\n# Set to -1 for T=250 and to 2 for T=500\nidx = -1\n\ntmp = (rho[idx+os]- rho[1+os]) / (T[idx+os] - T[1+os])\na.append((tmp - a[1]) / (T[idx+os] - T[0+os]))\n\nrho_T0_3rd = a[0] + a[1] * (T0 - T[0+os]) + a[2] * (T0 - T[0+os]) * (T0 - T[1+os])\ndisplay(Math(r\"\\text{{The air density at }} {} ^\\circ \\text{{C is }}{:0.4f} \\,kg\\, m^{{-3}}\".format(T0, rho_T0_3rd)))", "_____no_output_____" ] ], [ [ " When we add a third support point at $T=500^\\circ\\,$C we get the same result as for the case of the Lagrange interpolation in part (a), i.e. $\\rho=0.5676\\,$kg$\\,$m$^{-3}$. This is expected because the second order polynomial through the three support points is unique.\n\n*Remark:* When we instead add a third support point at $T=250^\\circ\\,$C we get $\\rho=0.5660\\,$kg$\\,$m$^{-3}$ which is slightly different from the other two interpolations. You should try this.", "_____no_output_____" ], [ "## Part c)\n\nUse the Python functions numpy.interp to interpolate the air density between $300^\\circ\\,$C and $500^\\circ\\,$C. Plot the interpolated air density and the measured air densities at the three support points.", "_____no_output_____" ] ], [ [ "import numpy as np\nT0 = 350\n\n# What are the next two lines doing?\nos = 4\nlength = 3\n\nrho_T0_np = np.interp(T0, T[os:os+length], rho[os:os+length])\ndisplay(Math(r\"\\text{{The air density at }} {} ^\\circ \\text{{C is }}{:0.4f} \\,kg\\, m^{{-3}}\".format(T0, rho_T0_np)))", "_____no_output_____" ], [ "x = np.linspace(300, 500, 201)\n\nplt.plot(T[os:os+length], rho[os:os+length], 'x', label=\"Measured points\")\nplt.plot(x, np.interp(x, T[os:os+length], rho[os:os+length]), 'r', label=\"Interpolation\")\nplt.xlabel(\"Temperature [$^\\circ\\,$C]\")\nplt.ylabel(\"Air density [kg$\\,$m$^{-3}$]\")\nplt.legend()\nplt.show()", "_____no_output_____" ] ], [ [ "The plot shows that the numpy.interp function uses only linear interpolation which we could have expected from the value at $T=350^\\circ\\,$C or if we had looked at the documentation: https://numpy.org/doc/stable/reference/generated/numpy.interp.html", "_____no_output_____" ], [ "## Part d) \n\nUse the Python functions scipy.interpolate.interp1d to interpolate the air density between $300^\\circ\\,$C and $500^\\circ\\,$C. Plot the interpolated air density and the measured air densities at the three support points.", "_____no_output_____" ] ], [ [ "from scipy import interpolate\n\nos = 4\nlength = 3\nf = interpolate.interp1d(T[os:os+length], rho[os:os+length], \"quadratic\", fill_value=\"extrapolate\")\nx = np.linspace(300, 500, 201)\n\nplt.plot(T[os:os+length], rho[os:os+length], 'x', label=\"Measured points\")\nplt.plot(x, f(x), 'r', label=\"Interpolation\")\nplt.xlabel(\"Temperature [$^\\circ\\,$C]\")\nplt.ylabel(\"Air density [kg$\\,$m$^{-3}$]\")\nplt.legend()\nplt.show()\ndisplay(Math(r\"\\text{{The air density at }} {} ^\\circ \\text{{C is }}{:0.4f} \\,kg\\, m^{{-3}}\".format(T0, f(T0))))", "_____no_output_____" ] ], [ [ "We can see that this interpolation is quadratic and produces the same interpolated value as the two examples in parts a) and the second interpolation in part b).", "_____no_output_____" ], [ "## Part e)\n\nCalculate the air density at $200^\\circ\\,$C from the interpolation and compare the value to the measured value. Extend the plot from part d) to $200^\\circ\\,$C.", "_____no_output_____" ] ], [ [ "from scipy import interpolate\n# Offset\nos = 2 \n# length\nlength = 5\n\nT0 = 200\nx = np.linspace(200, 500, 301)\n\nplt.plot(T[os:os+length], rho[os:os+length], 'x', label=\"Measured points\")\nplt.plot(x, f(x), 'r', label=\"Interpolation\")\nplt.xlabel(\"Temperature [$^\\circ\\,$C]\")\nplt.ylabel(\"Air density [kg$\\,$m$^{-3}$]\")\nplt.legend()\nplt.show()\ndisplay(Math(r\"\\text{{The air density at }} {} ^\\circ \\text{{C is }}{:0.4f} \\,kg\\, m^{{-3}}\".format(T0, f(T0))))", "_____no_output_____" ] ], [ [ "The interpolated value is $\\rho_{i,200}=0.73$ compared to the true value of $\\rho_{200}=0.746$. This gives a relative error of\n$$\ne = \\frac{0.746-0.73}{0.746} = 0.021\n$$\nIn this case, the extrapolation beyond the support point interval gives a reasonably good approximation. This is due to the fact that the exact data almost follows a quadratic curve.", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ] ]

4ad66a6b94dcefaa43ef7120e48ef0fedac9e090

3,982

ipynb

Jupyter Notebook

devel/python_vs_julia.ipynb

mjirik/LarSurf.jl

de2eaec62dfe8c63e7d621bc973aa01d8de019c6

[ "MIT" ]

2

2019-09-17T22:56:08.000Z

2020-01-04T09:50:42.000Z

devel/python_vs_julia.ipynb

mjirik/lario3d.jl

de2eaec62dfe8c63e7d621bc973aa01d8de019c6

[ "MIT" ]

1

2019-11-16T15:47:22.000Z

2019-11-18T17:43:46.000Z

devel/python_vs_julia.ipynb

mjirik/lario3d.jl

de2eaec62dfe8c63e7d621bc973aa01d8de019c6

[ "MIT" ]

1

2021-03-05T15:01:47.000Z

21.408602

287

0.503014

[ [ [ "def fib(n):\n if n<2:\n return n\n return fib(n-1)+fib(n-2)\n\n%timeit fib(20)", "3.68 ms ± 387 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)\n" ], [ "%load_ext Cython", "_____no_output_____" ], [ "\n%%cython\n\ncpdef long fib_cython_type(long n):\n if n<2:\n return n\n return fib_cython_type(n-1)+fib_cython_type(n-2)", "_____no_output_____" ], [ "from functools import lru_cache as cache\n\n@cache(maxsize=None)\ndef fib_cache(n):\n if n<2:\n return n\n return fib_cache(n-1)+fib_cache(n-2)\n%timeit fib_cache(20)", "95.3 ns ± 3.35 ns per loop (mean ± std. dev. of 7 runs, 10000000 loops each)\n" ], [ "# comprehensions", "_____no_output_____" ], [ "def add_one_for(n):\n some_list = []\n\n for i in range(n):\n some_list.append(i+1)\n%timeit add_one_for(100000)", "9.26 ms ± 115 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)\n" ], [ "def add_one_comp(n):\n some_list = [f(i) for i in range(n)]\n return some_list\n%timeit add_one_comp(100000)", "11.3 ms ± 202 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)\n" ], [ "\ndef add_one_to_mat(mat):\n if type(mat) == str:\n mat = read_file(mat)\n return mat + 1\n", "_____no_output_____" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad6740d522f29a2718e423eb37f00fd4cb9082c

2,055

ipynb

Jupyter Notebook

face extractor.ipynb

Shreyansh-Gupta/Open-contributions

e72a9ce2b0aa6a48081921bf8138b91ad259c422

[ "MIT" ]

61

2020-09-10T05:16:19.000Z

2021-11-07T00:22:46.000Z

face extractor.ipynb

Shreyansh-Gupta/Open-contributions

e72a9ce2b0aa6a48081921bf8138b91ad259c422

[ "MIT" ]

72

2020-09-12T09:34:19.000Z

2021-08-01T17:48:46.000Z

face extractor.ipynb

Shreyansh-Gupta/Open-contributions

e72a9ce2b0aa6a48081921bf8138b91ad259c422

[ "MIT" ]

571

2020-09-10T01:52:56.000Z

2022-03-26T17:26:23.000Z

26.688312

100

0.513869

[ [ [ "import cv2\nimport numpy as np\n\nface_classifier = cv2.CascadeClassifier('Haarcascades/haarcascade_frontalface_default.xml')\neye_classifier = cv2.CascadeClassifier('Haarcascades/haarcascade_eye.xml')\n\ndef face_detector(img, size=1):\n # Convert image to grayscale\n gray = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)\n faces = face_classifier.detectMultiScale(gray, 1.3, 5)\n if faces is ():\n return img\n \n for (x,y,w,h) in faces:\n cv2.rectangle(img,(x,y),(x+w,y+h),(255,0,0),2)\n roi_gray = gray[y:y+h, x:x+w]\n roi_color = img[y:y+h, x:x+w]\n eyes = eye_classifier.detectMultiScale(roi_gray)\n \n for (ex,ey,ew,eh) in eyes:\n cv2.rectangle(roi_color,(ex,ey),(ex+ew,ey+eh),(0,0,255),2) \n \n roi_color = cv2.flip(roi_color,1)\n return roi_color\n\ncap = cv2.VideoCapture(0)\n\nwhile True:\n\n ret, frame = cap.read()\n cv2.imshow('Our Face Extractor', face_detector(frame))\n if cv2.waitKey(1) == 13: #13 is the Enter Key\n break\n \ncap.release()\ncv2.destroyAllWindows() ", "_____no_output_____" ] ] ]

[ "code" ]

[ [ "code" ] ]

4ad6893116710549c690e0a818a734a34044a437

776,775

ipynb

Jupyter Notebook

etl/url_download.ipynb

conny-lin/python_lib

f31a93aab7331b5d1112db4282e5ec71c93f8869

[ "MIT" ]

null

etl/url_download.ipynb

conny-lin/python_lib

f31a93aab7331b5d1112db4282e5ec71c93f8869

[ "MIT" ]

null

etl/url_download.ipynb

conny-lin/python_lib

f31a93aab7331b5d1112db4282e5ec71c93f8869

[ "MIT" ]

null

3,119.578313

383,650

0.749298

[ [ [ "200000*1000\n", "_____no_output_____" ], [ "import requests\nimport numpy as np\nimport pandas as pd\n\n", "_____no_output_____" ], [ "url='https://www.dropbox.com/s/3e1kc6ofmafija2/allfilepaths.pickle?dl=0'\nr = requests.get(url, allow_redirects=True)\nr", "_____no_output_____" ], [ "r.json", "_____no_output_____" ], [ "import urllib.request\nimport pickle\n\nurl= 'https://www.dropbox.com/s/2zhpbusn7re4j08/legend_drunkposture.pickle?dl=1'\npickle.load(urllib.request.urlopen(url))\n", "_____no_output_____" ], [ "\npd.read_pickle(r.text)", "_____no_output_____" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code", "code", "code" ] ]

4ad69f9dd752662fd3d8888f95f243ced2ae2ae2

30,568

ipynb

Jupyter Notebook

nickel/.ipynb_checkpoints/A05_Bernstein_Vazirani_Algorithm_Solutions-checkpoint.ipynb

asif-saad/qNickel

34f60436769fe13b288706fb6abcfa084042a807

[ "Apache-2.0", "CC-BY-4.0" ]

null

nickel/.ipynb_checkpoints/A05_Bernstein_Vazirani_Algorithm_Solutions-checkpoint.ipynb

asif-saad/qNickel

34f60436769fe13b288706fb6abcfa084042a807

[ "Apache-2.0", "CC-BY-4.0" ]

null

nickel/.ipynb_checkpoints/A05_Bernstein_Vazirani_Algorithm_Solutions-checkpoint.ipynb

asif-saad/qNickel

34f60436769fe13b288706fb6abcfa084042a807

[ "Apache-2.0", "CC-BY-4.0" ]

null

121.784861

22,356

0.843398

[ [ [ "<table> <tr>\n <td style=\"background-color:#ffffff;\">\n <a href=\"http://qworld.lu.lv\" target=\"_blank\"><img src=\"..\\images\\qworld.jpg\" width=\"25%\" align=\"left\"> </a></td>\n <td style=\"background-color:#ffffff;vertical-align:bottom;text-align:right;\">\n prepared by Berat Yenilen, Utku Birkan, Arda Çınar and Özlem Salehi (<a href=\"http://qworld.lu.lv/index.php/qturkey/\" target=\"_blank\">QTurkey</a>)\n </td> \n</tr></table>", "_____no_output_____" ], [ "<table width=\"100%\"><tr><td style=\"color:#bbbbbb;background-color:#ffffff;font-size:11px;font-style:italic;text-align:right;\">This cell contains some macros. If there is a problem with displaying mathematical formulas, please run this cell to load these macros. </td></tr></table>\n$ \\newcommand{\\bra}[1]{\\langle #1|} $\n$ \\newcommand{\\ket}[1]{|#1\\rangle} $\n$ \\newcommand{\\braket}[2]{\\langle #1|#2\\rangle} $\n$ \\newcommand{\\dot}[2]{ #1 \\cdot #2} $\n$ \\newcommand{\\biginner}[2]{\\left\\langle #1,#2\\right\\rangle} $\n$ \\newcommand{\\mymatrix}[2]{\\left( \\begin{array}{#1} #2\\end{array} \\right)} $\n$ \\newcommand{\\myvector}[1]{\\mymatrix{c}{#1}} $\n$ \\newcommand{\\myrvector}[1]{\\mymatrix{r}{#1}} $\n$ \\newcommand{\\mypar}[1]{\\left( #1 \\right)} $\n$ \\newcommand{\\mybigpar}[1]{ \\Big( #1 \\Big)} $\n$ \\newcommand{\\sqrttwo}{\\frac{1}{\\sqrt{2}}} $\n$ \\newcommand{\\dsqrttwo}{\\dfrac{1}{\\sqrt{2}}} $\n$ \\newcommand{\\onehalf}{\\frac{1}{2}} $\n$ \\newcommand{\\donehalf}{\\dfrac{1}{2}} $\n$ \\newcommand{\\hadamard}{ \\mymatrix{rr}{ \\sqrttwo & \\sqrttwo \\\\ \\sqrttwo & -\\sqrttwo }} $\n$ \\newcommand{\\vzero}{\\myvector{1\\\\0}} $\n$ \\newcommand{\\vone}{\\myvector{0\\\\1}} $\n$ \\newcommand{\\stateplus}{\\myvector{ \\sqrttwo \\\\ \\sqrttwo } } $\n$ \\newcommand{\\stateminus}{ \\myrvector{ \\sqrttwo \\\\ -\\sqrttwo } } $\n$ \\newcommand{\\myarray}[2]{ \\begin{array}{#1}#2\\end{array}} $\n$ \\newcommand{\\X}{ \\mymatrix{cc}{0 & 1 \\\\ 1 & 0} } $\n$ \\newcommand{\\Z}{ \\mymatrix{rr}{1 & 0 \\\\ 0 & -1} } $\n$ \\newcommand{\\Htwo}{ \\mymatrix{rrrr}{ \\frac{1}{2} & \\frac{1}{2} & \\frac{1}{2} & \\frac{1}{2} \\\\ \\frac{1}{2} & -\\frac{1}{2} & \\frac{1}{2} & -\\frac{1}{2} \\\\ \\frac{1}{2} & \\frac{1}{2} & -\\frac{1}{2} & -\\frac{1}{2} \\\\ \\frac{1}{2} & -\\frac{1}{2} & -\\frac{1}{2} & \\frac{1}{2} } } $\n$ \\newcommand{\\CNOT}{ \\mymatrix{cccc}{1 & 0 & 0 & 0 \\\\ 0 & 1 & 0 & 0 \\\\ 0 & 0 & 0 & 1 \\\\ 0 & 0 & 1 & 0} } $\n$ \\newcommand{\\norm}[1]{ \\left\\lVert #1 \\right\\rVert } $\n$ \\newcommand{\\pstate}[1]{ \\lceil \\mspace{-1mu} #1 \\mspace{-1.5mu} \\rfloor } $", "_____no_output_____" ], [ "\n<h1> Solutions for Bernstein-Vazirani Problem </h1> ", "_____no_output_____" ], [ "<a id=\"task1\"></a>\n### Task 1\n\n- Using how many queries can you solve the problem clasically? How many queries if you use a probabilistic algorithm?\n- How many queries do you think we need to make if we are to solve the problem with a quantum computer? ", "_____no_output_____" ], [ "<h3>Solution</h3>\n\nLet's illustrate the solution over an example.\n\nLet $n$ = 4 and let's make the following queries to function $f$.\n\n\\begin{align*}\nf(1000) &= s_1.1 + s_2.0 + s_3.0 + s_4.0 = s_1\\\\\nf(0100) &= s_1.0 + s_2.1 + s_3.0 + s_4.0 = s_2\\\\\nf(0010) &= s_1.0 + s_2.0 + s_3.1 + s_4.0 = s_3\\\\\nf(0001) &= s_1.0 + s_2.0 + s_3.0 + s_4.1 = s_4\\\\\n\\end{align*}\n\nWe need $n$ queries and this is an optimal way of solving this problem by classical and probabilistic\nalgorithms. For further information about why classical and probabilistic algorithms can not perform better, please refer\nto Information Theory.", "_____no_output_____" ], [ "<a id=\"task2\"></a>\n### Task 2\n\nWhat can we say about the $f:\\{0,1\\}^n \\rightarrow \\{0,1\\}$ function if $s = 0^n$?", "_____no_output_____" ], [ "<h3>Solution</h3>\n\nIf $s=0^n$, then $f(x)=0$ for all $x$.", "_____no_output_____" ], [ "<a id=\"task3\"></a>\n### Task 3\n\nGiven an oracle function `bv_oracle()` that constructs a 6 qubit oracle circuit ($s$ has length 5) for $f$,construct a circuit that implements the algorithm described above to find out $s$.\n\nNote that qubit 5 is the output qubit.\n\nRun the following cell to load function `bv_oracle()`.", "_____no_output_____" ] ], [ [ "%run ../include/oracle.py", "_____no_output_____" ], [ "from qiskit import QuantumCircuit, execute, Aer\n\nn=5 \n\n#Create quantum circuit\nbv_circuit = QuantumCircuit(n+1, n)\n\n#Apply X gate to last qubit\nbv_circuit.x(n)\n\n#Apply Hadamard to all qubits\nbv_circuit.h(range(n+1))\n\n#Apply oracle\nbv_circuit += bv_oracle()\n\n#Apply Hadamard to all qubits\nbv_circuit.h(range(n))\n\n#Measure the first 4 qubits\nbv_circuit.measure(range(n), range(n))\n\n#Draw the circuit\nbv_circuit.draw(output=\"mpl\")", "_____no_output_____" ], [ "job = execute(bv_circuit, Aer.get_backend('qasm_simulator'),shots=10000)\ncounts = job.result().get_counts()\nfor outcome in counts:\n reverse_outcome = ''\n for i in outcome:\n reverse_outcome = i + reverse_outcome\n print(reverse_outcome,\"is observed\",counts[outcome],\"times\")", "01011 is observed 10000 times\n" ] ], [ [ "<a id=\"task4\"></a>\n### Task 4\n\nGiven $\\textbf{s} = 0110$, implement a function that returns an oracle for the function $ f(\\mathbf{x}) = \\mathbf{x} \\cdot \\mathbf{s} $. Note that $n=4$ and you will need a cirucit with 5 qubits where qubit 4 is the output qubit.", "_____no_output_____" ] ], [ [ "from qiskit import QuantumCircuit\n\ndef oracle():\n circuit = QuantumCircuit(5)\n circuit.barrier()\n \n circuit.cx(1, 4)\n circuit.cx(2, 4)\n \n circuit.barrier()\n return circuit", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ] ]

4ad6a6cdf7edb5866e94a11967cb1f084ea541db

5,649

ipynb

Jupyter Notebook

for-scripters/Python/jupyter-bridge-basic.ipynb

kozo2/cytoscape-automation

2252651795e0f38a46fd2e02afbb36a01e3c6bf3

[ "CC0-1.0" ]

null

for-scripters/Python/jupyter-bridge-basic.ipynb

kozo2/cytoscape-automation

2252651795e0f38a46fd2e02afbb36a01e3c6bf3

[ "CC0-1.0" ]

null

for-scripters/Python/jupyter-bridge-basic.ipynb

kozo2/cytoscape-automation

2252651795e0f38a46fd2e02afbb36a01e3c6bf3

[ "CC0-1.0" ]

null

32.653179

641

0.641175

[ [ [ "# Jupyter Bridge Basic\n## Yihang Xin and Alex Pico\n## 2021-04-04", "_____no_output_____" ], [ "# Why use Jupyter Bridge\n* Users do not need to worry about dependencies and environment.\n* Easily share notebook-based workflows and data sets\n* Workflows can reside in the cloud, access cloud resources, and yet still use Cytoscape features.\n", "_____no_output_____" ], [ "# How Jupyter Bridge works\nJupyter-Bridge enables a workflow running on remote Jupyter to execute functions on a PC-local Cytoscape – the remote Jupyter runs the request through Jupyter-Bridge, where it is picked up by Javascript code running on the Jupyter web page in the PC-local browser, which in turn calls Cytoscape. The Cytoscape response travels the reverse route.", "_____no_output_____" ], [ "![](https://raw.githubusercontent.com/cytoscape/jupyter-bridge/678fe9eab9d8c85d3fbb6ba879fe87b87d8e12b9/docs/images/Figure%202.svg)\n", "_____no_output_____" ], [ "Jupyter-Bridge allows a remote Jupyter Notebook to communicate with a workstation-based Cytoscape as if the Notebook were running on the Cytoscape workstation. A Jupyter Notebook passes a Cytoscape call to an independent Jupyter-Bridge server where it’s picked up by the Jupyter-Bridge browser component and is passed to Cytoscape. The Cytoscape response is returned via the opposite flow. As a result, workflows can reside in the cloud, access cloud resources, and yet still leverage Cytoscape features. Jupyter Bridge supports py4cytoscape first, and now RCy3 (R library for communicating with Cytoscape) also support Jupyter-Bridge.", "_____no_output_____" ], [ "Visit the [source code of Juputer Bridge](https://github.com/cytoscape/jupyter-bridge)\n for more information.", "_____no_output_____" ], [ "# Prerequisites (Local machine)\n## In addition to this package (py4cytoscape latest version 0.0.8), you will need:\n\n* Latest version of Cytoscape, which can be downloaded from https://cytoscape.org/download.html. Simply follow the installation instructions on screen.\n* Complete installation wizard\n* Launch Cytoscape\n* Install the filetransfer app from https://apps.cytoscape.org/apps/filetransfer\n", "_____no_output_____" ], [ "\n# Prerequisites (Cloud server)\n\nThere are a lot of cloud computing services online, such as Google Colab, Amazon EMR Notebook, Microsoft Azure, CoCalc and your own JupyterHub. You can choose your favorite one.\n\nHere we use Google Colab to demonstrate. Click this [link](https://colab.research.google.com/notebooks/empty.ipynb) to create a new empty Python notebook in the Google colab. \n\nCopy codes below to build connection between Jupyter notebook (cloud) and Cytoscape (local).\n\n Make sure to run code below in the cloud!!!", "_____no_output_____" ] ], [ [ "# Install and import required packages\n%%capture\n!pip install py4cytoscape\nimport IPython\nimport py4cytoscape as p4c", "_____no_output_____" ], [ "# Build connection between the cloud jupyter notebookand local Cytoscape\nbrowser_client_js = p4c.get_browser_client_js()\nIPython.display.Javascript(browser_client_js)", "_____no_output_____" ] ], [ [ "Then, launch Cytoscape and keep it running whenever using py4cytoscape and Jupyter Bridge. Confirm that you have everything installed and running:", "_____no_output_____" ] ], [ [ "# Check connection, you should see cytoscape version infomation\np4c.cytoscape_version_info()", "_____no_output_____" ] ], [ [ "Done! Now you can execute a workflow in a remote server-based Jupyter Notebook to leverage your workstation’s Cytoscape. You can also easily share notebook-based workflows and data sets.", "_____no_output_____" ], [ "For Jupyter Bridge workflow use case example, visit the Jupyter Bridge workflow notebook.", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ] ]

4ad6bd739f544e495c4433fbf6512fff73dedca8

188,121

ipynb

Jupyter Notebook

notebooks/stream_workspace.ipynb

norfordb/groundmotion

3f714894a34d9d37e1ac236f26b4366e25a05056

[ "CC0-1.0" ]

1

2019-08-16T16:11:23.000Z

notebooks/stream_workspace.ipynb

vinceq-usgs/gmprocess

2812dfd1803f8b2e0622c56cd8373b1c3fedb1a6

[ "CC0-1.0" ]

null

notebooks/stream_workspace.ipynb

vinceq-usgs/gmprocess

2812dfd1803f8b2e0622c56cd8373b1c3fedb1a6

[ "CC0-1.0" ]

null

375.491018

103,576

0.925197

[ [ [ "%matplotlib inline", "_____no_output_____" ], [ "import os.path\nimport pprint\n\nimport pandas as pd\n\nfrom gmprocess.io.asdf.stream_workspace import StreamWorkspace\nfrom gmprocess.io.test_utils import read_data_dir\nfrom gmprocess.io.read import read_data\nfrom gmprocess.streamcollection import StreamCollection\nfrom gmprocess.processing import process_streams\nfrom gmprocess.event import get_event_object\nfrom gmprocess.logging import setup_logger\n\n# Only log errors; this suppresses many warnings that are\n# distracting and not important.\nsetup_logger(level='error')", "_____no_output_____" ] ], [ [ "## Reading Data\nWe currently have a few different ways of reading in data. Here we use the `read_data_dir` helper function to quickly read streams and event (i.e., origin) information from the testing data in this repository.", "_____no_output_____" ] ], [ [ "datafiles, origin = read_data_dir('geonet', 'us1000778i', '*.V1A')", "_____no_output_____" ] ], [ [ "The read_data below finds the appropriate data reader for the format supplied.", "_____no_output_____" ] ], [ [ "tstreams = []\nfor dfile in datafiles:\n tstreams += read_data(dfile)", "_____no_output_____" ] ], [ [ "Note that `tstreams` is just a list of StationStream objects:", "_____no_output_____" ] ], [ [ "print(type(tstreams))\nprint(type(tstreams[0]))", "<class 'list'>\n<class 'gmprocess.stationstream.StationStream'>\n" ] ], [ [ "## gmprocess Subclasses of Obspy Classes\n\nThe `StationStream` class is a subclass of ObsPy's `Stream` class, which is effectively a list of `StationTrace` objects. The `StationTrace` class is, in turn, a subclass of ObsPy's `Trace` class. The motivation for these subclasses is primarily to enforce certain required metadata in the Trace stats dictionary (that ObsPy would generally store in their `Inventory` object).\n\nWe also have a `StreamCollection` class taht is effectively a list of `StationStream` objects, and enforces some rules that are required later for processing, such forcing all `StationTraces` in a `StationStream` be from the same network/station. The basic constructor for the StreamCollection class takes a list of streams:", "_____no_output_____" ] ], [ [ "sc = StreamCollection(tstreams)", "_____no_output_____" ] ], [ [ "The StreamCollection print method gives the number of StationStreams and the number that have passed/failed processing checks. Since we have not done any processing, all StationStreams should pass checks.", "_____no_output_____" ] ], [ [ "print(sc)", "3 StationStreams(s) in StreamCollection:\n 3 StationStreams(s) passed checks.\n 0 StationStreams(s) failed checks.\n\n" ] ], [ [ "More detailed information about the StreamCollection is given by the `describe` method:", "_____no_output_____" ] ], [ [ "sc.describe()", "3 StationStreams(s) in StreamCollection:\n 3 StationTrace(s) in StationStream (passed):\n NZ.HSES.--.HN1 | 2016-11-13T11:02:20.000000Z - 2016-11-13T11:07:47.675000Z | 200.0 Hz, 65536 samples (passed)\n NZ.HSES.--.HN2 | 2016-11-13T11:02:20.000000Z - 2016-11-13T11:07:47.675000Z | 200.0 Hz, 65536 samples (passed)\n NZ.HSES.--.HNZ | 2016-11-13T11:02:20.000000Z - 2016-11-13T11:07:47.675000Z | 200.0 Hz, 65536 samples (passed)\n 3 StationTrace(s) in StationStream (passed):\n NZ.WTMC.--.HN1 | 2016-11-13T11:02:19.000001Z - 2016-11-13T11:07:46.675001Z | 200.0 Hz, 65536 samples (passed)\n NZ.WTMC.--.HN2 | 2016-11-13T11:02:19.000001Z - 2016-11-13T11:07:46.675001Z | 200.0 Hz, 65536 samples (passed)\n NZ.WTMC.--.HNZ | 2016-11-13T11:02:19.000001Z - 2016-11-13T11:07:46.675001Z | 200.0 Hz, 65536 samples (passed)\n 3 StationTrace(s) in StationStream (passed):\n NZ.THZ.--.HN2 | 2016-11-13T11:02:33.000000Z - 2016-11-13T11:07:28.995000Z | 200.0 Hz, 59200 samples (passed)\n NZ.THZ.--.HN1 | 2016-11-13T11:02:33.000000Z - 2016-11-13T11:07:28.995000Z | 200.0 Hz, 59200 samples (passed)\n NZ.THZ.--.HNZ | 2016-11-13T11:02:33.000000Z - 2016-11-13T11:07:28.995000Z | 200.0 Hz, 59200 samples (passed)\n\n" ] ], [ [ "## Processing\nNote that processing options can be controlled in a config file that is installed in the user's home directory (`~/.gmprocess/config.yml`) and that event/origin information is required for processing:", "_____no_output_____" ] ], [ [ "pprint.pprint(origin)\nsc_processed = process_streams(sc, origin)\nprint(sc_processed)", "{'depth': 15.11,\n 'id': 'us1000778i',\n 'lat': -42.7373,\n 'lon': 173.054,\n 'magnitude': 7.8,\n 'time': '2016-11-13T11:02:56.340000'}\n3 StationStreams(s) in StreamCollection:\n 3 StationStreams(s) passed checks.\n 0 StationStreams(s) failed checks.\n\n" ] ], [ [ "Note that all checks have passed. When a stream does not pass a check, it is not deleted, but marked as failed and subsequent processing is aborted.\n\nProcessing steps are recorded according to the SEIS-PROV standard for each StationTrace. We log this information as a list of dictionaries, where each dictionary has keys `prov_id` and `prov_attributes`. This can be retrieved from each traces with the `getAllProvenance` method:", "_____no_output_____" ] ], [ [ "pprint.pprint(sc_processed[0][0].getAllProvenance())", "[{'prov_attributes': {'input_units': 'counts', 'output_units': 'cm/s^2'},\n 'prov_id': 'remove_response'},\n {'prov_attributes': {'detrending_method': 'linear'}, 'prov_id': 'detrend'},\n {'prov_attributes': {'detrending_method': 'demean'}, 'prov_id': 'detrend'},\n {'prov_attributes': {'new_end_time': UTCDateTime(2016, 11, 13, 11, 5, 44, 175000),\n 'new_start_time': UTCDateTime(2016, 11, 13, 11, 2, 58, 655000)},\n 'prov_id': 'cut'},\n {'prov_attributes': {'side': 'both',\n 'taper_width': 0.05,\n 'window_type': 'Hann'},\n 'prov_id': 'taper'},\n {'prov_attributes': {'corner_frequency': 0.08,\n 'filter_order': 5,\n 'filter_type': 'Butterworth',\n 'number_of_passes': 2},\n 'prov_id': 'highpass_filter'},\n {'prov_attributes': {'corner_frequency': 20.0,\n 'filter_order': 5,\n 'filter_type': 'Butterworth',\n 'number_of_passes': 2},\n 'prov_id': 'lowpass_filter'}]\n" ] ], [ [ "## Workspace\nWe use the ASDF format as a 'workspace' for saving data and metadata at all stages of processing/analysis.", "_____no_output_____" ] ], [ [ "outfile = os.path.join(os.path.expanduser('~'), 'geonet_test.hdf')\nif os.path.isfile(outfile):\n os.remove(outfile)\nworkspace = StreamWorkspace(outfile)\n# create an ObsPy event object from our dictionary\nevent = get_event_object(origin)\n\n# add the \"raw\" (GEONET actually pre-converts to gals) data\nworkspace.addStreams(event, sc, label='rawgeonet')\neventid = origin['id']\n\nworkspace.addStreams(event, sc_processed, label='processed')", "_____no_output_____" ] ], [ [ "## Creating and Retrieving Stream Metrics\nComputation of metrics requires specifying a list of requested intensity measure types (IMTs) and intensity measure components (IMCs). Not all IMT-IMC combinations are currently supported and in those cases the code returns NaNs. \n\nFor real uses (not just demonstration) it is probably more convenient to specify these values through the config file, which allows for specifying response spectral periods and Fourier amplitude spectra periods as linear or logspaced arrays.", "_____no_output_____" ] ], [ [ "imclist = [\n 'greater_of_two_horizontals',\n 'channels',\n 'rotd50',\n 'rotd100'\n]\nimtlist = [\n 'sa1.0',\n 'PGA',\n 'pgv',\n 'fas2.0',\n 'arias'\n]\nworkspace.setStreamMetrics(\n eventid, \n labels=['processed'], \n imclist=imclist,\n imtlist=imtlist\n)\ndf = workspace.getMetricsTable(\n eventid, \n labels=['processed']\n)", "_____no_output_____" ] ], [ [ "There are a lot of columns here, so we'll show them in sections:", "_____no_output_____" ] ], [ [ "pd.set_option('display.width', 1000)\nprint('ARIAS:')\nprint(df['ARIAS'])\nprint('\\nSpectral Acceleration (1 second)')\nprint(df['SA(1.0)'])\nprint('\\nFourier Amplitude Spectra (2 second)')\nprint(df['FAS(2.0)'])\nprint('\\nPGA')\nprint(df['PGA'])\nprint('\\nPGV')\nprint(df['PGV'])\n\nprint('\\nStation Information:')\nprint(df[['STATION', 'NAME', 'LAT', 'LON', 'SOURCE', 'NETID']])", "ARIAS:\n GEOMETRIC_MEAN GREATER_OF_TWO_HORIZONTALS HN1 HN2 HNZ ROTD100.0 ROTD50.0\n0 NaN 2.7744 2.2772 2.7744 0.6097 2.7921 2.5212\n1 NaN 0.1384 0.1183 0.1384 0.0567 0.1477 0.1282\n2 NaN 12.6649 12.6649 8.9070 12.8956 13.8093 10.8083\n\nSpectral Acceleration (1 second)\n GEOMETRIC_MEAN GREATER_OF_TWO_HORIZONTALS HN1 HN2 HNZ ROTD100.0 ROTD50.0\n0 NaN 42.0960 42.0960 41.8349 12.7510 47.8483 42.0960\n1 NaN 8.5429 8.5429 8.1002 5.1524 9.7224 8.4973\n2 NaN 134.0321 134.0321 81.5174 28.1952 144.8310 102.8883\n\nFourier Amplitude Spectra (2 second)\n GEOMETRIC_MEAN GREATER_OF_TWO_HORIZONTALS HN1 HN2 HNZ ROTD100.0 ROTD50.0\n0 0.5071 NaN NaN NaN NaN NaN NaN\n1 0.1745 NaN NaN NaN NaN NaN NaN\n2 1.1579 NaN NaN NaN NaN NaN NaN\n\nPGA\n GEOMETRIC_MEAN GREATER_OF_TWO_HORIZONTALS HN1 HN2 HNZ ROTD100.0 ROTD50.0\n0 NaN 26.8906 24.5105 26.8906 16.0941 27.3846 26.2241\n1 NaN 4.9814 4.9814 4.0292 2.5057 4.9823 4.5273\n2 NaN 94.6646 94.6646 86.7877 136.7054 100.6484 89.1401\n\nPGV\n GEOMETRIC_MEAN GREATER_OF_TWO_HORIZONTALS HN1 HN2 HNZ ROTD100.0 ROTD50.0\n0 NaN 31.6289 31.6289 31.2814 12.0136 33.4505 31.2814\n1 NaN 9.4751 9.4751 7.9116 7.1996 11.5000 8.5291\n2 NaN 108.1139 108.1139 69.0532 41.4889 120.0471 85.4965\n\nStation Information:\n STATION NAME LAT LON SOURCE NETID\n \n0 HSES Hanmer_Springs_Emergency_Centre -41.476667 172.830556 New Zealand Institute of Geological and Nuclea... NZ\n1 THZ Top_House -40.237500 172.905278 New Zealand Institute of Geological and Nuclea... NZ\n2 WTMC Te_Mara_Farm_Waiau -41.380556 173.053611 New Zealand Institute of Geological and Nuclea... NZ\n" ] ], [ [ "## Retrieving Streams", "_____no_output_____" ] ], [ [ "raw_hses = workspace.getStreams(\n eventid,\n stations=['hses'],\n labels=['rawgeonet'])[0]\nprocessed_hses = workspace.getStreams(\n eventid,\n stations=['hses'],\n labels=['processed'])[0]", "_____no_output_____" ], [ "raw_hses.plot()", "_____no_output_____" ], [ "processed_hses.plot()", "_____no_output_____" ] ] ]

[ "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ] ]

4ad6bf248ab0a54bd257bf56a667620d0a849896

432,546

ipynb

Jupyter Notebook

day07_Boosting/day07_boosting.ipynb

Burntt/MastersMachineLearning

1037c4520b1018365499661481d62dc5d28580e1

[ "MIT" ]

null

day07_Boosting/day07_boosting.ipynb

Burntt/MastersMachineLearning

1037c4520b1018365499661481d62dc5d28580e1

[ "MIT" ]

null

day07_Boosting/day07_boosting.ipynb

Burntt/MastersMachineLearning

1037c4520b1018365499661481d62dc5d28580e1

[ "MIT" ]

null

764.215548

147,964

0.952914

[ [ [ "# Day 6: Bagging and gradient boosting.\n\nThis practice notebook is based on Evgeny Sokolov's awesome [materials](https://github.com/esokolov/ml-course-hse/blob/master/2020-fall/seminars/sem09-gbm-part2.ipynb) and [this notebook](https://github.com/neychev/harbour_ml2020/blob/master/day07_Gradient_boosting/07_trees_boosting_ensembling.ipynb) from Harbour ML course.", "_____no_output_____" ], [ "# Part 1. Bagging and gradient boosting.\n\nLet's analyze how the performance of bagging and gradient boosting depends on the number of base learners in the ensemble.\n\nIn case of bagging, all the learners fit to different samples from the same data distribution $\\mathbb{X} \\times \\mathbb{Y}$. Some of them may be overfitted, nevertheless, subsequent averaging of their individual predictions allows to mitigate this effect. The reason for this is the fact that for uncorrelated algorithms the variance of their composition is $N$ times lower than the individual's. In other words, it's highly unlikely, that all the ensemble components would overfit to some atypical object from the training set (compared to one model). When the ensemble size $N$ becomes large enough, further addition of base learners does not increase the quality.\n\nIn boosting, each algorithm is being fit to the errors of the currently constructed composition, which allows the ensemble to gradually improve the quality of the data distribution approximation. However, the increase of ensemble size $N$ may lead to overfitting, as the addition of new models into the composition further fits the training data, and eventually may decrease the generalization ability.", "_____no_output_____" ] ], [ [ "import matplotlib.pyplot as plt\nimport numpy as np", "_____no_output_____" ] ], [ [ "Firstly, let's generate a synthetic dataset.", "_____no_output_____" ] ], [ [ "X_train = np.linspace(0, 1, 100)\nX_test = np.linspace(0, 1, 1000)\n\n\[email protected]\ndef target(x):\n return x > 0.5\n\n\nY_train = target(X_train) + np.random.randn(*X_train.shape) * 0.1\n\nplt.figure(figsize=(16, 9))\nplt.scatter(X_train, Y_train, s=50)\nplt.grid()\nplt.show()", "_____no_output_____" ] ], [ [ "Firstly, let's take bagging of decision trees algorithm.\n\nHere, the ensemble size is being gradually increased.\nLet's look at how the prediction depends on the size.", "_____no_output_____" ] ], [ [ "from sklearn.ensemble import BaggingRegressor, GradientBoostingRegressor\nfrom sklearn.tree import DecisionTreeRegressor", "_____no_output_____" ], [ "reg = BaggingRegressor(DecisionTreeRegressor(max_depth=2), warm_start=True)\n\nplt.figure(figsize=(20, 30))\nsizes = [1, 2, 5, 20, 100, 500, 1000, 2000]\n\nfor i, s in enumerate(sizes):\n reg.n_estimators = s\n reg.fit(X_train.reshape(-1, 1), Y_train)\n\n plt.subplot(4, 2, i + 1)\n plt.xlim([0, 1])\n plt.scatter(X_train, Y_train, s=30)\n plt.plot(X_test, reg.predict(X_test.reshape(-1, 1)), c=\"green\", linewidth=4)\n plt.title(\"{} trees\".format(s))", "_____no_output_____" ] ], [ [ "You can see that after a certain point the overall prediction does not change with the base learners' addition.\n\nNow let's do the same with the gradient boosting.", "_____no_output_____" ] ], [ [ "reg = GradientBoostingRegressor(max_depth=1, learning_rate=1, warm_start=True)\n\nplt.figure(figsize=(20, 30))\nsizes = [1, 2, 5, 20, 100, 500, 1000, 2000]\n\nfor i, s in enumerate(sizes):\n reg.n_estimators = s\n reg.fit(X_train.reshape(-1, 1), Y_train)\n\n plt.subplot(4, 2, i + 1)\n plt.xlim([0, 1])\n plt.scatter(X_train, Y_train, s=30)\n plt.plot(X_test, reg.predict(X_test.reshape(-1, 1)), c=\"green\", linewidth=4)\n plt.title(\"{} trees\".format(s))", "_____no_output_____" ] ], [ [ "Gradient boosting quickly captured the true dependency, but afterwards began overfitting towards individual objects from the training set. As a result, models with big ensemble sizes were severly overfitted.\n\nOne can tackle this problem by picking a very simple base learner, or intentionally lowering the weight of subsequent algorithms in the composition:\n$$a_N(x) = \\sum_{n=0}^N \\eta \\gamma_N b_n(x).$$\nHere, $\\eta$ is the step parameter, which controls the influence of new ensemble components.\n\nSuch an approach makes training slower compared to bagging, though, makes the final model less overfitted. Still, one should keep in mind that overfitting can happen for any $\\eta$ in the limit of infinite ensemble size.\n", "_____no_output_____" ] ], [ [ "reg = GradientBoostingRegressor(max_depth=1, learning_rate=0.1, warm_start=True)\n\nplt.figure(figsize=(20, 30))\nsizes = [1, 2, 5, 20, 100, 500, 1000, 2000]\n\nfor i, s in enumerate(sizes):\n reg.n_estimators = s\n reg.fit(X_train.reshape(-1, 1), Y_train)\n\n plt.subplot(4, 2, i + 1)\n plt.xlim([0, 1])\n plt.scatter(X_train, Y_train, s=30)\n plt.plot(X_test, reg.predict(X_test.reshape(-1, 1)), c=\"green\", linewidth=4)\n plt.title(\"{} trees\".format(s))", "_____no_output_____" ] ], [ [ "Let's look at the described phenomenon on a more realistic dataset.", "_____no_output_____" ] ], [ [ "from sklearn import datasets\nfrom sklearn.model_selection import train_test_split", "_____no_output_____" ], [ "ds = datasets.load_diabetes()\nX = ds.data\nY = ds.target\nX_train, X_test, Y_train, Y_test = train_test_split(X, Y, train_size=0.5, test_size=0.5)\n\nMAX_ESTIMATORS = 300\n\ngbclf = BaggingRegressor(warm_start=True)\nerr_train_bag = []\nerr_test_bag = []\nfor i in range(1, MAX_ESTIMATORS + 1):\n gbclf.n_estimators = i\n gbclf.fit(X_train, Y_train)\n err_train_bag.append(1 - gbclf.score(X_train, Y_train))\n err_test_bag.append(1 - gbclf.score(X_test, Y_test))\n\ngbclf = GradientBoostingRegressor(warm_start=True, max_depth=2, learning_rate=0.1)\nerr_train_gb = []\nerr_test_gb = []\nfor i in range(1, MAX_ESTIMATORS + 1):\n gbclf.n_estimators = i\n gbclf.fit(X_train, Y_train)\n err_train_gb.append(1 - gbclf.score(X_train, Y_train))\n err_test_gb.append(1 - gbclf.score(X_test, Y_test))\n\nplt.figure(figsize=(10, 4))\n\nplt.subplot(1, 2, 1)\nplt.plot(err_train_gb, label=\"Gradient Boosting\")\nplt.plot(err_train_bag, label=\"Bagging\")\nplt.legend()\nplt.title(\"Train\")\n\nplt.subplot(1, 2, 2)\nplt.plot(err_test_gb, label=\"Gradient Boosting\")\nplt.plot(err_test_bag, label=\"Bagging\")\nplt.legend()\nplt.title(\"Test\")\nplt.gcf().set_size_inches(15, 7)", "_____no_output_____" ] ], [ [ "# Part 2. Multilabel classification with Decision Trees, Random Forests, and Gradient Boosting.", "_____no_output_____" ], [ "In the second part of our practice session, we will apply each method to the classification task and do optimal model selection.", "_____no_output_____" ] ], [ [ "from sklearn.datasets import load_digits", "_____no_output_____" ], [ "data = load_digits()\nX = data.data\ny = data.target", "_____no_output_____" ] ], [ [ "We're going to use the digits dataset. This is a task of recognizing hand-written digits - a multilabel classification into 10 classes.", "_____no_output_____" ] ], [ [ "np.unique(y)", "_____no_output_____" ], [ "fig, axs = plt.subplots(3, 3, figsize=(12, 12))\nfig.suptitle(\"Training data examples\")\n\nfor i in range(9):\n img = X[i].reshape(8, 8)\n axs[i // 3, i % 3].imshow(img)\n axs[i // 3, i % 3].set_title(\"Class label: %s\" % y[i])", "_____no_output_____" ] ], [ [ "Firstly, split the dataset in order to be able to validate your model.\n\n**Hint**: use sklearn's `ShuffleSplit` or `train_test_split`.", "_____no_output_____" ] ], [ [ "# Split the dataset. Use any method you prefer", "_____no_output_____" ] ], [ [ "#### Decision trees", "_____no_output_____" ] ], [ [ "from sklearn.metrics import accuracy_score\nfrom sklearn.tree import DecisionTreeClassifier\n\n\n# Create and fit decision tree with the default parameters\n# Evaluate it on the validation set. Use accuracy", "_____no_output_____" ] ], [ [ "#### Random forest", "_____no_output_____" ] ], [ [ "from sklearn.ensemble import RandomForestClassifier\n\n\n# Create RandomForestClassifier with the default parameters\n# Fit and evaluate", "_____no_output_____" ], [ "# Now let's see how the quality depends on the number of models in the ensemble\n# For each value in [5, 10, 100, 500, 1000] create a random forest with the corresponding size, fit a model and evaluate\n# How does the quality change? What number is sufficient?\n# Please write you conslusions", "_____no_output_____" ] ], [ [ "#### Gradient boosting", "_____no_output_____" ] ], [ [ "from sklearn.ensemble import GradientBoostingClassifier\n\n\n# Create GradientBoostingClassifier with the default parameters\n# Fit and evaluate. Compare its quality to random forest", "_____no_output_____" ], [ "# Now let's see how the quality depends on the number of models in the ensemble\n# For each value in [5, 10, 100, 500, 1000] train a gradient boosting with the corresponding size\n# How does the quality change? What number is sufficient?\n# Please write you conslusions", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ] ]

4ad6c44838f7122db56358bafe4deb1a13ba8884

376,770

ipynb

Jupyter Notebook

Chapter01/Breast Cancer Detection with SVM (Jupyter Notebook).ipynb

hestrang1993/Machine-Learning-for-Healthcare-Analytics-Projects

70a2c81a427b664bf1f4eb594e30f2b43d280790

[ "MIT" ]

null

Chapter01/Breast Cancer Detection with SVM (Jupyter Notebook).ipynb

hestrang1993/Machine-Learning-for-Healthcare-Analytics-Projects

70a2c81a427b664bf1f4eb594e30f2b43d280790

[ "MIT" ]

null

Chapter01/Breast Cancer Detection with SVM (Jupyter Notebook).ipynb

hestrang1993/Machine-Learning-for-Healthcare-Analytics-Projects

70a2c81a427b664bf1f4eb594e30f2b43d280790

[ "MIT" ]

null

928.004926

331,220

0.936062

[ [ [ "# Check Python Version\nimport sys\nimport scipy\nimport numpy\nimport matplotlib\nimport pandas\nimport sklearn\n\nprint('Python: {}'.format(sys.version))\nprint('scipy: {}'.format(scipy.__version__))\nprint('numpy: {}'.format(numpy.__version__))\nprint('matplotlib: {}'.format(matplotlib.__version__))\nprint('pandas: {}'.format(pandas.__version__))\nprint('sklearn: {}'.format(sklearn.__version__))", "Python: 2.7.13 |Continuum Analytics, Inc.| (default, May 11 2017, 13:17:26) [MSC v.1500 64 bit (AMD64)]\nscipy: 1.0.0\nnumpy: 1.14.0\nmatplotlib: 2.1.0\npandas: 0.21.0\nsklearn: 0.19.1\n" ], [ "import numpy as np\nfrom sklearn import preprocessing, cross_validation\nfrom sklearn.neighbors import KNeighborsClassifier\nfrom sklearn.svm import SVC\nfrom sklearn import model_selection\nfrom sklearn.metrics import classification_report\nfrom sklearn.metrics import accuracy_score\nfrom pandas.plotting import scatter_matrix\nimport matplotlib.pyplot as plt\nimport pandas as pd", "C:\\Programdata\\anaconda2\\lib\\site-packages\\sklearn\\cross_validation.py:41: DeprecationWarning: This module was deprecated in version 0.18 in favor of the model_selection module into which all the refactored classes and functions are moved. Also note that the interface of the new CV iterators are different from that of this module. This module will be removed in 0.20.\n \"This module will be removed in 0.20.\", DeprecationWarning)\n" ], [ "# Load Dataset\nurl = \"https://archive.ics.uci.edu/ml/machine-learning-databases/breast-cancer-wisconsin/breast-cancer-wisconsin.data\"\nnames = ['id', 'clump_thickness', 'uniform_cell_size', 'uniform_cell_shape',\n 'marginal_adhesion', 'single_epithelial_size', 'bare_nuclei',\n 'bland_chromatin', 'normal_nucleoli', 'mitoses', 'class']\ndf = pd.read_csv(url, names=names)\n", "_____no_output_____" ], [ "# Preprocess the data\ndf.replace('?',-99999, inplace=True)\nprint(df.axes)\n\ndf.drop(['id'], 1, inplace=True)", "[RangeIndex(start=0, stop=699, step=1), Index([u'id', u'clump_thickness', u'uniform_cell_size', u'uniform_cell_shape',\n u'marginal_adhesion', u'single_epithelial_size', u'bare_nuclei',\n u'bland_chromatin', u'normal_nucleoli', u'mitoses', u'class'],\n dtype='object')]\n" ], [ "# Let explore the dataset and do a few visualizations\nprint(df.loc[10])\n\n# Print the shape of the dataset\nprint(df.shape)", "clump_thickness 1\nuniform_cell_size 1\nuniform_cell_shape 1\nmarginal_adhesion 1\nsingle_epithelial_size 1\nbare_nuclei 1\nbland_chromatin 3\nnormal_nucleoli 1\nmitoses 1\nclass 2\nName: 10, dtype: object\n(699, 10)\n" ], [ "# Describe the dataset\nprint(df.describe())", " clump_thickness uniform_cell_size uniform_cell_shape \\\ncount 699.000000 699.000000 699.000000 \nmean 4.417740 3.134478 3.207439 \nstd 2.815741 3.051459 2.971913 \nmin 1.000000 1.000000 1.000000 \n25% 2.000000 1.000000 1.000000 \n50% 4.000000 1.000000 1.000000 \n75% 6.000000 5.000000 5.000000 \nmax 10.000000 10.000000 10.000000 \n\n marginal_adhesion single_epithelial_size bland_chromatin \\\ncount 699.000000 699.000000 699.000000 \nmean 2.806867 3.216023 3.437768 \nstd 2.855379 2.214300 2.438364 \nmin 1.000000 1.000000 1.000000 \n25% 1.000000 2.000000 2.000000 \n50% 1.000000 2.000000 3.000000 \n75% 4.000000 4.000000 5.000000 \nmax 10.000000 10.000000 10.000000 \n\n normal_nucleoli mitoses class \ncount 699.000000 699.000000 699.000000 \nmean 2.866953 1.589413 2.689557 \nstd 3.053634 1.715078 0.951273 \nmin 1.000000 1.000000 2.000000 \n25% 1.000000 1.000000 2.000000 \n50% 1.000000 1.000000 2.000000 \n75% 4.000000 1.000000 4.000000 \nmax 10.000000 10.000000 4.000000 \n" ], [ "# Plot histograms for each variable\ndf.hist(figsize = (10, 10))\nplt.show()", "_____no_output_____" ], [ "# Create scatter plot matrix\nscatter_matrix(df, figsize = (18,18))\nplt.show()", "_____no_output_____" ], [ "# Create X and Y datasets for training\nX = np.array(df.drop(['class'], 1))\ny = np.array(df['class'])\n\nX_train, X_test, y_train, y_test = cross_validation.train_test_split(X, y, test_size=0.2)", "_____no_output_____" ], [ "# Testing Options\nseed = 8\nscoring = 'accuracy'", "_____no_output_____" ], [ "# Define models to train\nmodels = []\nmodels.append(('KNN', KNeighborsClassifier(n_neighbors = 5)))\nmodels.append(('SVM', SVC()))\n\n# evaluate each model in turn\nresults = []\nnames = []\n\nfor name, model in models:\n kfold = model_selection.KFold(n_splits=10, random_state = seed)\n cv_results = model_selection.cross_val_score(model, X_train, y_train, cv=kfold, scoring=scoring)\n results.append(cv_results)\n names.append(name)\n msg = \"%s: %f (%f)\" % (name, cv_results.mean(), cv_results.std())\n print(msg)", "KNN: 0.973117 (0.018442)\nSVM: 0.951558 (0.028352)\n" ], [ "# Make predictions on validation dataset\n\nfor name, model in models:\n model.fit(X_train, y_train)\n predictions = model.predict(X_test)\n print(name)\n print(accuracy_score(y_test, predictions))\n print(classification_report(y_test, predictions))\n \n# Accuracy - ratio of correctly predicted observation to the total observations. \n# Precision - (false positives) ratio of correctly predicted positive observations to the total predicted positive observations\n# Recall (Sensitivity) - (false negatives) ratio of correctly predicted positive observations to the all observations in actual class - yes.\n# F1 score - F1 Score is the weighted average of Precision and Recall. Therefore, this score takes both false positives and false ", "KNN\n0.9428571428571428\n precision recall f1-score support\n\n 2 0.94 0.96 0.95 83\n 4 0.95 0.91 0.93 57\n\navg / total 0.94 0.94 0.94 140\n\nSVM\n0.9571428571428572\n precision recall f1-score support\n\n 2 1.00 0.93 0.96 83\n 4 0.90 1.00 0.95 57\n\navg / total 0.96 0.96 0.96 140\n\n" ], [ "clf = SVC()\n\nclf.fit(X_train, y_train)\naccuracy = clf.score(X_test, y_test)\nprint(accuracy)\n\nexample_measures = np.array([[4,2,1,1,1,2,3,2,1]])\nexample_measures = example_measures.reshape(len(example_measures), -1)\nprediction = clf.predict(example_measures)\nprint(prediction)", "0.9571428571428572\n[2]\n" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad6c9ec4c43c9397120b6cb11776abb9c77221e

88,389

ipynb

Jupyter Notebook

Copy_of_Q&A_on_PDF_Files.ipynb

rizwandel/Auto_TS

89aabfe70cb8f8475c30f38421b5edf3ba3610dc

[ "Apache-2.0" ]

null

Copy_of_Q&A_on_PDF_Files.ipynb

rizwandel/Auto_TS

89aabfe70cb8f8475c30f38421b5edf3ba3610dc

[ "Apache-2.0" ]

null

Copy_of_Q&A_on_PDF_Files.ipynb

rizwandel/Auto_TS

89aabfe70cb8f8475c30f38421b5edf3ba3610dc

[ "Apache-2.0" ]

null

72.331424

8,380

0.625938

[ [ [ "<a href=\"https://colab.research.google.com/github/rizwandel/Auto_TS/blob/master/Copy_of_Q%26A_on_PDF_Files.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>", "_____no_output_____" ] ], [ [ "!pip install cdqa\n!pip install flask-ngrok\n", "Collecting cdqa\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/39/f5/af831b7ee653aa6bace99e39ec6b2754b1adb10bb60a1296f5e16f1f24ee/cdqa-1.3.9.tar.gz (45kB)\n\u001b[K |████████████████████████████████| 51kB 2.8MB/s \n\u001b[?25hCollecting Flask==1.1.1\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/9b/93/628509b8d5dc749656a9641f4caf13540e2cdec85276964ff8f43bbb1d3b/Flask-1.1.1-py2.py3-none-any.whl (94kB)\n\u001b[K |████████████████████████████████| 102kB 5.0MB/s \n\u001b[?25hCollecting flask_cors==3.0.8\n Downloading https://files.pythonhosted.org/packages/78/38/e68b11daa5d613e3a91e4bf3da76c94ac9ee0d9cd515af9c1ab80d36f709/Flask_Cors-3.0.8-py2.py3-none-any.whl\nCollecting joblib==0.13.2\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/cd/c1/50a758e8247561e58cb87305b1e90b171b8c767b15b12a1734001f41d356/joblib-0.13.2-py2.py3-none-any.whl (278kB)\n\u001b[K |████████████████████████████████| 286kB 21.1MB/s \n\u001b[?25hCollecting pandas==0.25.0\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/1d/9a/7eb9952f4b4d73fbd75ad1d5d6112f407e695957444cb695cbb3cdab918a/pandas-0.25.0-cp36-cp36m-manylinux1_x86_64.whl (10.5MB)\n\u001b[K |████████████████████████████████| 10.5MB 40.2MB/s \n\u001b[?25hCollecting prettytable==0.7.2\n Downloading https://files.pythonhosted.org/packages/ef/30/4b0746848746ed5941f052479e7c23d2b56d174b82f4fd34a25e389831f5/prettytable-0.7.2.tar.bz2\nCollecting transformers==2.1.1\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/fd/f9/51824e40f0a23a49eab4fcaa45c1c797cbf9761adedd0b558dab7c958b34/transformers-2.1.1-py3-none-any.whl (311kB)\n\u001b[K |████████████████████████████████| 317kB 32.4MB/s \n\u001b[?25hCollecting scikit_learn==0.21.2\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/85/04/49633f490f726da6e454fddc8e938bbb5bfed2001681118d3814c219b723/scikit_learn-0.21.2-cp36-cp36m-manylinux1_x86_64.whl (6.7MB)\n\u001b[K |████████████████████████████████| 6.7MB 35.5MB/s \n\u001b[?25hCollecting tika==1.19\n Downloading https://files.pythonhosted.org/packages/10/75/b566e446ffcf292f10c8d84c15a3d91615fe3d7ca8072a17c949d4e84b66/tika-1.19.tar.gz\nCollecting torch==1.2.0\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/30/57/d5cceb0799c06733eefce80c395459f28970ebb9e896846ce96ab579a3f1/torch-1.2.0-cp36-cp36m-manylinux1_x86_64.whl (748.8MB)\n\u001b[K |████████████████████████████████| 748.9MB 23kB/s \n\u001b[?25hCollecting markdown==3.1.1\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/c0/4e/fd492e91abdc2d2fcb70ef453064d980688762079397f779758e055f6575/Markdown-3.1.1-py2.py3-none-any.whl (87kB)\n\u001b[K |████████████████████████████████| 92kB 9.7MB/s \n\u001b[?25hCollecting tqdm==4.32.2\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/9f/3d/7a6b68b631d2ab54975f3a4863f3c4e9b26445353264ef01f465dc9b0208/tqdm-4.32.2-py2.py3-none-any.whl (50kB)\n\u001b[K |████████████████████████████████| 51kB 6.4MB/s \n\u001b[?25hCollecting wget==3.2\n Downloading https://files.pythonhosted.org/packages/47/6a/62e288da7bcda82b935ff0c6cfe542970f04e29c756b0e147251b2fb251f/wget-3.2.zip\nRequirement already satisfied: click>=5.1 in /usr/local/lib/python3.6/dist-packages (from Flask==1.1.1->cdqa) (7.1.2)\nRequirement already satisfied: Jinja2>=2.10.1 in /usr/local/lib/python3.6/dist-packages (from Flask==1.1.1->cdqa) (2.11.2)\nRequirement already satisfied: itsdangerous>=0.24 in /usr/local/lib/python3.6/dist-packages (from Flask==1.1.1->cdqa) (1.1.0)\nRequirement already satisfied: Werkzeug>=0.15 in /usr/local/lib/python3.6/dist-packages (from Flask==1.1.1->cdqa) (1.0.1)\nRequirement already satisfied: Six in /usr/local/lib/python3.6/dist-packages (from flask_cors==3.0.8->cdqa) (1.15.0)\nRequirement already satisfied: numpy>=1.13.3 in /usr/local/lib/python3.6/dist-packages (from pandas==0.25.0->cdqa) (1.18.5)\nRequirement already satisfied: python-dateutil>=2.6.1 in /usr/local/lib/python3.6/dist-packages (from pandas==0.25.0->cdqa) (2.8.1)\nRequirement already satisfied: pytz>=2017.2 in /usr/local/lib/python3.6/dist-packages (from pandas==0.25.0->cdqa) (2018.9)\nRequirement already satisfied: regex in /usr/local/lib/python3.6/dist-packages (from transformers==2.1.1->cdqa) (2019.12.20)\nRequirement already satisfied: requests in /usr/local/lib/python3.6/dist-packages (from transformers==2.1.1->cdqa) (2.23.0)\nCollecting sentencepiece\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/e5/2d/6d4ca4bef9a67070fa1cac508606328329152b1df10bdf31fb6e4e727894/sentencepiece-0.1.94-cp36-cp36m-manylinux2014_x86_64.whl (1.1MB)\n\u001b[K |████████████████████████████████| 1.1MB 34.0MB/s \n\u001b[?25hCollecting sacremoses\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/7d/34/09d19aff26edcc8eb2a01bed8e98f13a1537005d31e95233fd48216eed10/sacremoses-0.0.43.tar.gz (883kB)\n\u001b[K |████████████████████████████████| 890kB 34.8MB/s \n\u001b[?25hCollecting boto3\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/b7/df/6c160e21a5caa800de16f2aa859b92671623118b4d124639aeab06876c06/boto3-1.16.28-py2.py3-none-any.whl (129kB)\n\u001b[K |████████████████████████████████| 133kB 45.4MB/s \n\u001b[?25hRequirement already satisfied: scipy>=0.17.0 in /usr/local/lib/python3.6/dist-packages (from scikit_learn==0.21.2->cdqa) (1.4.1)\nRequirement already satisfied: setuptools in /usr/local/lib/python3.6/dist-packages (from tika==1.19->cdqa) (50.3.2)\nRequirement already satisfied: MarkupSafe>=0.23 in /usr/local/lib/python3.6/dist-packages (from Jinja2>=2.10.1->Flask==1.1.1->cdqa) (1.1.1)\nRequirement already satisfied: idna<3,>=2.5 in /usr/local/lib/python3.6/dist-packages (from requests->transformers==2.1.1->cdqa) (2.10)\nRequirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.6/dist-packages (from requests->transformers==2.1.1->cdqa) (2020.11.8)\nRequirement already satisfied: chardet<4,>=3.0.2 in /usr/local/lib/python3.6/dist-packages (from requests->transformers==2.1.1->cdqa) (3.0.4)\nRequirement already satisfied: urllib3!=1.25.0,!=1.25.1,<1.26,>=1.21.1 in /usr/local/lib/python3.6/dist-packages (from requests->transformers==2.1.1->cdqa) (1.24.3)\nCollecting botocore<1.20.0,>=1.19.28\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/ac/e0/966b82eb9eab5fe36e80bcbbfda0d3f49cdd9ec896ebe9edb9824f896cd7/botocore-1.19.28-py2.py3-none-any.whl (7.0MB)\n\u001b[K |████████████████████████████████| 7.0MB 35.9MB/s \n\u001b[?25hCollecting s3transfer<0.4.0,>=0.3.0\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/69/79/e6afb3d8b0b4e96cefbdc690f741d7dd24547ff1f94240c997a26fa908d3/s3transfer-0.3.3-py2.py3-none-any.whl (69kB)\n\u001b[K |████████████████████████████████| 71kB 7.5MB/s \n\u001b[?25hCollecting jmespath<1.0.0,>=0.7.1\n Downloading https://files.pythonhosted.org/packages/07/cb/5f001272b6faeb23c1c9e0acc04d48eaaf5c862c17709d20e3469c6e0139/jmespath-0.10.0-py2.py3-none-any.whl\nBuilding wheels for collected packages: cdqa, prettytable, tika, wget, sacremoses\n Building wheel for cdqa (setup.py) ... \u001b[?25l\u001b[?25hdone\n Created wheel for cdqa: filename=cdqa-1.3.9-cp36-none-any.whl size=47640 sha256=cdb55193a2747e8a61a6ce0b2a71d413a0d8ae2d0a6e4a0011c253f20a182706\n Stored in directory: /root/.cache/pip/wheels/8b/9a/68/d3f7651ea29c30d1bebc9e946bf5a8cf922e1c86fb6b8a33d9\n Building wheel for prettytable (setup.py) ... \u001b[?25l\u001b[?25hdone\n Created wheel for prettytable: filename=prettytable-0.7.2-cp36-none-any.whl size=13700 sha256=80123678fa07884d931b1bc097c1e44614ee3d350a32595456d60e877faacda3\n Stored in directory: /root/.cache/pip/wheels/80/34/1c/3967380d9676d162cb59513bd9dc862d0584e045a162095606\n Building wheel for tika (setup.py) ... \u001b[?25l\u001b[?25hdone\n Created wheel for tika: filename=tika-1.19-cp36-none-any.whl size=29223 sha256=099232cfcdcc83c585bf16bbd9ed0d23a953cdfc22aa90f58d5fb4fec210a3a0\n Stored in directory: /root/.cache/pip/wheels/b4/db/8a/3a3f0c0725448eaa92703e3dda71e29dc13a119ff6c1036848\n Building wheel for wget (setup.py) ... \u001b[?25l\u001b[?25hdone\n Created wheel for wget: filename=wget-3.2-cp36-none-any.whl size=9682 sha256=d65148a47e9250998b43ab5548a60b68b933e0078c7a231d0e2253d1770cb317\n Stored in directory: /root/.cache/pip/wheels/40/15/30/7d8f7cea2902b4db79e3fea550d7d7b85ecb27ef992b618f3f\n Building wheel for sacremoses (setup.py) ... \u001b[?25l\u001b[?25hdone\n Created wheel for sacremoses: filename=sacremoses-0.0.43-cp36-none-any.whl size=893257 sha256=ac4eb32a53b6cf0d23bff1b4e918265ba05cf1c54f01527b7350c67067f31ecb\n Stored in directory: /root/.cache/pip/wheels/29/3c/fd/7ce5c3f0666dab31a50123635e6fb5e19ceb42ce38d4e58f45\nSuccessfully built cdqa prettytable tika wget sacremoses\n\u001b[31mERROR: torchvision 0.8.1+cu101 has requirement torch==1.7.0, but you'll have torch 1.2.0 which is incompatible.\u001b[0m\n\u001b[31mERROR: spacy 2.2.4 has requirement tqdm<5.0.0,>=4.38.0, but you'll have tqdm 4.32.2 which is incompatible.\u001b[0m\n\u001b[31mERROR: google-colab 1.0.0 has requirement pandas~=1.1.0; python_version >= \"3.0\", but you'll have pandas 0.25.0 which is incompatible.\u001b[0m\n\u001b[31mERROR: fbprophet 0.7.1 has requirement pandas>=1.0.4, but you'll have pandas 0.25.0 which is incompatible.\u001b[0m\n\u001b[31mERROR: fbprophet 0.7.1 has requirement tqdm>=4.36.1, but you'll have tqdm 4.32.2 which is incompatible.\u001b[0m\n\u001b[31mERROR: botocore 1.19.28 has requirement urllib3<1.27,>=1.25.4; python_version != \"3.4\", but you'll have urllib3 1.24.3 which is incompatible.\u001b[0m\nInstalling collected packages: Flask, flask-cors, joblib, pandas, prettytable, sentencepiece, tqdm, sacremoses, jmespath, botocore, s3transfer, boto3, transformers, scikit-learn, tika, torch, markdown, wget, cdqa\n Found existing installation: Flask 1.1.2\n Uninstalling Flask-1.1.2:\n Successfully uninstalled Flask-1.1.2\n Found existing installation: joblib 0.17.0\n Uninstalling joblib-0.17.0:\n Successfully uninstalled joblib-0.17.0\n Found existing installation: pandas 1.1.4\n Uninstalling pandas-1.1.4:\n Successfully uninstalled pandas-1.1.4\n Found existing installation: prettytable 2.0.0\n Uninstalling prettytable-2.0.0:\n Successfully uninstalled prettytable-2.0.0\n Found existing installation: tqdm 4.41.1\n Uninstalling tqdm-4.41.1:\n Successfully uninstalled tqdm-4.41.1\n Found existing installation: scikit-learn 0.22.2.post1\n Uninstalling scikit-learn-0.22.2.post1:\n Successfully uninstalled scikit-learn-0.22.2.post1\n Found existing installation: torch 1.7.0+cu101\n Uninstalling torch-1.7.0+cu101:\n Successfully uninstalled torch-1.7.0+cu101\n Found existing installation: Markdown 3.3.3\n Uninstalling Markdown-3.3.3:\n Successfully uninstalled Markdown-3.3.3\nSuccessfully installed Flask-1.1.1 boto3-1.16.28 botocore-1.19.28 cdqa-1.3.9 flask-cors-3.0.8 jmespath-0.10.0 joblib-0.13.2 markdown-3.1.1 pandas-0.25.0 prettytable-0.7.2 s3transfer-0.3.3 sacremoses-0.0.43 scikit-learn-0.21.2 sentencepiece-0.1.94 tika-1.19 torch-1.2.0 tqdm-4.32.2 transformers-2.1.1 wget-3.2\n" ], [ "import os\nimport pandas as pd\nfrom ast import literal_eval\n\nfrom cdqa.utils.converters import pdf_converter\nfrom cdqa.pipeline import QAPipeline\nfrom cdqa.utils.download import download_model", "/usr/local/lib/python3.6/dist-packages/tqdm/autonotebook/__init__.py:18: TqdmExperimentalWarning: Using `tqdm.autonotebook.tqdm` in notebook mode. Use `tqdm.tqdm` instead to force console mode (e.g. in jupyter console)\n \" (e.g. in jupyter console)\", TqdmExperimentalWarning)\n" ], [ "download_model(model='bert-squad_1.1', dir='./models')", "\nDownloading trained model...\n" ], [ "!ls models", "bert_qa.joblib\n" ], [ "!mkdir docs", "_____no_output_____" ], [ "!wget -P ./docs/ https://s2.q4cdn.com/299287126/files/doc_financials/2020/q3/AMZN-Q3-2020-Earnings-Release.pdf\n!wget -P ./docs/ https://s2.q4cdn.com/299287126/files/doc_financials/2020/q2/Q2-2020-Amazon-Earnings-Release.pdf\n!wget -P ./docs/ https://s2.q4cdn.com/299287126/files/doc_financials/2020/Q1/AMZN-Q1-2020-Earnings-Release.pdf\n!wget -P ./docs/ https://s2.q4cdn.com/299287126/files/doc_news/archive/Amazon-Q4-2019-Earnings-Release.pdf\n!wget -P ./docs/ https://s2.q4cdn.com/299287126/files/doc_news/archive/Q3-2019-Amazon-Financial-Results.pdf", "--2020-12-02 18:17:27-- https://s2.q4cdn.com/299287126/files/doc_financials/2020/q3/AMZN-Q3-2020-Earnings-Release.pdf\nResolving s2.q4cdn.com (s2.q4cdn.com)... 161.202.39.238\nConnecting to s2.q4cdn.com (s2.q4cdn.com)|161.202.39.238|:443... connected.\nHTTP request sent, awaiting response... 200 OK\nLength: 945573 (923K) [application/pdf]\nSaving to: ‘./docs/AMZN-Q3-2020-Earnings-Release.pdf’\n\nAMZN-Q3-2020-Earnin 100%[===================>] 923.41K --.-KB/s in 0.08s \n\n2020-12-02 18:17:27 (10.7 MB/s) - ‘./docs/AMZN-Q3-2020-Earnings-Release.pdf’ saved [945573/945573]\n\n--2020-12-02 18:17:27-- https://s2.q4cdn.com/299287126/files/doc_financials/2020/q2/Q2-2020-Amazon-Earnings-Release.pdf\nResolving s2.q4cdn.com (s2.q4cdn.com)... 161.202.39.238\nConnecting to s2.q4cdn.com (s2.q4cdn.com)|161.202.39.238|:443... connected.\nHTTP request sent, awaiting response... 200 OK\nLength: 366951 (358K) [application/pdf]\nSaving to: ‘./docs/Q2-2020-Amazon-Earnings-Release.pdf’\n\nQ2-2020-Amazon-Earn 100%[===================>] 358.35K --.-KB/s in 0.05s \n\n2020-12-02 18:17:27 (6.50 MB/s) - ‘./docs/Q2-2020-Amazon-Earnings-Release.pdf’ saved [366951/366951]\n\n--2020-12-02 18:17:27-- https://s2.q4cdn.com/299287126/files/doc_financials/2020/Q1/AMZN-Q1-2020-Earnings-Release.pdf\nResolving s2.q4cdn.com (s2.q4cdn.com)... 161.202.39.238\nConnecting to s2.q4cdn.com (s2.q4cdn.com)|161.202.39.238|:443... connected.\nHTTP request sent, awaiting response... 200 OK\nLength: 536732 (524K) [application/pdf]\nSaving to: ‘./docs/AMZN-Q1-2020-Earnings-Release.pdf’\n\nAMZN-Q1-2020-Earnin 100%[===================>] 524.15K --.-KB/s in 0.07s \n\n2020-12-02 18:17:27 (6.94 MB/s) - ‘./docs/AMZN-Q1-2020-Earnings-Release.pdf’ saved [536732/536732]\n\n--2020-12-02 18:17:27-- https://s2.q4cdn.com/299287126/files/doc_news/archive/Amazon-Q4-2019-Earnings-Release.pdf\nResolving s2.q4cdn.com (s2.q4cdn.com)... 161.202.39.238\nConnecting to s2.q4cdn.com (s2.q4cdn.com)|161.202.39.238|:443... connected.\nHTTP request sent, awaiting response... 200 OK\nLength: 637064 (622K) [application/pdf]\nSaving to: ‘./docs/Amazon-Q4-2019-Earnings-Release.pdf’\n\nAmazon-Q4-2019-Earn 100%[===================>] 622.13K 536KB/s in 1.2s \n\n2020-12-02 18:17:30 (536 KB/s) - ‘./docs/Amazon-Q4-2019-Earnings-Release.pdf’ saved [637064/637064]\n\n--2020-12-02 18:17:30-- https://s2.q4cdn.com/299287126/files/doc_news/archive/Q3-2019-Amazon-Financial-Results.pdf\nResolving s2.q4cdn.com (s2.q4cdn.com)... 161.202.39.238\nConnecting to s2.q4cdn.com (s2.q4cdn.com)|161.202.39.238|:443... connected.\nHTTP request sent, awaiting response... 200 OK\nLength: 545031 (532K) [application/pdf]\nSaving to: ‘./docs/Q3-2019-Amazon-Financial-Results.pdf’\n\nQ3-2019-Amazon-Fina 100%[===================>] 532.26K 593KB/s in 0.9s \n\n2020-12-02 18:17:31 (593 KB/s) - ‘./docs/Q3-2019-Amazon-Financial-Results.pdf’ saved [545031/545031]\n\n" ], [ "df = pdf_converter(directory_path='./docs/')\ndf.head()", "2020-12-02 18:17:32,028 [MainThread ] [INFO ] Retrieving http://search.maven.org/remotecontent?filepath=org/apache/tika/tika-server/1.19/tika-server-1.19.jar to /tmp/tika-server.jar.\n2020-12-02 18:17:39,242 [MainThread ] [INFO ] Retrieving http://search.maven.org/remotecontent?filepath=org/apache/tika/tika-server/1.19/tika-server-1.19.jar.md5 to /tmp/tika-server.jar.md5.\n2020-12-02 18:17:40,998 [MainThread ] [WARNI] Failed to see startup log message; retrying...\n" ], [ "# pd.set_option('display.max_colwidth', -1)\ndf.head()", "_____no_output_____" ], [ "cdqa_pipeline = QAPipeline(reader='./models/bert_qa.joblib', max_df=1.0)\n\ncdqa_pipeline.fit_retriever(df=df)", "100%|██████████| 231508/231508 [00:00<00:00, 323263.68B/s]\n" ], [ "import joblib\n# cdqa_pipeline.to(\"cpu\")\njoblib.dump(cdqa_pipeline, './models/bert_qa_customc.joblib')", "_____no_output_____" ], [ "cdqa_pipeline=joblib.load('./models/bert_qa_customc.joblib')\ncdqa_pipeline", "_____no_output_____" ], [ "query = 'How much is increase in operating cash flow?'\nprediction = cdqa_pipeline.predict(query, 3)", "_____no_output_____" ], [ "cdqa_pipeline", "_____no_output_____" ], [ "prediction", "_____no_output_____" ], [ "query = 'What is latest earnings per share?'\ncdqa_pipeline.predict(query)", "_____no_output_____" ], [ "query = 'How many jobs are created in 2020?'\nprediction = cdqa_pipeline.predict(query)", "_____no_output_____" ], [ "print('query: {}'.format(query))\nprint('answer: {}'.format(prediction[0]))\nprint('title: {}'.format(prediction[1]))\nprint('paragraph: {}'.format(prediction[2]))", "query: How many jobs are created in 2020?\nanswer: 1.1 million\ntitle: Q2-2020-Amazon-Earnings-Release\nparagraph: Empowering Small and Medium-Sized Businesses• Amazon released its 2020 Small and Medium-Sized Business (SMB) Impact Report, highlighting how SMBs selling in its U.S. store have sold more than 3.4 billion products in the past year and created an estimated 1.1 million jobs.• In partnership with the British small business support network “Enterprise Nation,” Amazon launched the Amazon Small Business Accelerator, which aims to support more than 200,000 small businesses across the U.K. negatively impacted by the COVID-19 crisis. Amazon hosted a week-long boot camp in the U.K. to help 1,000 offline businesses get online, and offered free services, AWS credits, training, and support.• In the U.K., Amazon worked with the British Chambers of Commerce to give up to 1,000 businesses tours of Amazon fulfillment centers, helping other companies learn from the safety measures Amazon has put in place within its own operations so the businesses can re-open safely and kick-start the economy.• Amazon in Japan launched Global Selling to allow Japanese sellers to reach new customers across 16 Amazon sites worldwide. • Amazon in India announced plans to help digitally enable micro, small, and medium businesses across the country as part of a $1 billion investment pledge. Amazon launched Local Shops on Amazon.in, offering shopkeepers and retailers with physical stores the ability to register to serve more customers from their local areas. Since launch, more than 11,000 sellers have enrolled in the program. In addition, Amazon introduced seller registration and account management services in Hindi to help businesses overcome language barriers. Since launch, more than 10,000 sellers have used Hindi to register on Amazon.in. \n" ], [ "query = 'How many full time employees are on Amazon roll?'\nprediction = cdqa_pipeline.predict(query)", "_____no_output_____" ], [ "print('query: {}'.format(query))\nprint('answer: {}'.format(prediction[0]))\nprint('title: {}'.format(prediction[1]))\nprint('paragraph: {}'.format(prediction[2]))", "query: How many full time employees are on Amazon roll?\nanswer: 650,000\ntitle: Q2-2020-Amazon-Earnings-Release\nparagraph: Supporting Employees• Amazon’s top priority is providing for the health and safety of our employees and partners, and the company spent more than $4 billion in the second quarter on incremental COVID-19 related initiatives to help keep employees safe, provide additional compensation to our employees and delivery partners, and deliver products to customers.• Amazon provided a one-time Thank You bonus totaling over $500 million to all front-line employees and partners who were with the company throughout the month of June.• Amazon introduced a new family backup care benefit through Care.com to 650,000 full and part-time Amazon and Whole Foods Market employees in the U.S. This benefit provides employees with up to 10 days of company-subsidized emergency backup child or adult care.• Amazon introduced Distance Assistant to help keep employees safe by providing them with live feedback on their social distancing via a 50-inch monitor. Amazon made the software and AI behind this innovation available via open source so that anyone can create their own Distance Assistant at no cost and get up and running with just a monitor, computer, and camera.• Amazon is collaborating with national medical care group Crossover Health to pilot Amazon Neighborhood Health Centers, which are new medical facilities available to Amazon employees and their families. The centers will provide access to quality, convenient care while reducing health care costs for employees and Amazon. The pilot includes 20 Crossover Health branded and operated centers in Dallas/Fort Worth, Texas; Phoenix, Arizona; Louisville, Kentucky; Detroit, Michigan; and San Bernardino/Moreno Valley, California.\n" ], [ "query = 'General Availability of which AWS services were announced?'\nprediction = cdqa_pipeline.predict(query, n_predictions=5)", "_____no_output_____" ], [ "prediction", "_____no_output_____" ], [ "query = 'What is the impact of COVID on business?'\nprediction = cdqa_pipeline.predict(query, n_predictions=2)", "_____no_output_____" ], [ "print('query: {}'.format(query))\nprint('answer: {}'.format(prediction[0]))\nprint('title: {}'.format(prediction[1]))\n# print('paragraph: {}'.format(prediction[2]))", "query: What is the impact of COVID on business?\nanswer: ('negatively impacted by the COVID-19 crisis', 'Q2-2020-Amazon-Earnings-Release', 'Empowering Small and Medium-Sized Businesses• Amazon released its 2020 Small and Medium-Sized Business (SMB) Impact Report, highlighting how SMBs selling in its U.S. store have sold more than 3.4 billion products in the past year and created an estimated 1.1 million jobs.• In partnership with the British small business support network “Enterprise Nation,” Amazon launched the Amazon Small Business Accelerator, which aims to support more than 200,000 small businesses across the U.K. negatively impacted by the COVID-19 crisis. Amazon hosted a week-long boot camp in the U.K. to help 1,000 offline businesses get online, and offered free services, AWS credits, training, and support.• In the U.K., Amazon worked with the British Chambers of Commerce to give up to 1,000 businesses tours of Amazon fulfillment centers, helping other companies learn from the safety measures Amazon has put in place within its own operations so the businesses can re-open safely and kick-start the economy.• Amazon in Japan launched Global Selling to allow Japanese sellers to reach new customers across 16 Amazon sites worldwide. • Amazon in India announced plans to help digitally enable micro, small, and medium businesses across the country as part of a $1 billion investment pledge. Amazon launched Local Shops on Amazon.in, offering shopkeepers and retailers with physical stores the ability to register to serve more customers from their local areas. Since launch, more than 11,000 sellers have enrolled in the program. In addition, Amazon introduced seller registration and account management services in Hindi to help businesses overcome language barriers. Since launch, more than 10,000 sellers have used Hindi to register on Amazon.in. ', 5.994336808684102)\ntitle: ('lessen the impact that COVID-19 has on families, communities, and businesses', 'AMZN-Q1-2020-Earnings-Release', 'Amazon Small Business Academy to help entrepreneurs learn how to build their businesses online. • AWS is helping healthcare workers, medical researchers, scientists, and public health officials working to understand and fight COVID-19 by providing a centralized repository of curated, up-to-date, pre-processed, and publicly-readable datasets focused on the spread and characteristics of the virus. The AWS COVID-19 data lake, which includes data sets from Johns Hopkins University, Definitive Healthcare, Carnegie Mellon’s Delphi Research Group, and other sources, is available for anyone researching, tracking, deploying vital resources, or developing other helpful solutions and applications to combat COVID-19. • AWS is supporting the White House’s COVID-19 High Performance Computing Consortium, providing computing resources to advance research on diagnosis, treatment, and vaccines. • Customers are using AWS to lessen the impact that COVID-19 has on families, communities, and businesses. Examples include: • The New York City COVID-19 Rapid Response Coalition is using a conversational agent, which is running on AWS, to enable at-risk, elderly New Yorkers to receive accurate, timely information about medical needs. • The Los Angeles Unified School District is using AWS to power a new call center that is fielding IT questions, providing remote support, and enabling staff to answer calls around remote learning for 700,000 students. • Volunteer Surge, a nonprofit consortium, is running its online training platform on AWS to recruit, train, and deploy one million volunteer health workers. • The World Health Organization is using AWS to build large-scale data lakes, aggregate epidemiological country data, rapidly translate medical training videos into different languages, and help global healthcare workers better treat patients. • Cerner, a global healthcare technology company, is compiling de-identified patient data to help COVID-19 researchers and is leveraging AWS to secure and store critical information. The data is available free of charge and will support research, vaccine development, and new treatment options, allowing organizations to share information and accelerate understanding of the virus. • In England, the National Health Service is using AWS to analyze hospital occupancy levels, emergency room capacity, and patient wait times in order to help decide where best to allocate resources. • AWS is helping Kentucky and West Virginia authorities address the surge in call volumes to unemployment call centers by transitioning from legacy technology that often required agents to work in the states’ offices to ', 5.076033692769894)\n" ], [ "", "_____no_output_____" ], [ "from google.colab import drive\ndrive.mount('/content/drive')", "Go to this URL in a browser: https://accounts.google.com/o/oauth2/auth?client_id=947318989803-6bn6qk8qdgf4n4g3pfee6491hc0brc4i.apps.googleusercontent.com&redirect_uri=urn%3aietf%3awg%3aoauth%3a2.0%3aoob&scope=email%20https%3a%2f%2fwww.googleapis.com%2fauth%2fdocs.test%20https%3a%2f%2fwww.googleapis.com%2fauth%2fdrive%20https%3a%2f%2fwww.googleapis.com%2fauth%2fdrive.photos.readonly%20https%3a%2f%2fwww.googleapis.com%2fauth%2fpeopleapi.readonly%20https%3a%2f%2fwww.googleapis.com%2fauth%2fdrive.activity.readonly&response_type=code\n\nEnter your authorization code:\n\n" ], [ "!cp ./models/bert_qa_customc.joblib '/content/drive/MyDrive/'", "_____no_output_____" ], [ "\n\n'''import joblib\nimport requests\nimport streamlit as st\n\n# st.set_option('deprecation.showfileUploaderEncoding',False)\nst.title(\"Amazon QA WebApp\")\nst.text(\"What would you like to know regarding amazon today?\")\n\n\[email protected](allow_output_mutation=True)\ndef load_model():\n model=joblib.load('./models/bert_qa_customc.joblib')\n return model\nwith st.spinner('loading model into memory...'):\n model=load_model()\ntext=st.text_input(\"Enter question here\")\nif text:\n st.write('Response:- ')\n with st.spinner('Searching for answers'):\n prediction=model.predict(text)\n st.write('answer: {}'.format(prediction[0]))\n st.write('title: {}'.format(prediction[1]))\n st.write('paragraph: {}'.format(prediction[2]))\n st.write('')\n'''\n", "_____no_output_____" ], [ "\n\nimport joblib\nimport requests\n# import streamlit as st\nfrom flask import Flask, render_template, url_for, request\nimport pandas as pd\nimport pickle\nfrom flask_ngrok import run_with_ngrok\nmodel=joblib.load('./models/bert_qa_customc.joblib')\n# from sklearn.externals import joblib\n\napp = Flask(__name__)\nrun_with_ngrok(app)\n\n\[email protected]('/')\ndef home():\n return 'hii'\n\[email protected]('/predict',methods=['GET'])\ndef predict():\n \n message='amount of jobs created?'\n prediction = model.predict(message)\n return {'Question':message,'Prediction/Answer':list(prediction)}\n\nif __name__ == '__main__':\n app.run()", " * Serving Flask app \"__main__\" (lazy loading)\n * Environment: production\n\u001b[31m WARNING: This is a development server. Do not use it in a production deployment.\u001b[0m\n\u001b[2m Use a production WSGI server instead.\u001b[0m\n * Debug mode: off\n" ], [ "message='How many jobs are created in 2020?'\nprediction = model.predict(message)\n{'query':message,\n 'prediction':list(prediction)}", "_____no_output_____" ], [ "", "_____no_output_____" ] ] ]

[ "markdown", "code" ]

[ [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad6cb669f580077a49f8565f0527654af903593

14,316

ipynb

Jupyter Notebook

Small_Network_Examples/Lima/AequilibraE/GMNS_AE_Integrated.ipynb

zephyr-data-specs/gmns

ee12aa23ee31035a32d9a1ac0e5d1b142897e888

[ "Apache-2.0" ]

52

2020-01-13T13:50:47.000Z

2022-03-30T18:37:25.000Z

Small_Network_Examples/Lima/AequilibraE/GMNS_AE_Integrated.ipynb

zephyr-data-specs/gmns

ee12aa23ee31035a32d9a1ac0e5d1b142897e888

[ "Apache-2.0" ]

52

2020-03-16T17:08:28.000Z

2022-02-04T11:20:38.000Z

Small_Network_Examples/Lima/AequilibraE/GMNS_AE_Integrated.ipynb

zephyr-data-specs/gmns

ee12aa23ee31035a32d9a1ac0e5d1b142897e888

[ "Apache-2.0" ]

16

2020-01-10T19:57:03.000Z

2022-01-09T15:54:36.000Z

36.243038

413

0.545963

[ [ [ "# GMNS to AequilibraE example\n\n## Inputs\n1. Nodes as a .csv flat file in GMNS format\n2. Links as a .csv flat file in GMNS format\n3. Trips as a .csv flat file, with the following columns: orig_node, dest_node, trips\n4. Sqlite database used by AequilibraE\n\n## Steps\n1. Read the GMNS nodes\n - Place in SQLite database, then translate to AequilibraE nodes\n - Generate the dictionary of zones for the omx trip table (uses node_type = centroid)\n\n2. Read the GMNS links\n - Place in SQLite database, then translate to AequilibraE links\n\n3. Read the trips\n - Translate into .omx file\n\nA separate Jupyter notebook, Route, performs the following steps\n\n4. Run AequilibraE shortest path and routing\n\n5. Generate detail and summary outputs\n", "_____no_output_____" ] ], [ [ "#!/usr/bin/env python\n# coding: utf-8\nimport os\nimport numpy as np\nimport pandas as pd\nimport sqlite3\n#import shutil # needed?\nimport openmatrix as omx\nimport math\n#run_folder = 'C:/Users/Scott.Smith/GMNS/Lima'\nrun_folder = 'C:/Users/Scott/Documents/Work/AE/Lima' #Change to match your local environment\n#highest_centroid_node_number = 500 #we are now finding this from the nodes dataframe", "_____no_output_____" ] ], [ [ "## Read the nodes, and set up the dictionary of centroids\nThe dictionary of centroids is used later in setting up the omx trip table", "_____no_output_____" ] ], [ [ "#Read the nodes\nnode_csvfile = os.path.join(run_folder, 'GMNS_node.csv')\ndf_node = pd.read_csv(node_csvfile) #data already has headers\nprint(df_node.head()) #debugging\ndf_size = df_node.shape[0]\nprint(df_size)", " node_id name x_coord y_coord z_coord node_type ctrl_type zone_id \\\n0 1 NaN -84.106107 40.743322 NaN centroid NaN 1 \n1 2 NaN -84.104302 40.743319 NaN centroid NaN 2 \n2 3 NaN -84.102894 40.743284 NaN centroid NaN 3 \n3 4 NaN -84.102975 40.740703 NaN centroid NaN 4 \n4 5 NaN -84.104341 40.742056 NaN centroid NaN 5 \n\n parent_node_id \n0 NaN \n1 NaN \n2 NaN \n3 NaN \n4 NaN \n2232\n" ], [ "# Set up the dictionary of centroids\n# Assumption: the node_type = 'centroid' for centroid nodes\n# The centroid nodes are the lowest numbered nodes, at the beginning of the list of nodes,\n# but node numbers need not be consecutive\ntazdictrow = {}\nfor index in df_node.index:\n if df_node['node_type'][index]=='centroid':\n #DEBUG print(index, df_node['node_id'][index], df_node['node_type'][index])\n tazdictrow[df_node['node_id'][index]]=index\n#tazdictrow = {1:0,2:1,3:2,4:3,...,492:447,493:448}\ntaz_list = list(tazdictrow.keys())\nmatrix_size = len(tazdictrow) #Matches the number of nodes flagged as centroids\nprint(matrix_size) #DEBUG\nhighest_centroid_node_number = max(tazdictrow, key=tazdictrow.get) #for future use\nprint(highest_centroid_node_number) #DEBUG", "449\n493\n" ] ], [ [ "## Read the links", "_____no_output_____" ] ], [ [ "# Read the links\nlink_csvfile = os.path.join(run_folder, 'GMNS_link.csv')\ndf_link = pd.read_csv(link_csvfile) #data already has headers\n#print(df_node.head()) #debugging\n#df_size = df_link.shape[0]\nprint(df_link.shape[0]) #debug", "6095\n" ] ], [ [ "## Put nodes and links into SQLite. Then translate to AequilibraE 0.6.5 format\n1. Nodes are pushed into a table named GMNS_node\n2. node table used by AequilibraE is truncated, then filled with values from GMNS_node\n3. Centroid nodes are assumed to be the lowest numbered nodes, limited by the highest_centroid_node_number\n - Number of centroid nodes must equal matrix_size, the size of the trip OMX Matrix\n3. Links are pushed into a table named GMNS_link\n4. link table used by AequilibraE is truncated, then filled with values from GMNS_link\n\n### Some notes\n1. All the nodes whole id is <= highest_centroid_node_number are set as centroids\n2. GMNS capacity is in veh/hr/lane, AequilibraE is in veh/hr; hence, capacity * lanes in the insert statement\n3. free_flow_time (minutes) is assumed to be 60 (minutes/hr) * length (miles) / free_speed (miles/hr)", "_____no_output_____" ] ], [ [ "#Open the Sqlite database, and insert the nodes and links\nnetwork_db = os.path.join(run_folder,'1_project','Lima.sqlite')\nwith sqlite3.connect(network_db) as db_con:\n #nodes\n df_node.to_sql('GMNS_node',db_con, if_exists='replace',index=False)\n db_cur = db_con.cursor()\n sql0 = \"delete from nodes;\"\n db_cur.execute(sql0)\n sql1 = (\"insert into nodes(ogc_fid, node_id, x, y, is_centroid)\" +\n \" SELECT node_id, node_id, x_coord,y_coord,0 from \" +\n \" GMNS_node\")\n db_cur.execute(sql1)\n sql2 = (\"update nodes set is_centroid = 1 where ogc_fid <= \" + str(highest_centroid_node_number))\n db_cur.execute(sql2)\n \nwith sqlite3.connect(network_db) as db_con:\n df_link.to_sql('GMNS_link',db_con, if_exists='replace',index=False)\n db_cur = db_con.cursor()\n sql0 = \"delete from links;\"\n db_cur.execute(sql0)\n sql1 = (\"insert into links(ogc_fid, link_id, a_node, b_node, direction, distance, modes,\" +\n \" link_type, capacity_ab, speed_ab, free_flow_time) \" +\n \" SELECT link_id, link_id, from_node_id, to_node_id, directed, length, allowed_uses,\" +\n \" facility_type, capacity*lanes, free_speed, 60*length / free_speed\" +\n \" FROM GMNS_link where GMNS_link.capacity > 0\")\n db_cur.execute(sql1)\n sql2 = (\"update links set capacity_ba = 0, speed_ba = 0, b=0.15, power=4\")\n db_cur.execute(sql2)", "_____no_output_____" ] ], [ [ "Next step is to update the links with the parameters for the volume-delay function. This step is AequilibraE-specific and makes use of the link_types Sqlite table. This table is taken from v 0.7.1 of AequilibraE, to ease future compatibility. The link_types table expects at least one row with link_type = \"default\" to use for default values. The user may add other rows with the real link_types. \n\nIts CREATE statement is as follows\n\n```\nCREATE TABLE 'link_types' (link_type VARCHAR UNIQUE NOT NULL PRIMARY KEY,\n link_type_id VARCHAR UNIQUE NOT NULL,\n description VARCHAR,\n lanes NUMERIC,\n lane_capacity NUMERIC,\n alpha NUMERIC,\n beta NUMERIC,\n gamma NUMERIC,\n delta NUMERIC,\n epsilon NUMERIC,\n zeta NUMERIC,\n iota NUMERIC,\n sigma NUMERIC,\n phi NUMERIC,\n tau NUMERIC)\n```\n\n| link_type | link_type_id | description | lanes | lane_capacity | alpha | beta | other fields not used |\n| ----- | ----- | ----- | ----- | ----- |----- |----- |----- |\n| default | 99 | Default general link type | 2 | 900 | 0.15 | 4 | |", "_____no_output_____" ] ], [ [ "with sqlite3.connect(network_db) as db_con:\n db_cur = db_con.cursor()\n sql1 = \"update links set b = (select alpha from link_types where link_type = links.link_type)\"\n db_cur.execute(sql1)\n sql2 = (\"update links set b = (select alpha from link_types where link_type = 'default') where b is NULL\")\n db_cur.execute(sql2)\n sql3 = \"update links set power = (select beta from link_types where link_type = links.link_type)\"\n db_cur.execute(sql3)\n sql4 = (\"update links set power = (select beta from link_types where link_type = 'default') where power is NULL\")\n db_cur.execute(sql4)", "_____no_output_____" ] ], [ [ "## Read the trips, and translate to omx file", "_____no_output_____" ] ], [ [ "#Read a flat file trip table into pandas dataframe\ntrip_csvfile = os.path.join(run_folder, 'demand.csv')\ndf_trip = pd.read_csv(trip_csvfile) #data already has headers\nprint(df_trip.head()) #debugging\ndf_size = df_trip.shape[0]\nprint(df_size)\n#print(df.iloc[50]['o_zone_id'])\n#stuff for debugging\nprint(df_trip['total'].sum()) #for debugging: total number of trips\n#for k in range(df_size): #at most matrix_size*matrix_size\n# i = tazdictrow[df_trip.iloc[k]['orig_taz']]\n# j = tazdictrow[df_trip.iloc[k]['dest_taz']]\n# if k == 4: print(k,\" i=\",i,\" j=\",j) #debugging", " orig_taz dest_taz total\n0 1 57 1\n1 1 138 1\n2 2 2 1\n3 2 287 1\n4 3 67 1\n13000\n32041\n" ], [ "#Write the dataframe to an omx file\n# This makes use of tazdictrow and matrix_size, that was established earlier. \n# The rows are also written to a file that is used only for debugging\noutfile = os.path.join(run_folder, '0_tntp_data' ,'demand.omx') \noutdebugfile = open(os.path.join(run_folder,'debug_demand.txt'),\"w\")\noutput_demand = np.zeros((matrix_size,matrix_size))\nf_output = omx.open_file(outfile,'w')\n\nf_output.create_mapping('taz',taz_list)\n#write the data\nfor k in range(df_size): #at most matrix_size*matrix_size\n i = tazdictrow[df_trip.iloc[k]['orig_taz']]\n j = tazdictrow[df_trip.iloc[k]['dest_taz']]\n \n output_demand[i][j] = df_trip.iloc[k]['total']\n print('Row: ',df_trip.iloc[k]['orig_taz'],i,\" Col: \",df_trip.iloc[k]['dest_taz'],j,\" Output\",output_demand[i][j],file=outdebugfile)\n \nf_output['matrix'] = output_demand #puts the output_demand array into the omx matrix\nf_output.close()\noutdebugfile.close()\n#You may stop here", "_____no_output_____" ], [ "# Not needed except for debugging\n#Read the input omx trip table\ninfile = os.path.join(run_folder, '0_tntp_data' ,'demand.omx') \nf_input = omx.open_file(infile)\nm1 = f_input['matrix']\ninput_demand = np.array(m1)\n\nprint('Shape:',f_input.shape())\nprint('Number of tables',len(f_input))\nprint('Table names:',f_input.list_matrices())\nprint('attributes:',f_input.list_all_attributes())\nprint('sum of trips',np.sum(m1))\nf_input.close()", "Shape: (449, 449)\nNumber of tables 1\nTable names: ['matrix']\nattributes: []\nsum of trips 32041.0\n" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ] ]

4ad6ef2f72169e46c23a48fead0c91831bf40581

178,940

ipynb

Jupyter Notebook

notebooks/10k_Fillings_scrapper.ipynb

vaishalipp/Financial_Reports_NLP

dcd592d5e42cd2d7dca0e2375f608ad12485bdd3

[ "MIT" ]

null

notebooks/10k_Fillings_scrapper.ipynb

vaishalipp/Financial_Reports_NLP

dcd592d5e42cd2d7dca0e2375f608ad12485bdd3

[ "MIT" ]

null

notebooks/10k_Fillings_scrapper.ipynb

vaishalipp/Financial_Reports_NLP

dcd592d5e42cd2d7dca0e2375f608ad12485bdd3

[ "MIT" ]

1

2021-07-12T16:16:16.000Z

116.270305

2,919

0.716201

[ [ [ "import matplotlib.pyplot as plt\nimport requests\nfrom bs4 import BeautifulSoup\nimport pandas as pd\nfrom tqdm import tqdm\n\nfrom ratelimit import limits, sleep_and_retry", "_____no_output_____" ], [ "import Edgar_scrapper", "_____no_output_____" ], [ "import re", "_____no_output_____" ], [ "pd.set_option('display.max_colwidth',200)", "_____no_output_____" ], [ "edgar_access = Edgar_scrapper.EdgarAccess()", "_____no_output_____" ], [ "def get_fillings(fillings_ticker,fillings_cik, doc_type, start=0, count=60):\n fillings_url = 'https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK={}\\\n &type={}&start={}&count={}&owner=exclude&output=atom'.format(fillings_cik, doc_type, start, count)\n fillings_html = edgar_access.get(fillings_url)\n fillings_soup = BeautifulSoup(fillings_html, features=\"html.parser\")\n\n fillings_list = [\n (fillings_ticker,\n fillings_cik,\n link.find('filing-date').getText(),\n link.find('filing-href').getText())\n #link.find('filing-type').getText())\n for link in fillings_soup.find_all('entry')]\n\n return fillings_list", "_____no_output_____" ], [ "ticker_ciks = pd.read_csv('tickers.csv')", "_____no_output_____" ], [ "sample_ticker = pd.DataFrame(data={'ticker' : ['AMC'],'cik':[1411579]})", "_____no_output_____" ], [ "sec_fillings_df = pd.DataFrame(columns=['ticker','cik','date','annual_report_url'])", "_____no_output_____" ], [ "for ticker, cik in sample_ticker[:1].values:\n temp = pd.DataFrame(data = get_fillings(ticker,cik, '10-K'),columns=['ticker','cik','date','annual_report_url'])\n sec_fillings_df = sec_fillings_df.append(temp,ignore_index = True)\n\ndel(temp)", "_____no_output_____" ], [ "sec_fillings_df.annual_report_url = sec_fillings_df.annual_report_url.\\\n replace('-index.htm', '.txt',regex=True)\\\n .replace('.txtl', '.txt',regex=True) ", "_____no_output_____" ], [ "sec_fillings_df", "_____no_output_____" ], [ "sec_fillings_df['filling_text'] = [None] * len(sec_fillings_df)", "_____no_output_____" ], [ "for index,row in tqdm(sec_fillings_df.iterrows(),desc='Downloading Fillings', \\\n unit='filling',total=len(sec_fillings_df)):\n filing_href = row['annual_report_url'] \n report_txt= edgar_access.get(filing_href)\n report_soup = BeautifulSoup(report_txt, \"html\")\n for document in report_soup.find_all('TYPE'): \n if(re.match(r'\\s+10-K',document.prettify().splitlines()[1])):\n #if (document.find('html')):\n sec_fillings_df.iloc[index]['filling_text'] = document#.find('html')", "Downloading Fillings: 100%|██████████| 8/8 [01:24<00:00, 10.59s/filling]\n" ], [ "print(\"Memory consumption of sec_fillings_df is {:.2f}Mb\".format(sec_fillings_df.memory_usage().sum()/1024**2))", "Memory consumption of sec_fillings_df is 0.00Mb\n" ], [ "sec_fillings_df", "_____no_output_____" ], [ "sample_report_txt = sec_fillings_df.iloc[0]['filling_text']", "_____no_output_____" ], [ "result = [p_tag.getText() for p_tag in sample_report_txt.find_all('p',text=True) if re.match(r'\\w+',p_tag.getText())]", "_____no_output_____" ], [ "result[:2000]", "_____no_output_____" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad6f2b000e15460ec3a40469dbd5ffadcdf7c04

133,176

ipynb

Jupyter Notebook

Models/Model5.ipynb

AmiteshBadkul/eeg-signal-classification

5477f53981dafef71cecbe6129485df3976f734d

[ "MIT" ]

null

Models/Model5.ipynb

AmiteshBadkul/eeg-signal-classification

5477f53981dafef71cecbe6129485df3976f734d

[ "MIT" ]

null

Models/Model5.ipynb

AmiteshBadkul/eeg-signal-classification

5477f53981dafef71cecbe6129485df3976f734d

[ "MIT" ]

null

145.071895

67,502

0.811663

[ [ [ "%reset -f ", "_____no_output_____" ], [ "# libraries used\n# https://stats.stackexchange.com/questions/181/how-to-choose-the-number-of-hidden-layers-and-nodes-in-a-feedforward-neural-netw\nimport numpy as np\nimport pandas as pd\nimport seaborn as sns \nimport matplotlib.pyplot as plt\nfrom sklearn.model_selection import train_test_split\nfrom sklearn.metrics import confusion_matrix, classification_report\nfrom sklearn import preprocessing\nimport tensorflow as tf\nfrom tensorflow.keras import datasets, layers, models\nimport keras \nfrom keras.models import Sequential\nfrom keras.layers import Dense, Dropout\nimport itertools", "_____no_output_____" ], [ "emotions = pd.read_csv(\"drive/MyDrive/EEG/emotions.csv\")\nemotions.replace(['NEGATIVE', 'POSITIVE', 'NEUTRAL'], [2, 1, 0], inplace=True)\nemotions['label'].unique()", "_____no_output_____" ], [ "X = emotions.drop('label', axis=1).copy()\ny = (emotions['label'].copy())", "_____no_output_____" ], [ "# Splitting data into training and testing as 80-20 \nX_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=1)\n\nx = X_train #returns a numpy array\nmin_max_scaler = preprocessing.MinMaxScaler()\nx_scaled = min_max_scaler.fit_transform(x)\ndf = pd.DataFrame(x_scaled)", "_____no_output_____" ], [ "# resetting the data - https://www.tensorflow.org/api_docs/python/tf/keras/backend/clear_session\ntf.keras.backend.clear_session()", "_____no_output_____" ], [ "model = Sequential()\nmodel.add(Dense((2*X_train.shape[1]/3), input_dim=X_train.shape[1], activation='relu'))\nmodel.add(Dense((2*X_train.shape[1]/3), activation='relu'))\nmodel.add(Dense((1*X_train.shape[1]/3), activation='relu'))\nmodel.add(Dense((1*X_train.shape[1]/3), activation='relu'))\nmodel.add(Dense(3, activation='softmax'))\n#model.compile(loss='categorical_crossentropy', optimizer='adam', metrics=['accuracy'])\nprint(model.summary())\n\n\n\n", "Model: \"sequential\"\n_________________________________________________________________\nLayer (type) Output Shape Param # \n=================================================================\ndense (Dense) (None, 1698) 4328202 \n_________________________________________________________________\ndense_1 (Dense) (None, 1698) 2884902 \n_________________________________________________________________\ndense_2 (Dense) (None, 849) 1442451 \n_________________________________________________________________\ndense_3 (Dense) (None, 849) 721650 \n_________________________________________________________________\ndense_4 (Dense) (None, 3) 2550 \n=================================================================\nTotal params: 9,379,755\nTrainable params: 9,379,755\nNon-trainable params: 0\n_________________________________________________________________\nNone\n" ], [ "# for categorical entropy\n# https://stackoverflow.com/questions/63211181/error-while-using-categorical-crossentropy\nfrom tensorflow.keras.utils import to_categorical\nY_one_hot=to_categorical(y_train) # convert Y into an one-hot vector", "_____no_output_____" ], [ "# https://stackoverflow.com/questions/59737875/keras-change-learning-rate\n#optimizer = tf.keras.optimizers.Adam(0.001)\n#optimizer.learning_rate.assign(0.01)\n\nopt = keras.optimizers.Adamax(learning_rate=0.001)\n\nmodel.compile(\n optimizer=opt,\n loss='sparse_categorical_crossentropy',\n metrics=['accuracy']\n)\n\n# to be run for categorical cross entropy\n# model.compile(loss='categorical_crossentropy', optimizer=tf.keras.optimizers.Adam(lr=0.01), metrics=['accuracy'])\n", "_____no_output_____" ], [ "# make sure that the input data is shuffled before hand so that the model doesn't notice patterns and generalizes well\n# change y_train to y_hot_encoded when using categorical cross entorpy\nimport time\nstart_time = time.time()\nhistory = model.fit(\n df,\n y_train,\n validation_split=0.2,\n batch_size=32,\n epochs=75)", "Epoch 1/75\n43/43 [==============================] - 5s 95ms/step - loss: 1.5793 - accuracy: 0.4826 - val_loss: 0.5238 - val_accuracy: 0.8915\nEpoch 2/75\n43/43 [==============================] - 4s 89ms/step - loss: 0.4922 - accuracy: 0.8225 - val_loss: 0.3254 - val_accuracy: 0.8886\nEpoch 3/75\n43/43 [==============================] - 4s 102ms/step - loss: 0.4242 - accuracy: 0.8492 - val_loss: 0.5241 - val_accuracy: 0.7390\nEpoch 4/75\n43/43 [==============================] - 5s 124ms/step - loss: 0.3742 - accuracy: 0.8520 - val_loss: 0.2720 - val_accuracy: 0.8974\nEpoch 5/75\n43/43 [==============================] - 4s 93ms/step - loss: 0.3886 - accuracy: 0.8466 - val_loss: 0.2901 - val_accuracy: 0.8856\nEpoch 6/75\n43/43 [==============================] - 4s 92ms/step - loss: 0.3564 - accuracy: 0.8569 - val_loss: 0.2733 - val_accuracy: 0.9062\nEpoch 7/75\n43/43 [==============================] - 4s 86ms/step - loss: 0.3365 - accuracy: 0.8634 - val_loss: 0.4045 - val_accuracy: 0.8328\nEpoch 8/75\n43/43 [==============================] - 4s 96ms/step - loss: 0.3571 - accuracy: 0.8549 - val_loss: 0.2266 - val_accuracy: 0.9296\nEpoch 9/75\n43/43 [==============================] - 4s 88ms/step - loss: 0.2874 - accuracy: 0.9051 - val_loss: 0.6655 - val_accuracy: 0.7537\nEpoch 10/75\n43/43 [==============================] - 4s 85ms/step - loss: 0.3789 - accuracy: 0.8600 - val_loss: 0.3118 - val_accuracy: 0.8798\nEpoch 11/75\n43/43 [==============================] - 4s 98ms/step - loss: 0.2882 - accuracy: 0.9007 - val_loss: 0.3268 - val_accuracy: 0.8475\nEpoch 12/75\n43/43 [==============================] - 4s 92ms/step - loss: 0.3714 - accuracy: 0.8561 - val_loss: 0.1988 - val_accuracy: 0.9326\nEpoch 13/75\n43/43 [==============================] - 4s 92ms/step - loss: 0.2865 - accuracy: 0.8982 - val_loss: 0.3110 - val_accuracy: 0.8798\nEpoch 14/75\n43/43 [==============================] - 4s 85ms/step - loss: 0.3114 - accuracy: 0.8862 - val_loss: 0.2069 - val_accuracy: 0.9296\nEpoch 15/75\n43/43 [==============================] - 4s 85ms/step - loss: 0.3094 - accuracy: 0.8817 - val_loss: 0.2043 - val_accuracy: 0.9296\nEpoch 16/75\n43/43 [==============================] - 4s 85ms/step - loss: 0.2705 - accuracy: 0.8915 - val_loss: 0.2936 - val_accuracy: 0.8651\nEpoch 17/75\n43/43 [==============================] - 4s 93ms/step - loss: 0.2856 - accuracy: 0.9011 - val_loss: 0.5897 - val_accuracy: 0.7713\nEpoch 18/75\n43/43 [==============================] - 4s 87ms/step - loss: 0.4381 - accuracy: 0.8204 - val_loss: 0.2727 - val_accuracy: 0.8915\nEpoch 19/75\n43/43 [==============================] - 4s 91ms/step - loss: 0.2904 - accuracy: 0.8902 - val_loss: 0.1962 - val_accuracy: 0.9326\nEpoch 20/75\n43/43 [==============================] - 4s 95ms/step - loss: 0.2653 - accuracy: 0.9119 - val_loss: 0.2961 - val_accuracy: 0.8710\nEpoch 21/75\n43/43 [==============================] - 4s 95ms/step - loss: 0.2910 - accuracy: 0.8830 - val_loss: 0.2271 - val_accuracy: 0.9296\nEpoch 22/75\n43/43 [==============================] - 4s 101ms/step - loss: 0.2354 - accuracy: 0.9140 - val_loss: 0.2866 - val_accuracy: 0.8798\nEpoch 23/75\n43/43 [==============================] - 4s 102ms/step - loss: 0.3065 - accuracy: 0.8746 - val_loss: 0.3057 - val_accuracy: 0.8622\nEpoch 24/75\n43/43 [==============================] - 4s 92ms/step - loss: 0.2279 - accuracy: 0.9159 - val_loss: 0.1812 - val_accuracy: 0.9267\nEpoch 25/75\n43/43 [==============================] - 4s 89ms/step - loss: 0.2224 - accuracy: 0.9216 - val_loss: 0.1933 - val_accuracy: 0.9267\nEpoch 26/75\n43/43 [==============================] - 4s 101ms/step - loss: 0.1902 - accuracy: 0.9348 - val_loss: 0.2797 - val_accuracy: 0.8827\nEpoch 27/75\n43/43 [==============================] - 4s 102ms/step - loss: 0.2246 - accuracy: 0.9169 - val_loss: 0.2517 - val_accuracy: 0.8974\nEpoch 28/75\n43/43 [==============================] - 4s 87ms/step - loss: 0.2193 - accuracy: 0.9173 - val_loss: 0.1996 - val_accuracy: 0.9355\nEpoch 29/75\n43/43 [==============================] - 5s 106ms/step - loss: 0.2051 - accuracy: 0.9156 - val_loss: 0.1775 - val_accuracy: 0.9296\nEpoch 30/75\n43/43 [==============================] - 4s 87ms/step - loss: 0.2106 - accuracy: 0.9164 - val_loss: 0.1491 - val_accuracy: 0.9472\nEpoch 31/75\n43/43 [==============================] - 4s 86ms/step - loss: 0.1790 - accuracy: 0.9328 - val_loss: 0.1373 - val_accuracy: 0.9472\nEpoch 32/75\n43/43 [==============================] - 4s 93ms/step - loss: 0.1982 - accuracy: 0.9176 - val_loss: 0.3198 - val_accuracy: 0.8299\nEpoch 33/75\n43/43 [==============================] - 4s 89ms/step - loss: 0.2161 - accuracy: 0.9129 - val_loss: 0.1808 - val_accuracy: 0.9267\nEpoch 34/75\n43/43 [==============================] - 5s 108ms/step - loss: 0.2114 - accuracy: 0.9210 - val_loss: 0.2229 - val_accuracy: 0.9150\nEpoch 35/75\n43/43 [==============================] - 4s 92ms/step - loss: 0.1655 - accuracy: 0.9416 - val_loss: 0.1742 - val_accuracy: 0.9208\nEpoch 36/75\n43/43 [==============================] - 5s 106ms/step - loss: 0.1356 - accuracy: 0.9510 - val_loss: 0.1408 - val_accuracy: 0.9501\nEpoch 37/75\n43/43 [==============================] - 4s 99ms/step - loss: 0.1478 - accuracy: 0.9531 - val_loss: 0.1835 - val_accuracy: 0.9179\nEpoch 38/75\n43/43 [==============================] - 4s 86ms/step - loss: 0.1424 - accuracy: 0.9490 - val_loss: 0.1514 - val_accuracy: 0.9472\nEpoch 39/75\n43/43 [==============================] - 4s 85ms/step - loss: 0.1790 - accuracy: 0.9387 - val_loss: 0.1410 - val_accuracy: 0.9501\nEpoch 40/75\n43/43 [==============================] - 4s 93ms/step - loss: 0.1406 - accuracy: 0.9427 - val_loss: 0.1365 - val_accuracy: 0.9472\nEpoch 41/75\n43/43 [==============================] - 4s 94ms/step - loss: 0.1382 - accuracy: 0.9469 - val_loss: 0.1281 - val_accuracy: 0.9443\nEpoch 42/75\n43/43 [==============================] - 4s 95ms/step - loss: 0.1260 - accuracy: 0.9555 - val_loss: 0.1102 - val_accuracy: 0.9677\nEpoch 43/75\n43/43 [==============================] - 4s 86ms/step - loss: 0.1359 - accuracy: 0.9444 - val_loss: 0.1168 - val_accuracy: 0.9736\nEpoch 44/75\n43/43 [==============================] - 4s 97ms/step - loss: 0.1274 - accuracy: 0.9561 - val_loss: 0.1599 - val_accuracy: 0.9443\nEpoch 45/75\n43/43 [==============================] - 5s 107ms/step - loss: 0.1231 - accuracy: 0.9502 - val_loss: 0.1080 - val_accuracy: 0.9560\nEpoch 46/75\n43/43 [==============================] - 5s 110ms/step - loss: 0.1197 - accuracy: 0.9555 - val_loss: 0.2282 - val_accuracy: 0.9091\nEpoch 47/75\n43/43 [==============================] - 4s 101ms/step - loss: 0.1329 - accuracy: 0.9528 - val_loss: 0.0980 - val_accuracy: 0.9677\nEpoch 48/75\n43/43 [==============================] - 4s 90ms/step - loss: 0.0700 - accuracy: 0.9759 - val_loss: 0.0969 - val_accuracy: 0.9707\nEpoch 49/75\n43/43 [==============================] - 4s 87ms/step - loss: 0.0691 - accuracy: 0.9765 - val_loss: 0.1118 - val_accuracy: 0.9589\nEpoch 50/75\n43/43 [==============================] - 4s 93ms/step - loss: 0.0910 - accuracy: 0.9666 - val_loss: 0.1216 - val_accuracy: 0.9648\nEpoch 51/75\n43/43 [==============================] - 4s 90ms/step - loss: 0.0959 - accuracy: 0.9600 - val_loss: 0.1685 - val_accuracy: 0.9413\nEpoch 52/75\n43/43 [==============================] - 4s 95ms/step - loss: 0.1721 - accuracy: 0.9402 - val_loss: 0.1108 - val_accuracy: 0.9531\nEpoch 53/75\n43/43 [==============================] - 4s 101ms/step - loss: 0.0948 - accuracy: 0.9630 - val_loss: 0.0979 - val_accuracy: 0.9589\nEpoch 54/75\n43/43 [==============================] - 5s 108ms/step - loss: 0.0496 - accuracy: 0.9886 - val_loss: 0.0993 - val_accuracy: 0.9707\nEpoch 55/75\n43/43 [==============================] - 4s 98ms/step - loss: 0.0595 - accuracy: 0.9785 - val_loss: 0.2111 - val_accuracy: 0.9384\nEpoch 56/75\n43/43 [==============================] - 4s 103ms/step - loss: 0.1778 - accuracy: 0.9447 - val_loss: 0.1154 - val_accuracy: 0.9619\nEpoch 57/75\n43/43 [==============================] - 4s 93ms/step - loss: 0.0717 - accuracy: 0.9692 - val_loss: 0.0836 - val_accuracy: 0.9707\nEpoch 58/75\n43/43 [==============================] - 4s 96ms/step - loss: 0.0622 - accuracy: 0.9794 - val_loss: 0.1027 - val_accuracy: 0.9648\nEpoch 59/75\n43/43 [==============================] - 4s 96ms/step - loss: 0.0718 - accuracy: 0.9762 - val_loss: 0.0815 - val_accuracy: 0.9824\nEpoch 60/75\n43/43 [==============================] - 4s 86ms/step - loss: 0.0439 - accuracy: 0.9879 - val_loss: 0.1117 - val_accuracy: 0.9619\nEpoch 61/75\n43/43 [==============================] - 4s 92ms/step - loss: 0.1192 - accuracy: 0.9554 - val_loss: 0.1380 - val_accuracy: 0.9531\nEpoch 62/75\n43/43 [==============================] - 4s 87ms/step - loss: 0.0479 - accuracy: 0.9828 - val_loss: 0.3742 - val_accuracy: 0.8387\nEpoch 63/75\n43/43 [==============================] - 4s 93ms/step - loss: 0.1110 - accuracy: 0.9598 - val_loss: 0.0938 - val_accuracy: 0.9589\nEpoch 64/75\n43/43 [==============================] - 4s 89ms/step - loss: 0.0372 - accuracy: 0.9868 - val_loss: 0.1304 - val_accuracy: 0.9501\nEpoch 65/75\n43/43 [==============================] - 5s 110ms/step - loss: 0.0505 - accuracy: 0.9801 - val_loss: 0.1164 - val_accuracy: 0.9677\nEpoch 66/75\n43/43 [==============================] - 4s 93ms/step - loss: 0.0351 - accuracy: 0.9894 - val_loss: 0.1201 - val_accuracy: 0.9707\nEpoch 67/75\n43/43 [==============================] - 4s 91ms/step - loss: 0.1285 - accuracy: 0.9534 - val_loss: 0.1580 - val_accuracy: 0.9501\nEpoch 68/75\n43/43 [==============================] - 5s 111ms/step - loss: 0.0605 - accuracy: 0.9810 - val_loss: 0.0664 - val_accuracy: 0.9883\nEpoch 69/75\n43/43 [==============================] - 4s 103ms/step - loss: 0.0315 - accuracy: 0.9930 - val_loss: 0.2266 - val_accuracy: 0.9062\nEpoch 70/75\n43/43 [==============================] - 4s 86ms/step - loss: 0.1003 - accuracy: 0.9576 - val_loss: 0.0962 - val_accuracy: 0.9648\nEpoch 71/75\n43/43 [==============================] - 4s 93ms/step - loss: 0.0805 - accuracy: 0.9645 - val_loss: 0.1658 - val_accuracy: 0.9326\nEpoch 72/75\n43/43 [==============================] - 4s 87ms/step - loss: 0.0597 - accuracy: 0.9754 - val_loss: 0.1268 - val_accuracy: 0.9560\nEpoch 73/75\n43/43 [==============================] - 4s 86ms/step - loss: 0.0496 - accuracy: 0.9787 - val_loss: 0.0748 - val_accuracy: 0.9736\nEpoch 74/75\n43/43 [==============================] - 4s 99ms/step - loss: 0.0474 - accuracy: 0.9831 - val_loss: 0.3915 - val_accuracy: 0.8915\nEpoch 75/75\n43/43 [==============================] - 4s 99ms/step - loss: 0.1708 - accuracy: 0.9501 - val_loss: 0.1537 - val_accuracy: 0.9472\n" ], [ "history.history", "_____no_output_____" ], [ "print(\"--- %s seconds ---\" % (time.time() - start_time))\n", "--- 323.1897494792938 seconds ---\n" ], [ "x_test = X_test #returns a numpy array\nmin_max_scaler = preprocessing.MinMaxScaler()\nx_scaled_test = min_max_scaler.fit_transform(x_test)\ndf_test = pd.DataFrame(x_scaled_test)", "_____no_output_____" ], [ "predictions = model.predict(x=df_test, batch_size=32)", "_____no_output_____" ], [ "rounded_predictions = np.argmax(predictions, axis=-1)", "_____no_output_____" ], [ "cm = confusion_matrix(y_true=y_test, y_pred=rounded_predictions)", "_____no_output_____" ], [ "label_mapping = {'NEGATIVE': 0, 'NEUTRAL': 1, 'POSITIVE': 2}\n# for diff dataset\n# label_mapping = {'NEGATIVE': 0, 'POSITIVE': 1}\nplt.figure(figsize=(8, 8))\nsns.heatmap(cm, annot=True, vmin=0, fmt='g', cbar=False, cmap='Blues')\nclr = classification_report(y_test, rounded_predictions, target_names=label_mapping.keys())\nplt.xticks(np.arange(3) + 0.5, label_mapping.keys())\nplt.yticks(np.arange(3) + 0.5, label_mapping.keys())\nplt.xlabel(\"Predicted\")\nplt.ylabel(\"Actual\")\nplt.title(\"Confusion Matrix\")\nplt.show()\n\nprint(\"Classification Report:\\n----------------------\\n\", clr)\n\n", "_____no_output_____" ], [ "# https://stackoverflow.com/questions/26413185/how-to-recover-matplotlib-defaults-after-setting-stylesheet\nimport matplotlib as mpl\nmpl.rcParams.update(mpl.rcParamsDefault)", "_____no_output_____" ], [ "training_acc = history.history['accuracy']\nvalidation_acc = history.history['val_accuracy']\ntraining_loss = history.history['loss']\nvalidation_loss = history.history['val_loss']", "_____no_output_____" ], [ "epochs = history.epoch\nplt.plot(epochs, training_acc, color = '#17e6e6', label='Training Accuracy')\nplt.plot(epochs, validation_acc,color = '#e61771', label='Validation Accuracy')\nplt.title('Accuracy vs Epochs')\nplt.xlabel('Epochs')\nplt.ylabel('Accuracy')\nplt.legend()\nplt.savefig('AccuracyVsEpochs.png')\nplt.show()", "_____no_output_____" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad6fd09e6e4cbd021c54aa8ecf640e7d67463cc

146,145

ipynb

Jupyter Notebook

05_DeepLearning/02_Deep_Learning_for_Computer_Vision.ipynb

talitacardenas/The_Bridge_School_DataScience_PT

7c059d06a0eb53c0370d1db8868e0e7cb88c857b

[ "Apache-2.0" ]

null

05_DeepLearning/02_Deep_Learning_for_Computer_Vision.ipynb

talitacardenas/The_Bridge_School_DataScience_PT

7c059d06a0eb53c0370d1db8868e0e7cb88c857b

[ "Apache-2.0" ]

11

2021-05-10T19:16:00.000Z

2021-07-22T15:22:22.000Z

05_DeepLearning/02_Deep_Learning_for_Computer_Vision.ipynb

talitacardenas/The_Bridge_School_DataScience_PT

7c059d06a0eb53c0370d1db8868e0e7cb88c857b

[ "Apache-2.0" ]

8

2021-05-10T18:11:12.000Z

2021-11-23T14:31:38.000Z

78.277986

17,082

0.72697

[ [ [ "Clasificación de las imágenes.\n\nThe problem: clasificación del dataset MNIST\n- clasificación en escala de grises\n- dígitos handwritten\n- 28x28px\n- 10 categorías (0-9)", "_____no_output_____" ] ], [ [ "# tensorflow low level library\n# keras high level library\nfrom tensorflow import keras\nfrom tensorflow.keras import models\nfrom tensorflow.keras import layers\nimport matplotlib.pyplot as plt\nimport numpy as np", "_____no_output_____" ], [ "", "_____no_output_____" ], [ "# importamos el dataset\nfrom tensorflow.keras.datasets import mnist", "_____no_output_____" ], [ "(train_images, train_labels), (test_images, test_labels) = mnist.load_data()", "Downloading data from https://storage.googleapis.com/tensorflow/tf-keras-datasets/mnist.npz\n11493376/11490434 [==============================] - 0s 0us/step\n11501568/11490434 [==============================] - 0s 0us/step\n" ], [ "# Observamos el tipo de dato\ntype(train_images)", "_____no_output_____" ], [ "train_images.shape", "_____no_output_____" ], [ "train_images[0]", "_____no_output_____" ], [ "plt.imshow(train_images[0], cmap='gray')", "_____no_output_____" ], [ "train_labels[0]", "_____no_output_____" ], [ "# Imágenes en formato 28x28 pixel en formato matrices\ntrain_images", "_____no_output_____" ], [ "# Imágenes etiquetadas\ntrain_labels", "_____no_output_____" ], [ "# explorando las 10 imágenes\nn_images = 10\nfig, axs = plt.subplots(1, n_images, figsize=(20,20))\nfor i in range(n_images):\n axs[i].imshow(train_images[i], cmap='gray')", "_____no_output_____" ], [ "# explorando las 10 etiquetas\ntrain_labels[:n_images]", "_____no_output_____" ] ], [ [ "## Entrenando un modelo clásico de ML", "_____no_output_____" ] ], [ [ "from sklearn.ensemble import GradientBoostingClassifier\nfrom sklearn.model_selection import GridSearchCV", "_____no_output_____" ], [ "train_images[0].shape", "_____no_output_____" ], [ "X_train = train_images.reshape(train_images.shape[0], -1)\nX_test = test_images.reshape(test_images.shape[0], -1)", "_____no_output_____" ], [ "train_images.shape", "_____no_output_____" ], [ "train_images[0].shape", "_____no_output_____" ], [ "X_train.shape", "_____no_output_____" ], [ "X_train[0].shape", "_____no_output_____" ], [ "# training\nmodel = GradientBoostingClassifier(n_estimators=10,\n max_depth=5,\n max_features=0.1)", "_____no_output_____" ], [ "model.fit(X_train, train_labels)", "_____no_output_____" ], [ "# predicting\nmodel.predict(X_test[:10])", "_____no_output_____" ], [ "n_images = 10\nfig, axs = plt.subplots(1, n_images, figsize=(20,20))\nfor i in range(n_images):\n axs[i].imshow(test_images[i], cmap='gray')", "_____no_output_____" ], [ "# explorando las 10 etiquetas\ntest_labels[:n_images]", "_____no_output_____" ], [ "# Realizamos el check de los errores\nimport numpy as np\nerror_indices = np.argwhere(test_labels[:1000] != model.predict(X_test[:1000]))", "_____no_output_____" ], [ "n_images = 10\nfig, axs = plt.subplots(1, n_images, figsize=(20,20))\nfor i, index in zip(range(n_images), error_indices):\n axs[i].imshow(test_images[index][0], cmap='gray')", "_____no_output_____" ], [ "for i in error_indices[:10]:\n print(model.predict(X_test[i].reshape(1,-1)))", "[4]\n[4]\n[0]\n[3]\n[2]\n[2]\n[9]\n[7]\n[9]\n[8]\n" ], [ "# Accuracy Score\nmodel.score(X_train, train_labels)", "_____no_output_____" ], [ "# aplicando el score en los valores de test\nmodel.score(X_test, test_labels)", "_____no_output_____" ] ], [ [ "# Aplicando las ANN (Artificial Neural Network)\n\nSeguimos el workflow:\n- primero creamos el modelo con los datos de training (images, labels)\n- la red neuronal aprenderá de las imagénes y etiquetas\n- finalmente, creamos la predicción para los test_images\n- verificamos con las predicciones nuestras test_labels", "_____no_output_____" ] ], [ [ "# creamos nuestra red\nnetwork = models.Sequential()", "_____no_output_____" ] ], [ [ "# Construimos las capas\n\n- dos capas Dense layer, que estarán *densely-connected* (también 'fully-connected').\n- Una de las capas tendrá 10 salidas con la función 'softmax'\n- esta última devolverá un array de 10 scoring (sumando hará 1).\n- cada puntuacuón será la probabilidad de que el dígito real se encuentre entre las 10 clases.", "_____no_output_____" ] ], [ [ "network.add(layers.Dense(256, activation='relu', input_shape=(784,)))\nnetwork.add(layers.Dense(10, activation='softmax'))", "_____no_output_____" ], [ "# compilamos\nnetwork.compile(\n optimizer='rmsprop',\n loss='categorical_crossentropy',\n metrics=['accuracy']\n)", "_____no_output_____" ] ], [ [ "- necesitamos realizar el reshape de cada una de los 28x28 img a un vector de 784\n- escalar los valores en un intervalo [0, 1]", "_____no_output_____" ] ], [ [ "train_vectors = train_images.reshape((60000, 28 * 28)).astype('float32') / 255\ntest_vectors = test_images.reshape((10000, 28 * 28)).astype('float32') / 255", "_____no_output_____" ], [ "train_vectors.shape", "_____no_output_____" ], [ "# también necesitamos transformar las etiquetas en categóricas\nfrom tensorflow.keras.utils import to_categorical", "_____no_output_____" ], [ "train_labels", "_____no_output_____" ], [ "train_labels_hot = to_categorical(train_labels)\ntest_labels_hot = to_categorical(test_labels)", "_____no_output_____" ], [ "train_labels_hot", "_____no_output_____" ], [ "# comprobamos la transformación a categórica\ntrain_labels[10]", "_____no_output_____" ], [ "train_labels_hot[10]", "_____no_output_____" ], [ "# resumen de nuestra red neuronal\nnetwork.summary()", "Model: \"sequential\"\n_________________________________________________________________\n Layer (type) Output Shape Param # \n=================================================================\n dense (Dense) (None, 256) 200960 \n \n dense_1 (Dense) (None, 10) 2570 \n \n=================================================================\nTotal params: 203,530\nTrainable params: 203,530\nNon-trainable params: 0\n_________________________________________________________________\n" ], [ "%%time\nnetwork.fit(train_vectors, train_labels_hot,\n epochs=15, batch_size=128,\n validation_split=0.1)\n# no GPU - 42.6 sec", "Epoch 1/15\n422/422 [==============================] - 4s 5ms/step - loss: 0.3056 - accuracy: 0.9139 - val_loss: 0.1389 - val_accuracy: 0.9597\nEpoch 2/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.1361 - accuracy: 0.9599 - val_loss: 0.0953 - val_accuracy: 0.9722\nEpoch 3/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0922 - accuracy: 0.9730 - val_loss: 0.0834 - val_accuracy: 0.9758\nEpoch 4/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0692 - accuracy: 0.9790 - val_loss: 0.0780 - val_accuracy: 0.9767\nEpoch 5/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0533 - accuracy: 0.9843 - val_loss: 0.0764 - val_accuracy: 0.9798\nEpoch 6/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0430 - accuracy: 0.9871 - val_loss: 0.0769 - val_accuracy: 0.9783\nEpoch 7/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0346 - accuracy: 0.9899 - val_loss: 0.0725 - val_accuracy: 0.9797\nEpoch 8/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0284 - accuracy: 0.9918 - val_loss: 0.0770 - val_accuracy: 0.9798\nEpoch 9/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0229 - accuracy: 0.9936 - val_loss: 0.0817 - val_accuracy: 0.9805\nEpoch 10/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0184 - accuracy: 0.9949 - val_loss: 0.0798 - val_accuracy: 0.9803\nEpoch 11/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0153 - accuracy: 0.9959 - val_loss: 0.0756 - val_accuracy: 0.9808\nEpoch 12/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0122 - accuracy: 0.9969 - val_loss: 0.0770 - val_accuracy: 0.9820\nEpoch 13/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0100 - accuracy: 0.9976 - val_loss: 0.0865 - val_accuracy: 0.9792\nEpoch 14/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0077 - accuracy: 0.9981 - val_loss: 0.0879 - val_accuracy: 0.9798\nEpoch 15/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0066 - accuracy: 0.9983 - val_loss: 0.0866 - val_accuracy: 0.9812\nCPU times: user 29.6 s, sys: 5.08 s, total: 34.7 s\nWall time: 29.8 s\n" ], [ "%%time\nnetwork.fit(train_vectors, train_labels_hot,\n epochs=15, batch_size=128,\n validation_split=0.1)\n# with GPU - 29.6s", "Epoch 1/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0050 - accuracy: 0.9989 - val_loss: 0.0852 - val_accuracy: 0.9815\nEpoch 2/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0042 - accuracy: 0.9990 - val_loss: 0.0926 - val_accuracy: 0.9805\nEpoch 3/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0033 - accuracy: 0.9993 - val_loss: 0.1016 - val_accuracy: 0.9798\nEpoch 4/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0027 - accuracy: 0.9994 - val_loss: 0.0983 - val_accuracy: 0.9827\nEpoch 5/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0024 - accuracy: 0.9994 - val_loss: 0.0938 - val_accuracy: 0.9818\nEpoch 6/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0018 - accuracy: 0.9996 - val_loss: 0.0976 - val_accuracy: 0.9830\nEpoch 7/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0014 - accuracy: 0.9997 - val_loss: 0.1083 - val_accuracy: 0.9827\nEpoch 8/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0012 - accuracy: 0.9997 - val_loss: 0.1112 - val_accuracy: 0.9803\nEpoch 9/15\n422/422 [==============================] - 2s 4ms/step - loss: 0.0011 - accuracy: 0.9998 - val_loss: 0.1170 - val_accuracy: 0.9812\nEpoch 10/15\n422/422 [==============================] - 2s 4ms/step - loss: 7.8512e-04 - accuracy: 0.9998 - val_loss: 0.1137 - val_accuracy: 0.9820\nEpoch 11/15\n422/422 [==============================] - 2s 4ms/step - loss: 5.0787e-04 - accuracy: 0.9999 - val_loss: 0.1205 - val_accuracy: 0.9822\nEpoch 12/15\n422/422 [==============================] - 2s 5ms/step - loss: 4.8522e-04 - accuracy: 0.9999 - val_loss: 0.1201 - val_accuracy: 0.9817\nEpoch 13/15\n422/422 [==============================] - 2s 4ms/step - loss: 3.9790e-04 - accuracy: 0.9999 - val_loss: 0.1198 - val_accuracy: 0.9822\nEpoch 14/15\n422/422 [==============================] - 2s 4ms/step - loss: 2.9533e-04 - accuracy: 1.0000 - val_loss: 0.1259 - val_accuracy: 0.9807\nEpoch 15/15\n422/422 [==============================] - 2s 4ms/step - loss: 2.6651e-04 - accuracy: 1.0000 - val_loss: 0.1225 - val_accuracy: 0.9827\nCPU times: user 29.2 s, sys: 4.98 s, total: 34.2 s\nWall time: 41.1 s\n" ], [ "plt.plot(network.history.history['accuracy'], label='train')\nplt.plot(network.history.history['val_accuracy'], label='validation')\nplt.legend()", "_____no_output_____" ], [ "_, test_acc = network.evaluate(test_vectors, test_labels_hot)", "313/313 [==============================] - 1s 3ms/step - loss: 0.1168 - accuracy: 0.9818\n" ], [ "test_acc", "_____no_output_____" ], [ "# comprobaciones\nn_images = 10\nfig, axs = plt.subplots(1, n_images, figsize=(20,20))\nfor i in range(n_images):\n axs[i].imshow(test_images[i], cmap='gray')", "_____no_output_____" ], [ "test_labels[:10]", "_____no_output_____" ], [ "np.argmax(network.predict(test_vectors), axis=-1)[:10]", "_____no_output_____" ], [ "# Comprobación de los errores\nerrores_indices = np.argwhere(test_labels[:1000] != np.argmax(network.predict(test_vectors), axis=-1)[:1000]).flatten()", "_____no_output_____" ], [ "errores_indices", "_____no_output_____" ], [ "n_images = 10\nfig, axs = plt.subplots(1, n_images, figsize=(20,20))\nfor i, index in zip(range(n_images), errores_indices):\n axs[i].imshow(test_images[index], cmap='gray')", "_____no_output_____" ], [ "test_labels[errores_indices][:10]", "_____no_output_____" ], [ "np.argmax(network.predict(test_vectors[errores_indices]), axis=-1)[:10]", "_____no_output_____" ] ], [ [ "# CNN - Convolutional neural network", "_____no_output_____" ] ], [ [ "model = models.Sequential()", "_____no_output_____" ], [ "# creamos las capas ocultas\nmodel.add(layers.Conv2D(32, (3,3), activation='relu', input_shape=(28,28,1)))\nmodel.add(layers.MaxPooling2D((2,2)))\nmodel.add(layers.Conv2D(64, (3,3), activation='relu'))\nmodel.add(layers.MaxPooling2D((2,2)))\nmodel.add(layers.Conv2D(64, (3,3), activation='relu'))", "_____no_output_____" ], [ "model.add(layers.Flatten())\nmodel.add(layers.Dense(64, activation='relu'))\nmodel.add(layers.Dense(10, activation='softmax'))", "_____no_output_____" ], [ "model.summary()", "Model: \"sequential_1\"\n_________________________________________________________________\n Layer (type) Output Shape Param # \n=================================================================\n conv2d (Conv2D) (None, 26, 26, 32) 320 \n \n max_pooling2d (MaxPooling2D (None, 13, 13, 32) 0 \n ) \n \n conv2d_1 (Conv2D) (None, 11, 11, 64) 18496 \n \n max_pooling2d_1 (MaxPooling (None, 5, 5, 64) 0 \n 2D) \n \n conv2d_2 (Conv2D) (None, 3, 3, 64) 36928 \n \n flatten (Flatten) (None, 576) 0 \n \n dense_2 (Dense) (None, 64) 36928 \n \n dense_3 (Dense) (None, 10) 650 \n \n=================================================================\nTotal params: 93,322\nTrainable params: 93,322\nNon-trainable params: 0\n_________________________________________________________________\n" ], [ "train_images = train_images.reshape((60000, 28, 28, 1))\ntrain_images = train_images.astype('float32') / 255\n\ntest_images = test_images.reshape((10000, 28, 28, 1))\ntest_images = test_images.astype('float32') / 255", "_____no_output_____" ], [ "model.compile(\n optimizer='rmsprop',\n loss='categorical_crossentropy',\n metrics=['accuracy']\n)", "_____no_output_____" ], [ "train_images.shape", "_____no_output_____" ], [ "train_labels_hot", "_____no_output_____" ], [ "model.fit(train_images, train_labels_hot, epochs=10, batch_size=128, validation_split=0.1)", "Epoch 1/10\n" ], [ "", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad7003910f4815d71939be1e4ec1cb8cd3d1305

48,333

ipynb

Jupyter Notebook

examples/ex3_ula_performance.ipynb

jtang10/doatools.py

c9f2a8dddeb2ea9bf96b956f03a1eb57d10cb850

[ "MIT" ]

86

2018-09-16T05:35:32.000Z

2022-03-28T14:16:30.000Z

examples/ex3_ula_performance.ipynb

jtang10/doatools.py

c9f2a8dddeb2ea9bf96b956f03a1eb57d10cb850

[ "MIT" ]

1

2019-06-15T08:51:31.000Z

examples/ex3_ula_performance.ipynb

jtang10/doatools.py

c9f2a8dddeb2ea9bf96b956f03a1eb57d10cb850

[ "MIT" ]

33

2018-12-10T08:38:40.000Z

2022-03-29T19:45:46.000Z

232.370192

41,824

0.907123

[ [ [ "# Performance analysis of a uniform linear array\n\nWe compare the MSE of MUSIC with the CRB for a uniform linear array (ULA).", "_____no_output_____" ] ], [ [ "import numpy as np\nimport doatools.model as model\nimport doatools.estimation as estimation\nimport doatools.performance as perf\nimport matplotlib.pyplot as plt\n%matplotlib inline", "_____no_output_____" ], [ "wavelength = 1.0 # normalized\nd0 = wavelength / 2\n\n# Create a 12-element ULA.\nula = model.UniformLinearArray(12, d0)\n# Place 8 sources uniformly within (-pi/3, pi/4)\nsources = model.FarField1DSourcePlacement(\n np.linspace(-np.pi/3, np.pi/4, 8)\n)\n# All sources share the same power.\npower_source = 1 # Normalized\nsource_signal = model.ComplexStochasticSignal(sources.size, power_source)\n# 200 snapshots.\nn_snapshots = 200\n# We use root-MUSIC.\nestimator = estimation.RootMUSIC1D(wavelength)", "_____no_output_____" ] ], [ [ "We vary the SNR from -20 dB to 20 dB. Here the SNR is defined as:\n\\begin{equation}\n \\mathrm{SNR} = 10\\log_{10}\\frac{\\min_i p_i}{\\sigma^2_{\\mathrm{n}}},\n\\end{equation}\nwhere $p_i$ is the power of the $i$-th source, and $\\sigma^2_{\\mathrm{n}}$ is the noise power.", "_____no_output_____" ] ], [ [ "snrs = np.linspace(-20, 10, 20)\n# 300 Monte Carlo runs for each SNR\nn_repeats = 300\n\nmses = np.zeros((len(snrs),))\ncrbs_sto = np.zeros((len(snrs),))\ncrbs_det = np.zeros((len(snrs),))\ncrbs_stouc = np.zeros((len(snrs),))\n\nfor i, snr in enumerate(snrs):\n power_noise = power_source / (10**(snr / 10))\n noise_signal = model.ComplexStochasticSignal(ula.size, power_noise)\n # The squared errors and the deterministic CRB varies\n # for each run. We need to compute the average.\n cur_mse = 0.0\n cur_crb_det = 0.0\n for r in range(n_repeats):\n # Stochastic signal model.\n A = ula.steering_matrix(sources, wavelength)\n S = source_signal.emit(n_snapshots)\n N = noise_signal.emit(n_snapshots)\n Y = A @ S + N\n Rs = (S @ S.conj().T) / n_snapshots\n Ry = (Y @ Y.conj().T) / n_snapshots\n resolved, estimates = estimator.estimate(Ry, sources.size, d0)\n # In practice, you should check if `resolved` is true.\n # We skip the check here.\n cur_mse += np.mean((estimates.locations - sources.locations)**2)\n B_det = perf.ecov_music_1d(ula, sources, wavelength, Rs, power_noise,\n n_snapshots)\n cur_crb_det += np.mean(np.diag(B_det))\n # Update the results.\n B_sto = perf.crb_sto_farfield_1d(ula, sources, wavelength, power_source,\n power_noise, n_snapshots)\n B_stouc = perf.crb_stouc_farfield_1d(ula, sources, wavelength, power_source,\n power_noise, n_snapshots)\n mses[i] = cur_mse / n_repeats\n crbs_sto[i] = np.mean(np.diag(B_sto))\n crbs_det[i] = cur_crb_det / n_repeats\n crbs_stouc[i] = np.mean(np.diag(B_stouc))\n print('Completed SNR = {0:.2f} dB'.format(snr))", "Completed SNR = -20.00 dB\nCompleted SNR = -18.42 dB\nCompleted SNR = -16.84 dB\nCompleted SNR = -15.26 dB\nCompleted SNR = -13.68 dB\nCompleted SNR = -12.11 dB\nCompleted SNR = -10.53 dB\nCompleted SNR = -8.95 dB\nCompleted SNR = -7.37 dB\nCompleted SNR = -5.79 dB\nCompleted SNR = -4.21 dB\nCompleted SNR = -2.63 dB\nCompleted SNR = -1.05 dB\nCompleted SNR = 0.53 dB\nCompleted SNR = 2.11 dB\nCompleted SNR = 3.68 dB\nCompleted SNR = 5.26 dB\nCompleted SNR = 6.84 dB\nCompleted SNR = 8.42 dB\nCompleted SNR = 10.00 dB\n" ] ], [ [ "We plot the results below.\n\n* The MSE should approach the stochastic CRBs in high SNR regions.\n* The stochastic CRB should be tighter than the deterministic CRB.\n* With the additional assumption of uncorrelated sources, we expect a even lower CRB.\n* All three CRBs should converge together as the SNR approaches infinity.", "_____no_output_____" ] ], [ [ "plt.figure(figsize=(8, 6))\nplt.semilogy(\n snrs, mses, '-x',\n snrs, crbs_sto, '--',\n snrs, crbs_det, '--',\n snrs, crbs_stouc, '--'\n)\nplt.xlabel('SNR (dB)')\nplt.ylabel(r'MSE / $\\mathrm{rad}^2$')\nplt.grid(True)\nplt.legend(['MSE', 'Stochastic CRB', 'Deterministic CRB', \n 'Stochastic CRB (Uncorrelated)'])\nplt.title('MSE vs. CRB')\nplt.margins(x=0)\nplt.show()", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad70cf66c4c8291544af44cc5808099121603ed

16,080

ipynb

Jupyter Notebook

site/en/tutorials/customization/basics.ipynb

prasoonkottarathil/docs

5c4ef179c339b0a09377ffe1bdbdf1e0609d047c

[ "Apache-2.0" ]

1

2020-01-12T10:42:48.000Z

site/en/tutorials/customization/basics.ipynb

prasoonkottarathil/docs

5c4ef179c339b0a09377ffe1bdbdf1e0609d047c

[ "Apache-2.0" ]

null

site/en/tutorials/customization/basics.ipynb

prasoonkottarathil/docs

5c4ef179c339b0a09377ffe1bdbdf1e0609d047c

[ "Apache-2.0" ]

2

2020-01-15T21:50:31.000Z

2020-01-15T21:56:30.000Z

34.655172

627

0.543657

[ [ [ "##### Copyright 2018 The TensorFlow Authors.", "_____no_output_____" ] ], [ [ "#@title Licensed under the Apache License, Version 2.0 (the \"License\");\n# you may not use this file except in compliance with the License.\n# You may obtain a copy of the License at\n#\n# https://www.apache.org/licenses/LICENSE-2.0\n#\n# Unless required by applicable law or agreed to in writing, software\n# distributed under the License is distributed on an \"AS IS\" BASIS,\n# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.\n# See the License for the specific language governing permissions and\n# limitations under the License.", "_____no_output_____" ] ], [ [ "# Customization basics: tensors and operations", "_____no_output_____" ], [ "<table class=\"tfo-notebook-buttons\" align=\"left\">\n <td>\n <a target=\"_blank\" href=\"https://www.tensorflow.org/tutorials/customization/basics\"><img src=\"https://www.tensorflow.org/images/tf_logo_32px.png\" />View on TensorFlow.org</a>\n </td>\n <td>\n <a target=\"_blank\" href=\"https://colab.research.google.com/github/tensorflow/docs/blob/master/site/en/tutorials/customization/basics.ipynb\"><img src=\"https://www.tensorflow.org/images/colab_logo_32px.png\" />Run in Google Colab</a>\n </td>\n <td>\n <a target=\"_blank\" href=\"https://github.com/tensorflow/docs/blob/master/site/en/tutorials/customization/basics.ipynb\"><img src=\"https://www.tensorflow.org/images/GitHub-Mark-32px.png\" />View source on GitHub</a>\n </td>\n <td>\n <a href=\"https://storage.googleapis.com/tensorflow_docs/docs/site/en/tutorials/customization/basics.ipynb\"><img src=\"https://www.tensorflow.org/images/download_logo_32px.png\" />Download notebook</a>\n </td>\n</table>", "_____no_output_____" ], [ "This is an introductory TensorFlow tutorial that shows how to:\n\n* Import the required package\n* Create and use tensors\n* Use GPU acceleration\n* Demonstrate `tf.data.Dataset`", "_____no_output_____" ] ], [ [ "from __future__ import absolute_import, division, print_function, unicode_literals\n\ntry:\n # %tensorflow_version only exists in Colab.\n %tensorflow_version 2.x\nexcept Exception:\n pass\n", "_____no_output_____" ] ], [ [ "## Import TensorFlow\n\nTo get started, import the `tensorflow` module. As of TensorFlow 2, eager execution is turned on by default. This enables a more interactive frontend to TensorFlow, the details of which we will discuss much later.", "_____no_output_____" ] ], [ [ "import tensorflow as tf", "_____no_output_____" ] ], [ [ "## Tensors\n\nA Tensor is a multi-dimensional array. Similar to NumPy `ndarray` objects, `tf.Tensor` objects have a data type and a shape. Additionally, `tf.Tensor`s can reside in accelerator memory (like a GPU). TensorFlow offers a rich library of operations ([tf.add](https://www.tensorflow.org/api_docs/python/tf/add), [tf.matmul](https://www.tensorflow.org/api_docs/python/tf/matmul), [tf.linalg.inv](https://www.tensorflow.org/api_docs/python/tf/linalg/inv) etc.) that consume and produce `tf.Tensor`s. These operations automatically convert native Python types, for example:\n", "_____no_output_____" ] ], [ [ "print(tf.add(1, 2))\nprint(tf.add([1, 2], [3, 4]))\nprint(tf.square(5))\nprint(tf.reduce_sum([1, 2, 3]))\n\n# Operator overloading is also supported\nprint(tf.square(2) + tf.square(3))", "_____no_output_____" ] ], [ [ "Each `tf.Tensor` has a shape and a datatype:", "_____no_output_____" ] ], [ [ "x = tf.matmul([[1]], [[2, 3]])\nprint(x)\nprint(x.shape)\nprint(x.dtype)", "_____no_output_____" ] ], [ [ "The most obvious differences between NumPy arrays and `tf.Tensor`s are:\n\n1. Tensors can be backed by accelerator memory (like GPU, TPU).\n2. Tensors are immutable.", "_____no_output_____" ], [ "### NumPy Compatibility\n\nConverting between a TensorFlow `tf.Tensor`s and a NumPy `ndarray` is easy:\n\n* TensorFlow operations automatically convert NumPy ndarrays to Tensors.\n* NumPy operations automatically convert Tensors to NumPy ndarrays.\n\nTensors are explicitly converted to NumPy ndarrays using their `.numpy()` method. These conversions are typically cheap since the array and `tf.Tensor` share the underlying memory representation, if possible. However, sharing the underlying representation isn't always possible since the `tf.Tensor` may be hosted in GPU memory while NumPy arrays are always backed by host memory, and the conversion involves a copy from GPU to host memory.", "_____no_output_____" ] ], [ [ "import numpy as np\n\nndarray = np.ones([3, 3])\n\nprint(\"TensorFlow operations convert numpy arrays to Tensors automatically\")\ntensor = tf.multiply(ndarray, 42)\nprint(tensor)\n\n\nprint(\"And NumPy operations convert Tensors to numpy arrays automatically\")\nprint(np.add(tensor, 1))\n\nprint(\"The .numpy() method explicitly converts a Tensor to a numpy array\")\nprint(tensor.numpy())", "_____no_output_____" ] ], [ [ "## GPU acceleration\n\nMany TensorFlow operations are accelerated using the GPU for computation. Without any annotations, TensorFlow automatically decides whether to use the GPU or CPU for an operation—copying the tensor between CPU and GPU memory, if necessary. Tensors produced by an operation are typically backed by the memory of the device on which the operation executed, for example:", "_____no_output_____" ] ], [ [ "x = tf.random.uniform([3, 3])\n\nprint(\"Is there a GPU available: \"),\nprint(tf.config.experimental.list_physical_devices(\"GPU\"))\n\nprint(\"Is the Tensor on GPU #0: \"),\nprint(x.device.endswith('GPU:0'))", "_____no_output_____" ] ], [ [ "### Device Names\n\nThe `Tensor.device` property provides a fully qualified string name of the device hosting the contents of the tensor. This name encodes many details, such as an identifier of the network address of the host on which this program is executing and the device within that host. This is required for distributed execution of a TensorFlow program. The string ends with `GPU:<N>` if the tensor is placed on the `N`-th GPU on the host.", "_____no_output_____" ], [ "\n\n### Explicit Device Placement\n\nIn TensorFlow, *placement* refers to how individual operations are assigned (placed on) a device for execution. As mentioned, when there is no explicit guidance provided, TensorFlow automatically decides which device to execute an operation and copies tensors to that device, if needed. However, TensorFlow operations can be explicitly placed on specific devices using the `tf.device` context manager, for example:", "_____no_output_____" ] ], [ [ "import time\n\ndef time_matmul(x):\n start = time.time()\n for loop in range(10):\n tf.matmul(x, x)\n\n result = time.time()-start\n\n print(\"10 loops: {:0.2f}ms\".format(1000*result))\n\n# Force execution on CPU\nprint(\"On CPU:\")\nwith tf.device(\"CPU:0\"):\n x = tf.random.uniform([1000, 1000])\n assert x.device.endswith(\"CPU:0\")\n time_matmul(x)\n\n# Force execution on GPU #0 if available\nif tf.config.experimental.list_physical_devices(\"GPU\"):\n print(\"On GPU:\")\n with tf.device(\"GPU:0\"): # Or GPU:1 for the 2nd GPU, GPU:2 for the 3rd etc.\n x = tf.random.uniform([1000, 1000])\n assert x.device.endswith(\"GPU:0\")\n time_matmul(x)", "_____no_output_____" ] ], [ [ "## Datasets\n\nThis section uses the [`tf.data.Dataset` API](https://www.tensorflow.org/guide/datasets) to build a pipeline for feeding data to your model. The `tf.data.Dataset` API is used to build performant, complex input pipelines from simple, re-usable pieces that will feed your model's training or evaluation loops.", "_____no_output_____" ], [ "### Create a source `Dataset`\n\nCreate a *source* dataset using one of the factory functions like [`Dataset.from_tensors`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#from_tensors), [`Dataset.from_tensor_slices`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#from_tensor_slices), or using objects that read from files like [`TextLineDataset`](https://www.tensorflow.org/api_docs/python/tf/data/TextLineDataset) or [`TFRecordDataset`](https://www.tensorflow.org/api_docs/python/tf/data/TFRecordDataset). See the [TensorFlow Dataset guide](https://www.tensorflow.org/guide/datasets#reading_input_data) for more information.", "_____no_output_____" ] ], [ [ "ds_tensors = tf.data.Dataset.from_tensor_slices([1, 2, 3, 4, 5, 6])\n\n# Create a CSV file\nimport tempfile\n_, filename = tempfile.mkstemp()\n\nwith open(filename, 'w') as f:\n f.write(\"\"\"Line 1\nLine 2\nLine 3\n \"\"\")\n\nds_file = tf.data.TextLineDataset(filename)", "_____no_output_____" ] ], [ [ "### Apply transformations\n\nUse the transformations functions like [`map`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#map), [`batch`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#batch), and [`shuffle`](https://www.tensorflow.org/api_docs/python/tf/data/Dataset#shuffle) to apply transformations to dataset records.", "_____no_output_____" ] ], [ [ "ds_tensors = ds_tensors.map(tf.square).shuffle(2).batch(2)\n\nds_file = ds_file.batch(2)", "_____no_output_____" ] ], [ [ "### Iterate\n\n`tf.data.Dataset` objects support iteration to loop over records:", "_____no_output_____" ] ], [ [ "print('Elements of ds_tensors:')\nfor x in ds_tensors:\n print(x)\n\nprint('\\nElements in ds_file:')\nfor x in ds_file:\n print(x)", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad71f3ae394a70cf8ae7b12fec97524228eea73

63,642

ipynb

Jupyter Notebook

module1-afirstlookatdata/Kole_Goldsberry_LS_DSPT3_111_A_First_Look_at_Data.ipynb

Tenntucky/DS-Unit-1-Sprint-1-Dealing-With-Data

9274d813881175c68bc09ad3f18680dfc0d6186e

[ "MIT" ]

null

module1-afirstlookatdata/Kole_Goldsberry_LS_DSPT3_111_A_First_Look_at_Data.ipynb

Tenntucky/DS-Unit-1-Sprint-1-Dealing-With-Data

9274d813881175c68bc09ad3f18680dfc0d6186e

[ "MIT" ]

null

module1-afirstlookatdata/Kole_Goldsberry_LS_DSPT3_111_A_First_Look_at_Data.ipynb

Tenntucky/DS-Unit-1-Sprint-1-Dealing-With-Data

9274d813881175c68bc09ad3f18680dfc0d6186e

[ "MIT" ]

null

40.71785

11,968

0.531017

[ [ [ "<a href=\"https://colab.research.google.com/github/Tenntucky/DS-Unit-1-Sprint-1-Dealing-With-Data/blob/master/module1-afirstlookatdata/Kole_Goldsberry_LS_DSPT3_111_A_First_Look_at_Data.ipynb\" target=\"_parent\"><img src=\"https://colab.research.google.com/assets/colab-badge.svg\" alt=\"Open In Colab\"/></a>", "_____no_output_____" ], [ "# Lambda School Data Science - A First Look at Data\n\n", "_____no_output_____" ], [ "## Lecture - let's explore Python DS libraries and examples!\n\nThe Python Data Science ecosystem is huge. You've seen some of the big pieces - pandas, scikit-learn, matplotlib. What parts do you want to see more of?", "_____no_output_____" ] ], [ [ "2 + 2", "_____no_output_____" ], [ "def helloworld():\n return print('Hello World!')", "_____no_output_____" ], [ "helloworld()", "Hello World!\n" ], [ "import pandas as pd\n\ndf = pd.DataFrame({'a': [1, 2, 3, 4, 5], \n 'b': [5, 4, 3, 2, 1]})\ndf.head()", "_____no_output_____" ], [ "df.plot.scatter('a', 'b');", "_____no_output_____" ], [ "joe = {'name': 'Joe', 'is_female': False, 'age': 19}\nalice = {'name': 'Alice', 'is_female': True, 'age': 20}\nsarah = {'name': 'Sarah', 'is_female': True, 'age': 20}\nstudents = [joe, alice, sarah]", "_____no_output_____" ], [ "import numpy as np", "_____no_output_____" ], [ "np.random.randint(0, 10, size=10)", "_____no_output_____" ], [ "import matplotlib.pyplot as plt", "_____no_output_____" ], [ "x = [1, 2, 3, 4]\ny = [2, 4, 6, 10]\nprint(x, y)", "[1, 2, 3, 4] [2, 4, 6, 10]\n" ], [ "plt.scatter(x, y, color='r');\nplt.plot(x, y, color='g');", "_____no_output_____" ], [ "df = pd.DataFrame({'first_col': x, 'second_col': y})\ndf", "_____no_output_____" ], [ "df['first_col']", "_____no_output_____" ], [ "df['second_col']", "_____no_output_____" ], [ "df.shape", "_____no_output_____" ], [ "df['third_col'] = df['first_col'] + 2*df['second_col']", "_____no_output_____" ], [ "df", "_____no_output_____" ], [ "df.shape", "_____no_output_____" ], [ "arr_1 = np.random.randint(low=0, high=100, size=10000)\narr_2 = np.random.randint(low=0, high=100, size=10000)", "_____no_output_____" ], [ "arr_1.shape", "_____no_output_____" ], [ "arr_2.shape", "_____no_output_____" ], [ "arr_1 + arr_2", "_____no_output_____" ], [ "x + y", "_____no_output_____" ], [ "type(arr_1)\ntype(arr_2)", "_____no_output_____" ], [ "type(x)", "_____no_output_____" ], [ "df", "_____no_output_____" ], [ "df['fourth_col'] = df['third_col'] > 10", "_____no_output_____" ], [ "df", "_____no_output_____" ], [ "df.shape", "_____no_output_____" ], [ "df[df['second_col'] < 10]", "_____no_output_____" ], [ "print(students)", "[{'name': 'Joe', 'is_female': False, 'age': 19}, {'name': 'Alice', 'is_female': True, 'age': 20}, {'name': 'Sarah', 'is_female': True, 'age': 20}]\n" ], [ "df_1 = pd.DataFrame(students)", "_____no_output_____" ], [ "df_1", "_____no_output_____" ], [ "df_1['legal_drinker'] = df_1['age'] > 21", "_____no_output_____" ], [ "df_1", "_____no_output_____" ], [ "df_1[df_1['name'] == 'Alice']", "_____no_output_____" ] ], [ [ "## Assignment - now it's your turn\n\nPick at least one Python DS library, and using documentation/examples reproduce in this notebook something cool. It's OK if you don't fully understand it or get it 100% working, but do put in effort and look things up.", "_____no_output_____" ] ], [ [ "# TODO - your code here\n# Use what we did live in lecture as an example\n\n# 1.\n# Above I recreated what Alex Kim did yesterday in lecture. I decided to look\n# at the day one notes where they create a dictionary of three peoples names. \n# I decided to add some columns with legal drinking status (worked) and one with\n# the eligibility of said person on playing for the United States Womens National \n# Soccer Team (didn't work). I couldn't get it to read a boolean (pass/fail) \n# so I would like to figure that out. \n\n# 2.\n# I believe the boolean aspect will be the hardest. Also, changing some of the \n# data types is going to be challenging.\n\n# 3. \n# I'm really hoping that about halfway through the NFL season I will be able\n# to use this to predict where the teams currently sit then. So, appending a \n# data frame based off of previous data.\n\n# 4.\n# I would like to continue exploring dataframes. The idea of a large subset of \n# values that can be quickly explored with key words then used to give a \n# summation of that data. I would like to be able to predict possible values of \n# players versus teams in order to offer insight into personel decisions.", "_____no_output_____" ] ], [ [ "### Assignment questions\n\nAfter you've worked on some code, answer the following questions in this text block:\n\n1. Describe in a paragraph of text what you did and why, as if you were writing an email to somebody interested but nontechnical.\n\n2. What was the most challenging part of what you did?\n\n3. What was the most interesting thing you learned?\n\n4. What area would you like to explore with more time?\n\n\n", "_____no_output_____" ], [ "## Stretch goals and resources\n\nFollowing are *optional* things for you to take a look at. Focus on the above assignment first, and make sure to commit and push your changes to GitHub (and since this is the first assignment of the sprint, open a PR as well).\n\n- [pandas documentation](https://pandas.pydata.org/pandas-docs/stable/)\n- [scikit-learn documentation](http://scikit-learn.org/stable/documentation.html)\n- [matplotlib documentation](https://matplotlib.org/contents.html)\n- [Awesome Data Science](https://github.com/bulutyazilim/awesome-datascience) - a list of many types of DS resources\n\nStretch goals:\n\n- Find and read blogs, walkthroughs, and other examples of people working through cool things with data science - and share with your classmates!\n- Write a blog post (Medium is a popular place to publish) introducing yourself as somebody learning data science, and talking about what you've learned already and what you're excited to learn more about.", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown", "markdown", "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ] ]

4ad73b9d878540a22fd955139639a9b11e7f3cb1

184,357

ipynb

Jupyter Notebook

Gender Recognition by Voice Project.ipynb

pth051001/Gender-Recognition-by-Voice-with-different-ML-classification-methods

5158f951d8df52eeff229c80555194d50aeca41e

[ "MIT" ]

null

Gender Recognition by Voice Project.ipynb

pth051001/Gender-Recognition-by-Voice-with-different-ML-classification-methods

5158f951d8df52eeff229c80555194d50aeca41e

[ "MIT" ]

null

Gender Recognition by Voice Project.ipynb

pth051001/Gender-Recognition-by-Voice-with-different-ML-classification-methods

5158f951d8df52eeff229c80555194d50aeca41e

[ "MIT" ]

null

73.683853

34,420

0.718676

[ [ [ "## Gender Recognition by Voice Project\nIn this project, we will classify a person's gender by his/her various aspects of voice using different classification methods like logistic regression, k-nearest neighbors, Naive Bayes. These methods will be completely implemented from sratch using pure Python and related mathematical concepts. For each method, we'll compare it with its built-in version in sklearn library to see if there are any differences in results. In addition, other methods like SVM, Decision Tree, Random Forest are also used from sklearn to compare the accuracy among methods. Data were downloaded from Kaggle", "_____no_output_____" ], [ "## Imports\n", "_____no_output_____" ] ], [ [ "import pandas as pd\nimport seaborn as sns\nimport matplotlib.pyplot as plt\nimport numpy as np\n%matplotlib inline", "_____no_output_____" ] ], [ [ "## Read and explore data", "_____no_output_____" ] ], [ [ "df = pd.read_csv(\"voice.csv\")", "_____no_output_____" ], [ "df.head()", "_____no_output_____" ], [ "df.describe()", "_____no_output_____" ], [ "sns.countplot(x='label', data=df)", "_____no_output_____" ], [ "sns.heatmap(df.drop('label', axis=1).corr(), square = True, cmap=\"YlGnBu\", linecolor='black')", "_____no_output_____" ], [ "sns.FacetGrid(df, hue='label', size=5).map(plt.scatter, 'meandom','meanfun').add_legend()", "C:\\Users\\phant\\anaconda3\\lib\\site-packages\\seaborn\\axisgrid.py:243: UserWarning: The `size` parameter has been renamed to `height`; please update your code.\n warnings.warn(msg, UserWarning)\n" ] ], [ [ "## Standardize data\nData need to be standardized to a smaller scale to calculate more easily.", "_____no_output_____" ] ], [ [ "from sklearn.preprocessing import StandardScaler", "_____no_output_____" ], [ "scaler = StandardScaler()", "_____no_output_____" ], [ "scaler.fit(df.drop('label',axis=1))", "_____no_output_____" ], [ "scaled_features = scaler.transform(df.drop('label',axis=1))", "_____no_output_____" ], [ "# feat all columns except the \"Type\" one\ndf_feat = pd.DataFrame(scaled_features,columns=df.columns[:-1])\ndf_feat.head()", "_____no_output_____" ] ], [ [ "## Split the data", "_____no_output_____" ] ], [ [ "from sklearn.model_selection import train_test_split", "_____no_output_____" ], [ "# encoding label column\ndf['label'] = df['label'].replace(['male', 'female'], [0,1])", "_____no_output_____" ], [ "X = df_feat\ny = df['label']\nX_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.33, random_state=42)\n\n# df to compare results among methods\ncomparison = pd.DataFrame(columns=['Name', 'accuracy'])", "_____no_output_____" ] ], [ [ "## Use Logistic Regression from sklearn library", "_____no_output_____" ] ], [ [ "from sklearn.linear_model import LogisticRegression", "_____no_output_____" ], [ "# Train the model\nlogmodel = LogisticRegression()\nlogmodel.fit(X_train,y_train)", "_____no_output_____" ], [ "# Get the prediction for X_test\ny_logSK_pred = logmodel.predict(X_test)", "_____no_output_____" ], [ "from sklearn.metrics import confusion_matrix, classification_report, accuracy_score", "_____no_output_____" ], [ "print(classification_report(y_test,y_logSK_pred))", " precision recall f1-score support\n\n 0 0.97 0.98 0.97 547\n 1 0.97 0.97 0.97 499\n\n accuracy 0.97 1046\n macro avg 0.97 0.97 0.97 1046\nweighted avg 0.97 0.97 0.97 1046\n\n" ], [ "print(confusion_matrix(y_test,y_logSK_pred))", "[[534 13]\n [ 17 482]]\n" ] ], [ [ "## Implementing Logistic Regression from scratch", "_____no_output_____" ] ], [ [ "# Create a logistic regression class to import from (LogisticRegressionFromScratch.py)\n\nclass LogisticRegressionFromScratch:\n def __init__(self, X, y):\n # Add a column of zeros for X\n zeros_col = np.ones((X.shape[0],1))\n X = np.append(zeros_col,X,axis=1)\n # Initialize variables\n self.X = X\n self.y = y\n self.m = X.shape[0]\n self.n = X.shape[1]\n # Randomize values for theta\n self.theta = np.random.randn(X.shape[1],1)\n \n def sigmoid(self, z):\n return 1/(1 + np.exp(-z))\n \n def costFunction(self):\n # Calculate predicted h then cost value\n h = self.sigmoid(np.matmul(self.X, self.theta))\n self.J = (1/self.m)*(-self.y.T.dot(np.log(h)) - (1 - self.y).T.dot(np.log(1 - h)))\n return self.J\n \n def gradientDescent(self, alpha, num_iters):\n # Keep records of cost values and thetas\n self.J_history = []\n self.theta_history = []\n for i in range (num_iters):\n # Calculate new value for h then update J_history\n h = self.sigmoid(np.matmul(self.X, self.theta))\n self.J_history.append(self.costFunction())\n self.theta_history.append(self.theta)\n self.theta = self.theta - (alpha/self.m)*(self.X.T.dot(h-self.y))\n return self.J_history, self.theta_history, self.theta\n \n def predict(self, X_test, y_test):\n # Add a column of zeros for X_test\n zeros_col = np.ones((X_test.shape[0],1))\n X_test = np.append(zeros_col, X_test, axis = 1)\n # Calculate final predicted y values after using gradient descent to update theta\n cal_sigmoid = self.sigmoid(np.matmul(X_test, self.theta))\n self.y_pred = []\n for value in cal_sigmoid:\n if value >= 0.5:\n self.y_pred.append(1)\n else:\n self.y_pred.append(0) \n return self.y_pred ", "_____no_output_____" ], [ "from LogisticRegressionFromScratch import LogisticRegressionFromScratch", "_____no_output_____" ], [ "lmFromScratch = LogisticRegressionFromScratch(X_train, y_train.to_numpy().reshape(y_train.shape[0],1))\n# PREDICT USING GRADIENT DESCENT\n# set up number of iterations and learning rate\nnum_iters = 15000\nalpha = 0.01 \n# update theta value and get predicted y\nj_hist, theta_hist, theta = lmFromScratch.gradientDescent(alpha, num_iters)\ny_logScratch_pred = lmFromScratch.predict(X_test, y_test.to_numpy().reshape(y_test.shape[0],1))", "_____no_output_____" ], [ "print(confusion_matrix(y_test,y_logScratch_pred))", "[[535 12]\n [ 19 480]]\n" ], [ "print(classification_report(y_test,y_logScratch_pred))", " precision recall f1-score support\n\n 0 0.97 0.98 0.97 547\n 1 0.98 0.96 0.97 499\n\n accuracy 0.97 1046\n macro avg 0.97 0.97 0.97 1046\nweighted avg 0.97 0.97 0.97 1046\n\n" ], [ "new_data = {'Name': 'Logistic Regression', 'accuracy': accuracy_score(y_test,y_logScratch_pred)}\ncomparison = comparison.append(new_data, ignore_index=True)", "_____no_output_____" ] ], [ [ "## Use KNN from sklearn library", "_____no_output_____" ] ], [ [ "from sklearn.neighbors import KNeighborsClassifier", "_____no_output_____" ], [ "# start with k = 1\nknn = KNeighborsClassifier(n_neighbors=1)\nknn.fit(X_train, y_train)\ny_knnSK_pred = knn.predict(X_test)", "_____no_output_____" ], [ "print(confusion_matrix(y_test, y_knnSK_pred))", "[[532 15]\n [ 11 488]]\n" ], [ "print(classification_report(y_test,y_knnSK_pred))", " precision recall f1-score support\n\n 0 0.98 0.97 0.98 547\n 1 0.97 0.98 0.97 499\n\n accuracy 0.98 1046\n macro avg 0.97 0.98 0.98 1046\nweighted avg 0.98 0.98 0.98 1046\n\n" ], [ "# plot out the error vs k-value graph to choose the best k value\nerror_rate = []\n\nfor i in range(1,40):\n knn = KNeighborsClassifier(n_neighbors=i)\n knn.fit(X_train, y_train)\n pred_i = knn.predict(X_test)\n error_rate.append(np.mean(pred_i != y_test))", "_____no_output_____" ], [ "plt.figure(figsize = (10,6))\nplt.plot(range(1,40), error_rate, color = 'blue', linestyle = '--', marker = 'o', markerfacecolor = 'red', markersize = 10)", "_____no_output_____" ], [ "# From above plot, k = 3 is the value gives us the lowest error, so we'll retrain knn model with k = 3\nknn = KNeighborsClassifier(n_neighbors=3)\nknn.fit(X_train, y_train)\npred = knn.predict(X_test)", "_____no_output_____" ], [ "print(confusion_matrix(y_test, y_knnSK_pred))", "[[532 15]\n [ 11 488]]\n" ], [ "print(classification_report(y_test,y_knnSK_pred))", " precision recall f1-score support\n\n 0 0.98 0.97 0.98 547\n 1 0.97 0.98 0.97 499\n\n accuracy 0.98 1046\n macro avg 0.97 0.98 0.98 1046\nweighted avg 0.98 0.98 0.98 1046\n\n" ] ], [ [ "There are no difference in results when changing k to 3 because the errors when k = 1 and k = 3 is too small and inevitable.", "_____no_output_____" ], [ "## Implementing KNN from scratch", "_____no_output_____" ] ], [ [ "import numpy as np\n\nclass KNNFromScratch():\n def __init__(self, k):\n self.k = k\n \n # get training data\n def train(self, X, y):\n self.X_train = X\n self.y_train = y\n \n def predict(self, X_test):\n dist = self.compute_dist(X_test) \n return self.predict_label(dist)\n \n # compute distance between each sample in X_test and X_train\n def compute_dist(self, X_test):\n test_size = X_test.shape[0]\n train_size = self.X_train.shape[0]\n dist = np.zeros((test_size, train_size))\n for i in range(test_size):\n for j in range(train_size):\n dist[i, j] = np.sqrt(np.sum((X_test[i,:] - self.X_train[j,:])**2))\n return dist\n \n # return predicted label with given distance of X_test\n def predict_label(self, dist):\n test_size = dist.shape[0]\n y_pred = np.zeros(test_size)\n for i in range(test_size):\n y_indices = np.argsort(dist[i, :])\n k_closest = self.y_train[y_indices[: self.k]].astype(int)\n y_pred[i] = np.argmax(np.bincount(k_closest))\n return y_pred", "_____no_output_____" ], [ "from KNNFromScratch import KNNFromScratch", "_____no_output_____" ], [ "# train with k=3\nknnFromScratch = KNNFromScratch(3)\nknnFromScratch.train(X_train.to_numpy(), y_train.to_numpy())\ny_knnScratch_pred = knnFromScratch.predict(X_test.to_numpy())", "_____no_output_____" ] ], [ [ "The result is slightly better than using sklearn library", "_____no_output_____" ] ], [ [ "print(confusion_matrix(y_test, y_knnScratch_pred))", "[[536 11]\n [ 12 487]]\n" ], [ "print(classification_report(y_test,y_knnScratch_pred))", " precision recall f1-score support\n\n 0 0.98 0.98 0.98 547\n 1 0.98 0.98 0.98 499\n\n accuracy 0.98 1046\n macro avg 0.98 0.98 0.98 1046\nweighted avg 0.98 0.98 0.98 1046\n\n" ], [ "new_data = {'Name': 'KNN', 'accuracy': accuracy_score(y_test,y_knnScratch_pred)}\ncomparison = comparison.append(new_data, ignore_index=True)", "_____no_output_____" ] ], [ [ "## Use Naive Bayes from sklearn library", "_____no_output_____" ] ], [ [ "from sklearn.naive_bayes import GaussianNB", "_____no_output_____" ], [ "naiveBayes = GaussianNB()\nnaiveBayes.fit(X_train, y_train)\ny_nbSK_pred = naiveBayes.predict(X_test)", "_____no_output_____" ], [ "print(confusion_matrix(y_test, y_nbSK_pred))", "[[489 58]\n [ 51 448]]\n" ], [ "print(classification_report(y_test,y_nbSK_pred))", " precision recall f1-score support\n\n 0 0.91 0.89 0.90 547\n 1 0.89 0.90 0.89 499\n\n accuracy 0.90 1046\n macro avg 0.90 0.90 0.90 1046\nweighted avg 0.90 0.90 0.90 1046\n\n" ] ], [ [ "## Implementing Naive Bayes from scratch", "_____no_output_____" ] ], [ [ "import numpy as np\nimport matplotlib.pyplot as plt\nfrom scipy.stats import multivariate_normal\n\nclass NaiveBayesFromScratch():\n def __init__(self, X, y):\n self.num_examples, self.num_features = X.shape\n self.num_classes = len(np.unique(y))\n\n def fit(self, X, y):\n self.classes_mean = {}\n self.classes_variance = {}\n self.classes_prior = {}\n \n # calculate the mean, variance, prior of each class\n for c in range(self.num_classes):\n X_c = X[y == c]\n\n self.classes_mean[str(c)] = np.mean(X_c, axis=0)\n self.classes_variance[str(c)] = np.var(X_c, axis=0)\n self.classes_prior[str(c)] = X_c.shape[0] / X.shape[0]\n \n # predict using Naive Bayes Gaussian formula\n def predict(self, X):\n probs = np.zeros((X.shape[0], self.num_classes))\n for c in range(self.num_classes):\n prior = self.classes_prior[str(c)]\n probs_c = multivariate_normal.pdf(X, mean=self.classes_mean[str(c)], cov=self.classes_variance[str(c)])\n probs[:,c] = probs_c*prior\n return np.argmax(probs, 1)\n ", "_____no_output_____" ], [ "from NaiveBayesFromScratch import NaiveBayesFromScratch", "_____no_output_____" ], [ "naiveBayesFromScratch = NaiveBayesFromScratch(X_train.to_numpy(), y_train.to_numpy())\nnaiveBayesFromScratch.fit(X_train.to_numpy(), y_train.to_numpy())\ny_nbScratch_pred = naiveBayesFromScratch.predict(X_test.to_numpy())", "_____no_output_____" ], [ "print(confusion_matrix(y_test, y_nbScratch_pred))", "[[489 58]\n [ 51 448]]\n" ], [ "print(classification_report(y_test,y_nbScratch_pred))", " precision recall f1-score support\n\n 0 0.91 0.89 0.90 547\n 1 0.89 0.90 0.89 499\n\n accuracy 0.90 1046\n macro avg 0.90 0.90 0.90 1046\nweighted avg 0.90 0.90 0.90 1046\n\n" ], [ "new_data = {'Name': 'Naive Bayes', 'accuracy': accuracy_score(y_test,y_nbScratch_pred)}\ncomparison = comparison.append(new_data, ignore_index=True)", "_____no_output_____" ] ], [ [ "## Use SVM from sklearn", "_____no_output_____" ] ], [ [ "from sklearn.svm import SVC\nsvm = SVC()\nsvm.fit(X_train,y_train)\nsvm_pred = svm.predict(X_test)", "_____no_output_____" ], [ "print(confusion_matrix(y_test,svm_pred))", "[[535 12]\n [ 8 491]]\n" ], [ "print(classification_report(y_test,svm_pred))", " precision recall f1-score support\n\n 0 0.99 0.98 0.98 547\n 1 0.98 0.98 0.98 499\n\n accuracy 0.98 1046\n macro avg 0.98 0.98 0.98 1046\nweighted avg 0.98 0.98 0.98 1046\n\n" ], [ "new_data = {'Name': 'SVM', 'accuracy': accuracy_score(y_test,svm_pred)}\ncomparison = comparison.append(new_data, ignore_index=True)", "_____no_output_____" ], [ "# Grid Search\nparam_grid = {'C': [0.1,1, 10, 100, 1000], 'gamma': [1,0.1,0.01,0.001,0.0001], 'kernel': ['rbf']} ", "_____no_output_____" ], [ "from sklearn.model_selection import GridSearchCV", "_____no_output_____" ], [ "grid = GridSearchCV(SVC(),param_grid,refit=True,verbose=3)\ngrid.fit(X_train,y_train)", "Fitting 5 folds for each of 25 candidates, totalling 125 fits\n[CV] C=0.1, gamma=1, kernel=rbf ......................................\n[CV] .......... C=0.1, gamma=1, kernel=rbf, score=0.694, total= 0.1s\n[CV] C=0.1, gamma=1, kernel=rbf ......................................\n" ], [ "grid.best_params_", "_____no_output_____" ], [ "grid_pred = grid.predict(X_test)", "_____no_output_____" ] ], [ [ "The result is not better than the default one", "_____no_output_____" ] ], [ [ "print(confusion_matrix(y_test,grid_pred))", "[[533 14]\n [ 9 490]]\n" ], [ "print(classification_report(y_test,grid_pred))", " precision recall f1-score support\n\n 0 0.98 0.97 0.98 547\n 1 0.97 0.98 0.98 499\n\n accuracy 0.98 1046\n macro avg 0.98 0.98 0.98 1046\nweighted avg 0.98 0.98 0.98 1046\n\n" ] ], [ [ "## Use Decision Tree from sklearn", "_____no_output_____" ] ], [ [ "from sklearn.tree import DecisionTreeClassifier", "_____no_output_____" ], [ "# train the model and get predicted results for test set\ndtree = DecisionTreeClassifier()\ndtree.fit(X_train,y_train)\ndtre_pred = dtree.predict(X_test)", "_____no_output_____" ], [ "print(classification_report(y_test,dtre_pred))", " precision recall f1-score support\n\n 0 0.97 0.96 0.97 547\n 1 0.95 0.97 0.96 499\n\n accuracy 0.96 1046\n macro avg 0.96 0.96 0.96 1046\nweighted avg 0.96 0.96 0.96 1046\n\n" ], [ "print(confusion_matrix(y_test,dtre_pred))", "[[524 23]\n [ 15 484]]\n" ], [ "new_data = {'Name': 'Decision Tree', 'accuracy': accuracy_score(y_test,dtre_pred)}\ncomparison = comparison.append(new_data, ignore_index=True)", "_____no_output_____" ] ], [ [ "## Use Random Forest from sklearn", "_____no_output_____" ] ], [ [ "from sklearn.ensemble import RandomForestClassifier", "_____no_output_____" ], [ "# train the model and get predicted results for test set\nrfc = RandomForestClassifier()\nrfc.fit(X_train, y_train)\nrfc_pred = rfc.predict(X_test)", "_____no_output_____" ], [ "print(classification_report(y_test,rfc_pred))", " precision recall f1-score support\n\n 0 0.99 0.97 0.98 547\n 1 0.97 0.98 0.98 499\n\n accuracy 0.98 1046\n macro avg 0.98 0.98 0.98 1046\nweighted avg 0.98 0.98 0.98 1046\n\n" ], [ "print(confusion_matrix(y_test,rfc_pred))", "[[532 15]\n [ 8 491]]\n" ], [ "new_data = {'Name': 'Random Forest', 'accuracy': accuracy_score(y_test,rfc_pred)}\ncomparison = comparison.append(new_data, ignore_index=True)", "_____no_output_____" ] ], [ [ "## Results comparison among methods", "_____no_output_____" ] ], [ [ "comparison", "_____no_output_____" ], [ "sns.barplot(data=comparison, x='accuracy', y='Name')", "_____no_output_____" ] ], [ [ "## Conclusion", "_____no_output_____" ], [ "In conclusion, results of methods implemented from scratch are quite similar to those from sklearn library. Among classification methods, SVM is the best method with highest accuracy.", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ] ]

4ad77442966356d65522427d48708689d053897b

9,323

ipynb

Jupyter Notebook

01_Introduction.ipynb

mue94/oddstradamus

29a97b933652bdeef21d9c9dcce06bf87ba4f1f0

[ "MIT" ]

null

01_Introduction.ipynb

mue94/oddstradamus

29a97b933652bdeef21d9c9dcce06bf87ba4f1f0

[ "MIT" ]

null

01_Introduction.ipynb

mue94/oddstradamus

29a97b933652bdeef21d9c9dcce06bf87ba4f1f0

[ "MIT" ]

null

34.150183

654

0.58951

[ [ [ "# Oddstradamus\n### Good odds and where to find them", "_____no_output_____" ], [ "### Introduction", "_____no_output_____" ], [ "In the long run, the bookmaker always wins. The aim of this project is to disprove exactly this. We are in the football sports betting market and are trying to develop a strategy that is profitable in the long term and which will make the bookmaker leave the pitch as the loser. There are three aspects to this strategy that need to be optimised. \n\nThese are:\n\n- the selection of suitable football matches\n- the prediction of the corresponding outcome\n- and the determination of the optimal stake per bet.\n\nIn order to achieve this goal, a data set is compiled containing data from almost 60,000 football matches from 22 different leagues. This data set is processed, evaluated and then used to develop the long-term strategy with the help of selected machine learning algorithms. \n\nThe data comes from the following source: [Data source](https://www.football-data.co.uk/downloadm.php)", "_____no_output_____" ], [ "### Merging the data", "_____no_output_____" ], [ "The first step is to read the data from 264 .csv files and combine them appropriately. Before the data set is saved, an additional column with information about the season of the match is created to ensure a unique allocation.", "_____no_output_____" ] ], [ [ "# import packages\nimport glob\nimport os\nimport pandas as pd", "_____no_output_____" ], [ "# loading the individual datasets of the different seasons\nfile_type = 'csv'\nseperator =','\ndf_20_21 = pd.concat([pd.read_csv(f, sep=seperator) for f in glob.glob('20:21' + \"/*.\"+file_type)],ignore_index=True)\ndf_19_20 = pd.concat([pd.read_csv(f, sep=seperator) for f in glob.glob('19:20' + \"/*.\"+file_type)],ignore_index=True)\ndf_18_19 = pd.concat([pd.read_csv(f, sep=seperator) for f in glob.glob('18:19' + \"/*.\"+file_type)],ignore_index=True)\ndf_17_18 = pd.concat([pd.read_csv(f, sep=seperator) for f in glob.glob('17:18' + \"/*.\"+file_type)],ignore_index=True)\ndf_16_17 = pd.concat([pd.read_csv(f, sep=seperator) for f in glob.glob('16:17' + \"/*.\"+file_type)],ignore_index=True)\ndf_15_16 = pd.concat([pd.read_csv(f, sep=seperator) for f in glob.glob('15:16' + \"/*.\"+file_type)],ignore_index=True)\ndf_14_15 = pd.concat([pd.read_csv(f, sep=seperator) for f in glob.glob('14:15' + \"/*.\"+file_type)],ignore_index=True)\ndf_13_14 = pd.concat([pd.read_csv(f, sep=seperator) for f in glob.glob('13:14' + \"/*.\"+file_type)],ignore_index=True)", "_____no_output_____" ], [ "# add a column of the season for clear assignment\ndf_20_21['Season'] = '20/21'\ndf_19_20['Season'] = '19/20'\ndf_18_19['Season'] = '18/19'\ndf_17_18['Season'] = '17/18'\ndf_16_17['Season'] = '16/17'\ndf_15_16['Season'] = '15/16'\ndf_14_15['Season'] = '14/15'\ndf_13_14['Season'] = '13/14'", "_____no_output_____" ], [ "# combining the individual datasets into one\ndfs = [df_14_15, df_15_16, df_16_17, df_17_18, df_18_19, df_19_20, df_20_21]\nresults = df_13_14.append(dfs, sort=False)", "_____no_output_____" ], [ "# saving the merged dataframe for processing\nresults.to_csv(\"Data/Results2013_2021.csv\")", "_____no_output_____" ] ], [ [ "### Quick Overview", "_____no_output_____" ] ], [ [ "# output of the data shape\nresults.shape", "_____no_output_____" ] ], [ [ "In its initial state, the data set comprises almost 60000 rows and 133 columns. In addition to information on league affiliation, the season of the match and the team constellation, information on the final result is available in the form of the number of goals, shots, shots on target, corners, fouls and yellow and red cards for home and away teams. In addition, the dataset contains information on betting odds from a large number of bookmakers.\n\nAs a large proportion of the columns are only sporadically filled, especially with regard to the betting odds, those bookmakers whose odds are available for the 60,000 matches were filtered. This procedure alone reduced the data set from 133 to 31 columns. ", "_____no_output_____" ] ], [ [ "# selecting the necessary columns of the original data set\nresults = results[['Div', 'Season', 'HomeTeam','AwayTeam', 'FTHG', 'FTAG', 'FTR', 'HS', 'AS', 'HST', 'AST', 'HF', 'AF', 'HC',\n 'AC', 'HY', 'AY', 'HR', 'AR','B365H','B365D','B365A', 'BWH','BWD','BWA', 'IWH', 'IWD', 'IWA', 'WHH', 'WHD', 'WHA']]\nresults.shape", "_____no_output_____" ] ], [ [ "Die verbleibenden Spalten werden in der folgenden Tabelle kurz erläutert:\n\n| Column | Description |\n| - | - |\n| `Div` | League Division |\n| `Season` | Season in which the match took place |\n| `HomeTeam` | Home Team |\n| `AwayTeam` | Away Team |\n| `FTHG` | Full Time Home Team Goals |\n| `FTAG`| Full Time Away Team Goals |\n| `FTR` | Full Time Result (H=Home Win, D=Draw, A=Away Win) |\n| `HS` | Home Team Shots |\n| `AS` | Away Team Shots |\n| `HST` | Home Team Shots on Target |\n| `AST` | Away Team Shots on Target |\n| `HF` | Home Team Fouls Committed |\n| `AF` | Away Team Fouls Committed |\n| `HC` | Home Team Corners |\n| `AC` | Away Team Corners |\n| `HY` | Home Team Yellow Cards |\n| `AY` | Away Team Yellow Cards |\n| `HR`| Home Team Red Cards |\n| `AR` | Away Team Red Cards |\n| `B365H` | Bet365 Home Win Odds |\n| `B365D` | Bet365 Draw Odds |\n| `B365A` | Bet365 Away Win Odds |\n| `BWH` | Bet&Win Home Win Odds |\n| `BWD` | Bet&Win Draw Odds |\n| `BWA` | Bet&Win Away Win Odds |\n| `IWH` | Interwetten Home Win Odds |\n| `IWD` | Interwetten Draw Odds |\n| `IWA` | Interwetten Away Win Odds |\n| `WHH` | William Hill Home Win Odds |\n| `WHD` | William Hill Draw Odds |\n| `WHA` | William Hill Away Win Odds |", "_____no_output_____" ], [ "Since one aspect of the objective is to use the data to predict football matches, it must be noted that, with the exception of the league, the season, the team constellation and the betting odds, this is exclusively information that only becomes known after the end of the match. Accordingly, the data in its present form cannot be used without further ado to predict the outcome of the match. In the [following notebook](https://github.com/mue94/oddstradamus/blob/main/02_Data_Processing.ipynb), the corresponding data is processed and transformed in such a way that it can contribute to the prediction without hesitation and without data leakage.", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown", "markdown", "markdown", "markdown", "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ] ]

4ad7751f5e13bfc2f9fe5f6ac2b206bc002549b5

45,601

ipynb

Jupyter Notebook

.ipynb_checkpoints/dataExploration[02]-checkpoint.ipynb

k----n/yeg311DataMining

7a08defaf89e03d38e12ed7813c4b2245875590c

[ "Apache-2.0" ]

null

.ipynb_checkpoints/dataExploration[02]-checkpoint.ipynb

k----n/yeg311DataMining

7a08defaf89e03d38e12ed7813c4b2245875590c

[ "Apache-2.0" ]

null

.ipynb_checkpoints/dataExploration[02]-checkpoint.ipynb

k----n/yeg311DataMining

7a08defaf89e03d38e12ed7813c4b2245875590c

[ "Apache-2.0" ]

null

285.00625

24,716

0.922589

[ [ [ "%matplotlib inline\nimport pandas as pd\ndf = pd.read_csv('data/311_Explorer_01.csv', parse_dates=\\\n ['Date Created','Date Closed'], \\\n low_memory = False)\ndf.columns", "/home/k/anaconda3/lib/python3.5/site-packages/matplotlib/font_manager.py:273: UserWarning: Matplotlib is building the font cache using fc-list. This may take a moment.\n warnings.warn('Matplotlib is building the font cache using fc-list. This may take a moment.')\n/home/k/anaconda3/lib/python3.5/site-packages/matplotlib/font_manager.py:273: UserWarning: Matplotlib is building the font cache using fc-list. This may take a moment.\n warnings.warn('Matplotlib is building the font cache using fc-list. This may take a moment.')\n" ], [ "import matplotlib.pyplot as plt \n\ndf['Days to Resolution'].plot.hist()\nplt.xlabel(\"Days to Resolve\")\nplt.show()", "_____no_output_____" ], [ "a = df['Days to Resolution'].value_counts()\nprint(\"Number of different values:\",len(a))\nprint(\"Number of values >= 200 days to resolve:\", len(a[a>=200]))\n\n# Most issues are resolved in the first 200 days", "Number of different values: 976\nNumber of values >= 200 days to resolve: 106\n" ], [ "# get the days of the week to plot\ndf['Day of Week Ticket Created'] = df['Date Created'].dt.weekday_name\n\ndf['Day of Week Ticket Created'].value_counts().plot.bar()\n\nplt.ylabel(\"Number of Tickets\")\nplt.xlabel(\"Day of Week Ticket Created\")\nplt.show()\n", "_____no_output_____" ], [ "df2 = pd.read_csv('data/2013-2017_Council_And_Committee_Meetings_-_Meeting_Details.csv', parse_dates=['MEETING_DATE'], low_memory = False)\ndf2['MEETING_TYPE'].value_counts()", "_____no_output_____" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code", "code" ] ]

4ad7790d32f0098c304909cc53ae17254bcc8462

28,193

ipynb

Jupyter Notebook

Python/tigre/demos/d03_simple_image_reconstruction.ipynb

sjha2048/TIGRE

343476921d5b137c7c0a5aa1f31b9c3a22f626a9

[ "BSD-3-Clause" ]

null

Python/tigre/demos/d03_simple_image_reconstruction.ipynb

sjha2048/TIGRE

343476921d5b137c7c0a5aa1f31b9c3a22f626a9

[ "BSD-3-Clause" ]

null

Python/tigre/demos/d03_simple_image_reconstruction.ipynb

sjha2048/TIGRE

343476921d5b137c7c0a5aa1f31b9c3a22f626a9

[ "BSD-3-Clause" ]

null

174.030864

23,973

0.894406

[ [ [ "# Demo 3: Simple Image reconstruction", "_____no_output_____" ] ], [ [ "This demo will show you a simple image reconstruction can be performed, by using OS_SART and FDK.\nNOTE: if you havent already downloaded the tigre_demo_file and navigated to the correct directory, do so before continuing with this demo. \n--------------------------------------------------------------------------\n--------------------------------------------------------------------------\nThis file is part of the TIGRE Toolbox\n\nCopyright (c) 2015, University of Bath and\n CERN-European Organization for Nuclear Research\n All rights reserved.\n\nLicense: Open Source under BSD.\n See the full license at\n https://github.com/CERN/TIGRE/license.txt\n\nContact: [email protected]\nCodes: https://github.com/CERN/TIGRE/\n--------------------------------------------------------------------------\nCoded by: MATLAB (original code): Ander Biguri\n PYTHON : Reuben Lindroos,Sam Loescher", "_____no_output_____" ] ], [ [ "## Define geometry", "_____no_output_____" ] ], [ [ "import tigre\ngeo = tigre.geometry_default(high_quality=False)", "_____no_output_____" ] ], [ [ "## Load data and generate projections", "_____no_output_____" ] ], [ [ "import numpy as np\nfrom tigre.Ax import Ax\nfrom Test_data import data_loader\n# define angles\nangles=np.linspace(0,2*np.pi,dtype=np.float32)\n# load head phantom data\nhead=data_loader.load_head_phantom(number_of_voxels=geo.nVoxel)\n# generate projections\nprojections=Ax(head,geo,angles,'interpolated')", "_____no_output_____" ] ], [ [ "## Reconstruct image using OS-SART and FDK", "_____no_output_____" ] ], [ [ "# OS_SART\nniter=50\nimgOSSART=tigre.algorithms.ossart(projections,geo,angles,niter, **dict(blocksize=20))\n\n# FDK \nimgfdk=tigre.algorithms.FDK(projections,geo,angles)\n\n# Show the results\nfrom tigre.utilities.plotImg import plotImg\nplotImg(np.hstack((imgOSSART,imgfdk)),slice=32,dim='x')", "Algorithm in progress.\nEsitmated time until completetion (s): 14.657517\n" ] ] ]

[ "markdown", "raw", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "raw" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad77ef3672154c701c6398576d00c8091c03650

245,864

ipynb

Jupyter Notebook

notebooks/benchmark-python.ipynb

ei-grad/fast-bernoulli

e992f5fc91c6a118e0fa441f1162b5334fbd92cd

[ "BSD-3-Clause" ]

5

2019-10-13T17:29:09.000Z

2021-09-28T13:27:22.000Z

notebooks/benchmark-python.ipynb

ei-grad/fast-bernoulli

e992f5fc91c6a118e0fa441f1162b5334fbd92cd

[ "BSD-3-Clause" ]

1

2022-01-21T15:56:29.000Z

2022-02-03T16:28:14.000Z

notebooks/benchmark-python.ipynb

ei-grad/fast-bernoulli

e992f5fc91c6a118e0fa441f1162b5334fbd92cd

[ "BSD-3-Clause" ]

2

2019-10-18T16:23:17.000Z

2022-01-21T15:39:41.000Z

664.497297

90,756

0.947988

[ [ [ "# Fast Bernoulli: Benchmark Python", "_____no_output_____" ], [ "In this notebooks we will measure performance of generating sequencies of Bernoulli-distributed random varibales in Python without and within LLVM JIT compiler. The baseline generator is based on top of expression `random.uniform() < p`.", "_____no_output_____" ] ], [ [ "import numpy as np\nimport matplotlib.pyplot as plt", "_____no_output_____" ], [ "from random import random\nfrom typing import List", "_____no_output_____" ], [ "from bernoulli import LLVMBernoulliGenerator, PyBernoulliGenerator\nfrom tqdm import tqdm", "_____no_output_____" ] ], [ [ "## Benchmarking", "_____no_output_____" ], [ "As it was mentioned above, the baseline generator is just thresholding a uniform-distributed random variable.", "_____no_output_____" ] ], [ [ "class BaselineBernoulliGenerator:\n \n def __init__(self, probability: float, tolerance: float = float('nan'), seed: int = None):\n self.prob = probability\n\n def __call__(self, nobits: int = 32):\n return [int(random() <= self.prob) for _ in range(nobits)]", "_____no_output_____" ] ], [ [ "Here we define some routines for benchmarking.", "_____no_output_____" ] ], [ [ "def benchmark(cls, nobits_list: List[int], probs: List[float], tol: float = 1e-6) -> np.ndarray:\n timings = np.empty((len(probs), len(nobits_list)))\n with tqdm(total=timings.size, unit='bench') as progress:\n for i, prob in enumerate(probs):\n generator = cls(prob, tol)\n for j, nobits in enumerate(nobits_list):\n try:\n timing = %timeit -q -o generator(nobits)\n timings[i, j] = timing.average\n except Exception as e:\n # Here we catch the case when number of bits is not enough \n # to obtain desirable precision.\n timings[i, j] = float('nan')\n progress.update()\n return timings", "_____no_output_____" ] ], [ [ "The proposed Bernoulli generator has two parameters. The first one is well-known that is probability of success $p$. The second one is precision of quantization.", "_____no_output_____" ] ], [ [ "NOBITS = [1, 2, 4, 8, 16, 32]\nPROBAS = [1 / 2 ** n for n in range(1, 8)]", "_____no_output_____" ] ], [ [ "Now, start benchmarking!", "_____no_output_____" ] ], [ [ "baseline = benchmark(BaselineBernoulliGenerator, NOBITS, PROBAS)", "100%|██████████| 42/42 [05:07<00:00, 6.98s/bench]" ], [ "py = benchmark(PyBernoulliGenerator, NOBITS, PROBAS)", "100%|██████████| 42/42 [02:54<00:00, 6.58s/bench]\u001b[A" ], [ "llvm = benchmark(LLVMBernoulliGenerator, NOBITS, PROBAS)", "100%|██████████| 42/42 [02:31<00:00, 3.01s/bench]\u001b[A\u001b[A" ] ], [ [ "Multiplication by factor $10^6$ corresponds to changing units from seconds to microseconds.", "_____no_output_____" ] ], [ [ "baseline *= 1e6\npy *= 1e6\nllvm *= 1e6", "_____no_output_____" ] ], [ [ "Save timings for the future.", "_____no_output_____" ] ], [ [ "np.save('../data/benchmark-data-baseline.npy', baseline)\nnp.save('../data/benchmark-data-py.npy', py)\nnp.save('../data/benchmark-data-llvm.npy', llvm)", "_____no_output_____" ] ], [ [ "## Visualization\n\nOn the figures below we depics how timings (or bitrate) depends on algorithm parameters.", "_____no_output_____" ] ], [ [ "fig = plt.figure(figsize=(14, 6))\n\nax = fig.add_subplot(1, 1, 1)\nax.grid()\n\nfor i, proba in enumerate(PROBAS):\n ax.loglog(NOBITS, baseline[i, :], '-x',label=f'baseline p={proba}')\n ax.loglog(NOBITS, py[i, :], '-+',label=f'python p={proba}')\n ax.loglog(NOBITS, llvm[i, :], '-o',label=f'llvm p={proba}')\n \nax.legend(loc='center left', bbox_to_anchor=(1, 0.5))\nax.set_xlabel('Sequence length, bit')\nax.set_ylabel('Click time, $\\mu s$')\n\nplt.show()", "_____no_output_____" ], [ "fig = plt.figure(figsize=(14, 6))\n\nax = fig.add_subplot(1, 1, 1)\nax.grid()\n\nfor i, proba in enumerate(PROBAS):\n ax.loglog(NOBITS, NOBITS / baseline[i, :], '-x',label=f'baseline p={proba}')\n ax.loglog(NOBITS, NOBITS / py[i, :], '-+',label=f'python p={proba}')\n ax.loglog(NOBITS, NOBITS / llvm[i, :], '-o',label=f'llvm p={proba}')\n \nax.legend(loc='center left', bbox_to_anchor=(1, 0.5))\nax.set_xlabel('Sequence length, bit')\nax.set_ylabel('Bit rate, Mbit per s')\n\nplt.show()", "_____no_output_____" ], [ "fig = plt.figure(figsize=(14, 6))\n\nax = fig.add_subplot(1, 1, 1)\nax.grid()\n\nfor j, nobits in enumerate(NOBITS):\n ax.loglog(PROBAS, nobits / baseline[:, j], '-x',label=f'baseline block={nobits}')\n ax.loglog(PROBAS, nobits / py[:, j], '-+',label=f'python block={nobits}')\n ax.loglog(PROBAS, nobits / llvm[:, j], '-o',label=f'llvm block={nobits}')\n \nax.legend(loc='center left', bbox_to_anchor=(1, 0.5))\nax.set_xlabel('Bernoulli parameter')\nax.set_ylabel('Bitrate, Mbit / sec')\n\nplt.show()", "_____no_output_____" ] ], [ [ "## Comments and Discussions\n\nOn the figures above one can see that direct implementation of the algorithm does not improve bitrate. Also, we can see that the statement is true for implementaion with LLVM as well as without LLVM. This means that overhead is too large.\n\nNevertheless, the third figure is worth to note that bitrate scales note very well for baseline generator. The bitrate of baseline drops dramatically while the bitrates of the others decrease much lesser.\n\nSuch benchmarking like this has unaccounted effects like different implementation levels (IR and Python), expansion of bit block to list of bits, overhead of Python object system.", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown", "markdown" ], [ "code", "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ] ]

4ad780231502ce1f75c089df41b5800c89ee2660

305,963

ipynb

Jupyter Notebook

Proyecto 3. Equipo.ipynb

Carlos2212/Proyecto3_Gonz-lezCarlos-Gonz-lezRaul-SerranoJuan

291b8aec2e67fdee02efc6e4c0a0500af5e65264

[ "MIT" ]

null

Proyecto 3. Equipo.ipynb

Carlos2212/Proyecto3_Gonz-lezCarlos-Gonz-lezRaul-SerranoJuan

291b8aec2e67fdee02efc6e4c0a0500af5e65264

[ "MIT" ]

null

Proyecto 3. Equipo.ipynb

Carlos2212/Proyecto3_Gonz-lezCarlos-Gonz-lezRaul-SerranoJuan

291b8aec2e67fdee02efc6e4c0a0500af5e65264

[ "MIT" ]

null

163.616578

93,588

0.850322

[ [ [ "# Proyecto 3.\n- Carlos González Mendoza\n- Raul Enrique González Paz\n- Juan Andres Serrano Rivera", "_____no_output_____" ], [ "Para este proyecto revisaremos los precios ajustados de las empresas **ADIDAS**, **NIKE** y **UNDER ARMOUR**, ya que son las dos empresas deportivas mas grandes del mundo. Y aparte son un trio de empresas con gran impacto en todo el mundo. \n\n<img style=\"float: right; margin: 0px 0px 10px 10px;\" src=\"http://content.nike.com/content/dam/one-nike/globalAssets/social_media_images/nike_swoosh_logo_black.png\" width=\"300px\" height=\"125px\" />\n\n<img style=\"float: right; margin: 0px 0px 10px 10px;\" src=\"https://upload.wikimedia.org/wikipedia/commons/thumb/2/20/Adidas_Logo.svg/2000px-Adidas_Logo.svg.png\" width=\"300px\" height=\"125px\" />\n\n<img style=\"float: right; margin: 0px 0px 10px 10px;\" src=\"https://upload.wikimedia.org/wikipedia/commons/thumb/4/44/Under_armour_logo.svg/2000px-Under_armour_logo.svg.png\" width=\"300px\" height=\"125px\" />", "_____no_output_____" ] ], [ [ "# Primero importamos todas las librerias que ocuparemos.\nimport pandas as pd\npd.core.common.is_list_like = pd.api.types.is_list_like\nimport pandas_datareader.data as web\nimport numpy as np\nimport matplotlib.pyplot as plt\n%matplotlib inline", "_____no_output_____" ], [ "# Despues creamos la funcion con la cual descargaremos los datos\ndef get_closes(tickers, start_date=None, end_date=None, freq='d'):\n # Por defecto la fecha de inicio es el 01 de Enero de 2010 y fecha de termino por defecto es hasta el dia de hoy.\n # Frecuencia de muestreo por defecto es 'd'.\n # Creamos DataFrame vacío de precios, con el índice de las fechas que necesitamos.\n closes = pd.DataFrame(columns = tickers, index=web.YahooDailyReader(symbols=tickers[0], start=start_date, end=end_date, interval=freq).read().index)\n # Una vez creado el DataFrame, agregamos cada uno de los precios con YahooDailyReader\n for ticker in tickers:\n df = web.YahooDailyReader(symbols=ticker, start=start_date, end=end_date, interval=freq).read()\n closes[ticker]=df['Adj Close']\n closes.index_name = 'Date'\n closes = closes.sort_index()\n return closes", "_____no_output_____" ], [ "# Instrumentos a descargar (Apple, Walmart, IBM, Nike)\nnames = ['NKE', 'ADDYY', 'UAA']\n# Fechas: inicios 2011 a finales de 2015\nstart, end = '2011-01-01', '2017-12-31'", "_____no_output_____" ] ], [ [ "- Entonces ya realizado lo anterior, podremos obtener los precios ajustados de las empresas **ADIDAS** y **NIKE**", "_____no_output_____" ] ], [ [ "closes = get_closes(names, start, end)\ncloses", "_____no_output_____" ], [ "closes.plot(figsize=(15,10))", "_____no_output_____" ], [ "closes.describe()", "_____no_output_____" ] ], [ [ "- Obtendremos los rendimientos diarios", "_____no_output_____" ] ], [ [ "closes.shift()", "_____no_output_____" ] ], [ [ "- Entonces se calculan los rendimientos diarios de la siguiente manera.", "_____no_output_____" ] ], [ [ "rend = ((closes-closes.shift())/closes.shift()).dropna()\nrend", "_____no_output_____" ] ], [ [ "- Y los rendimientos se grafican de la siguiente manera.", "_____no_output_____" ] ], [ [ "rend.plot(figsize=(15,10))", "_____no_output_____" ] ], [ [ "- Sacamos las medias y las desviaciones estandar de cada una de las empresas.", "_____no_output_____" ] ], [ [ "mu_NKE, mu_ADDYY, mu_UAA = rend.mean().NKE, rend.mean().ADDYY, rend.mean().UAA\ns_NKE, s_ADDYY, s_UAA = rend.std().NKE, rend.std().ADDYY, rend.std().UAA", "_____no_output_____" ] ], [ [ "- Simulamos 10000 escenarios de rendimientos diarios para el 2017 para las 3 empresas deportivas.", "_____no_output_____" ] ], [ [ "def rend_sim(mu, sigma, ndays, nscen, start_date):\n dates = pd.date_range(start=start_date,periods=ndays)\n return pd.DataFrame(data = sigma*np.random.randn(ndays, nscen)+mu, index = dates)", "_____no_output_____" ], [ "simrend_NKE = rend_sim(mu_NKE, s_NKE, 252, 10000, '2018-01-01')\nsimrend_ADDYY = rend_sim(mu_ADDYY, s_ADDYY, 252, 10000, '2018-01-01')\nsimrend_UAA = rend_sim(mu_NKE, s_ADDYY, 252, 10000, '2018-01-01')", "_____no_output_____" ], [ "simcloses_NKE = closes.iloc[-1].NKE*((1+simrend_NKE).cumprod())\nsimcloses_ADDYY = closes.iloc[-1].ADDYY*((1+simrend_ADDYY).cumprod())\nsimcloses_UAA = closes.iloc[-1].UAA*((1+simrend_UAA).cumprod())", "_____no_output_____" ] ], [ [ "- Calculamos la probabilidad de que el precio incremente un 10% el siguiente año", "_____no_output_____" ] ], [ [ "K_NKE = (1+0.05)*closes.iloc[-1].NKE\nprob_NKE = pd.DataFrame((simcloses_NKE>K_NKE).sum(axis=1)/10000)\nprob_NKE.plot(figsize=(10,6), grid=True, color = 'r');\n\nK_ADDYY = (1+0.05)*closes.iloc[-1].ADDYY\nprob_ADDYY = pd.DataFrame((simcloses_ADDYY>K_ADDYY).sum(axis=1)/10000)\nprob_ADDYY.plot(figsize=(10,6), grid=True, color = 'g');\n\nK_UAA = (1+0.05)*closes.iloc[-1].UAA\nprob_UAA = pd.DataFrame((simcloses_UAA>K_UAA).sum(axis=1)/10000)\nprob_UAA.plot(figsize=(10,6), grid=True);", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ] ]

4ad79bbcbd1a971b1ecf437fc5c7a1c73b7e82dd

197,252

ipynb

Jupyter Notebook

source_code/Plant_Pathology_Modeling_PyTorch_V1.ipynb

nhutnamhcmus/ml-lab-02-classification

0130e89eef80e9f69d2a6c578313841c770bc7f7

[ "MIT" ]

1

2021-06-02T13:06:21.000Z

source_code/Plant_Pathology_Modeling_PyTorch_V1.ipynb

nhutnamhcmus/ml-lab-02-classification

0130e89eef80e9f69d2a6c578313841c770bc7f7

[ "MIT" ]

null

source_code/Plant_Pathology_Modeling_PyTorch_V1.ipynb

nhutnamhcmus/ml-lab-02-classification

0130e89eef80e9f69d2a6c578313841c770bc7f7

[ "MIT" ]

1

2021-11-05T03:38:51.000Z

60.543892

28,666

0.563629

[ [ [ "# Install dependencies", "_____no_output_____" ] ], [ [ "!pip install pretrainedmodels\n!pip install albumentations==0.4.5", "Collecting pretrainedmodels\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/84/0e/be6a0e58447ac16c938799d49bfb5fb7a80ac35e137547fc6cee2c08c4cf/pretrainedmodels-0.7.4.tar.gz (58kB)\n\r\u001b[K |█████▋ | 10kB 17.9MB/s eta 0:00:01\r\u001b[K |███████████▏ | 20kB 23.0MB/s eta 0:00:01\r\u001b[K |████████████████▊ | 30kB 22.2MB/s eta 0:00:01\r\u001b[K |██████████████████████▎ | 40kB 17.3MB/s eta 0:00:01\r\u001b[K |███████████████████████████▉ | 51kB 8.5MB/s eta 0:00:01\r\u001b[K |████████████████████████████████| 61kB 4.0MB/s \n\u001b[?25hRequirement already satisfied: torch in /usr/local/lib/python3.7/dist-packages (from pretrainedmodels) (1.8.1+cu101)\nRequirement already satisfied: torchvision in /usr/local/lib/python3.7/dist-packages (from pretrainedmodels) (0.9.1+cu101)\nCollecting munch\n Downloading https://files.pythonhosted.org/packages/cc/ab/85d8da5c9a45e072301beb37ad7f833cd344e04c817d97e0cc75681d248f/munch-2.5.0-py2.py3-none-any.whl\nRequirement already satisfied: tqdm in /usr/local/lib/python3.7/dist-packages (from pretrainedmodels) (4.41.1)\nRequirement already satisfied: numpy in /usr/local/lib/python3.7/dist-packages (from torch->pretrainedmodels) (1.19.5)\nRequirement already satisfied: typing-extensions in /usr/local/lib/python3.7/dist-packages (from torch->pretrainedmodels) (3.7.4.3)\nRequirement already satisfied: pillow>=4.1.1 in /usr/local/lib/python3.7/dist-packages (from torchvision->pretrainedmodels) (7.1.2)\nRequirement already satisfied: six in /usr/local/lib/python3.7/dist-packages (from munch->pretrainedmodels) (1.15.0)\nBuilding wheels for collected packages: pretrainedmodels\n Building wheel for pretrainedmodels (setup.py) ... \u001b[?25l\u001b[?25hdone\n Created wheel for pretrainedmodels: filename=pretrainedmodels-0.7.4-cp37-none-any.whl size=60963 sha256=30bc806145b3e8feb8ac06611a23dfcf0503aa3d54d17ca23bec3a9429015671\n Stored in directory: /root/.cache/pip/wheels/69/df/63/62583c096289713f22db605aa2334de5b591d59861a02c2ecd\nSuccessfully built pretrainedmodels\nInstalling collected packages: munch, pretrainedmodels\nSuccessfully installed munch-2.5.0 pretrainedmodels-0.7.4\nCollecting albumentations==0.4.5\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/8d/40/a343ecacc7e22fe52ab9a16b84dc6165ba05ee17e3729adeb3e2ffa2b37b/albumentations-0.4.5.tar.gz (116kB)\n\u001b[K |████████████████████████████████| 122kB 7.2MB/s \n\u001b[?25hRequirement already satisfied: numpy>=1.11.1 in /usr/local/lib/python3.7/dist-packages (from albumentations==0.4.5) (1.19.5)\nRequirement already satisfied: scipy in /usr/local/lib/python3.7/dist-packages (from albumentations==0.4.5) (1.4.1)\nCollecting imgaug<0.2.7,>=0.2.5\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/ad/2e/748dbb7bb52ec8667098bae9b585f448569ae520031932687761165419a2/imgaug-0.2.6.tar.gz (631kB)\n\u001b[K |████████████████████████████████| 634kB 12.0MB/s \n\u001b[?25hRequirement already satisfied: PyYAML in /usr/local/lib/python3.7/dist-packages (from albumentations==0.4.5) (3.13)\nRequirement already satisfied: opencv-python>=4.1.1 in /usr/local/lib/python3.7/dist-packages (from albumentations==0.4.5) (4.1.2.30)\nRequirement already satisfied: scikit-image>=0.11.0 in /usr/local/lib/python3.7/dist-packages (from imgaug<0.2.7,>=0.2.5->albumentations==0.4.5) (0.16.2)\nRequirement already satisfied: six in /usr/local/lib/python3.7/dist-packages (from imgaug<0.2.7,>=0.2.5->albumentations==0.4.5) (1.15.0)\nRequirement already satisfied: PyWavelets>=0.4.0 in /usr/local/lib/python3.7/dist-packages (from scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations==0.4.5) (1.1.1)\nRequirement already satisfied: matplotlib!=3.0.0,>=2.0.0 in /usr/local/lib/python3.7/dist-packages (from scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations==0.4.5) (3.2.2)\nRequirement already satisfied: networkx>=2.0 in /usr/local/lib/python3.7/dist-packages (from scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations==0.4.5) (2.5.1)\nRequirement already satisfied: pillow>=4.3.0 in /usr/local/lib/python3.7/dist-packages (from scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations==0.4.5) (7.1.2)\nRequirement already satisfied: imageio>=2.3.0 in /usr/local/lib/python3.7/dist-packages (from scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations==0.4.5) (2.4.1)\nRequirement already satisfied: cycler>=0.10 in /usr/local/lib/python3.7/dist-packages (from matplotlib!=3.0.0,>=2.0.0->scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations==0.4.5) (0.10.0)\nRequirement already satisfied: pyparsing!=2.0.4,!=2.1.2,!=2.1.6,>=2.0.1 in /usr/local/lib/python3.7/dist-packages (from matplotlib!=3.0.0,>=2.0.0->scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations==0.4.5) (2.4.7)\nRequirement already satisfied: kiwisolver>=1.0.1 in /usr/local/lib/python3.7/dist-packages (from matplotlib!=3.0.0,>=2.0.0->scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations==0.4.5) (1.3.1)\nRequirement already satisfied: python-dateutil>=2.1 in /usr/local/lib/python3.7/dist-packages (from matplotlib!=3.0.0,>=2.0.0->scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations==0.4.5) (2.8.1)\nRequirement already satisfied: decorator<5,>=4.3 in /usr/local/lib/python3.7/dist-packages (from networkx>=2.0->scikit-image>=0.11.0->imgaug<0.2.7,>=0.2.5->albumentations==0.4.5) (4.4.2)\nBuilding wheels for collected packages: albumentations, imgaug\n Building wheel for albumentations (setup.py) ... \u001b[?25l\u001b[?25hdone\n Created wheel for albumentations: filename=albumentations-0.4.5-cp37-none-any.whl size=64378 sha256=595bc1dcfd86e12f30807299092978000597c4f50a03636d7a45344d53189278\n Stored in directory: /root/.cache/pip/wheels/f0/a0/61/e50f93165a5ec7e7f5d65064e513239505bc4c06d2289557d3\n Building wheel for imgaug (setup.py) ... \u001b[?25l\u001b[?25hdone\n Created wheel for imgaug: filename=imgaug-0.2.6-cp37-none-any.whl size=654019 sha256=9bdfe6fad796a688884730d599491a37f67da626e9e0339f38cef64c4e748ad4\n Stored in directory: /root/.cache/pip/wheels/97/ec/48/0d25896c417b715af6236dbcef8f0bed136a1a5e52972fc6d0\nSuccessfully built albumentations imgaug\nInstalling collected packages: imgaug, albumentations\n Found existing installation: imgaug 0.2.9\n Uninstalling imgaug-0.2.9:\n Successfully uninstalled imgaug-0.2.9\n Found existing installation: albumentations 0.1.12\n Uninstalling albumentations-0.1.12:\n Successfully uninstalled albumentations-0.1.12\nSuccessfully installed albumentations-0.4.5 imgaug-0.2.6\n" ], [ "!pip install transformers", "Collecting transformers\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/d5/43/cfe4ee779bbd6a678ac6a97c5a5cdeb03c35f9eaebbb9720b036680f9a2d/transformers-4.6.1-py3-none-any.whl (2.2MB)\n\u001b[K |████████████████████████████████| 2.3MB 7.4MB/s \n\u001b[?25hRequirement already satisfied: regex!=2019.12.17 in /usr/local/lib/python3.7/dist-packages (from transformers) (2019.12.20)\nRequirement already satisfied: requests in /usr/local/lib/python3.7/dist-packages (from transformers) (2.23.0)\nRequirement already satisfied: tqdm>=4.27 in /usr/local/lib/python3.7/dist-packages (from transformers) (4.41.1)\nRequirement already satisfied: numpy>=1.17 in /usr/local/lib/python3.7/dist-packages (from transformers) (1.19.5)\nRequirement already satisfied: filelock in /usr/local/lib/python3.7/dist-packages (from transformers) (3.0.12)\nRequirement already satisfied: packaging in /usr/local/lib/python3.7/dist-packages (from transformers) (20.9)\nRequirement already satisfied: importlib-metadata; python_version < \"3.8\" in /usr/local/lib/python3.7/dist-packages (from transformers) (4.0.1)\nCollecting huggingface-hub==0.0.8\n Downloading https://files.pythonhosted.org/packages/a1/88/7b1e45720ecf59c6c6737ff332f41c955963090a18e72acbcbeac6b25e86/huggingface_hub-0.0.8-py3-none-any.whl\nCollecting tokenizers<0.11,>=0.10.1\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/d4/e2/df3543e8ffdab68f5acc73f613de9c2b155ac47f162e725dcac87c521c11/tokenizers-0.10.3-cp37-cp37m-manylinux_2_5_x86_64.manylinux1_x86_64.manylinux_2_12_x86_64.manylinux2010_x86_64.whl (3.3MB)\n\u001b[K |████████████████████████████████| 3.3MB 42.6MB/s \n\u001b[?25hCollecting sacremoses\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/75/ee/67241dc87f266093c533a2d4d3d69438e57d7a90abb216fa076e7d475d4a/sacremoses-0.0.45-py3-none-any.whl (895kB)\n\u001b[K |████████████████████████████████| 901kB 50.2MB/s \n\u001b[?25hRequirement already satisfied: urllib3!=1.25.0,!=1.25.1,<1.26,>=1.21.1 in /usr/local/lib/python3.7/dist-packages (from requests->transformers) (1.24.3)\nRequirement already satisfied: chardet<4,>=3.0.2 in /usr/local/lib/python3.7/dist-packages (from requests->transformers) (3.0.4)\nRequirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.7/dist-packages (from requests->transformers) (2020.12.5)\nRequirement already satisfied: idna<3,>=2.5 in /usr/local/lib/python3.7/dist-packages (from requests->transformers) (2.10)\nRequirement already satisfied: pyparsing>=2.0.2 in /usr/local/lib/python3.7/dist-packages (from packaging->transformers) (2.4.7)\nRequirement already satisfied: zipp>=0.5 in /usr/local/lib/python3.7/dist-packages (from importlib-metadata; python_version < \"3.8\"->transformers) (3.4.1)\nRequirement already satisfied: typing-extensions>=3.6.4; python_version < \"3.8\" in /usr/local/lib/python3.7/dist-packages (from importlib-metadata; python_version < \"3.8\"->transformers) (3.7.4.3)\nRequirement already satisfied: joblib in /usr/local/lib/python3.7/dist-packages (from sacremoses->transformers) (1.0.1)\nRequirement already satisfied: six in /usr/local/lib/python3.7/dist-packages (from sacremoses->transformers) (1.15.0)\nRequirement already satisfied: click in /usr/local/lib/python3.7/dist-packages (from sacremoses->transformers) (7.1.2)\nInstalling collected packages: huggingface-hub, tokenizers, sacremoses, transformers\nSuccessfully installed huggingface-hub-0.0.8 sacremoses-0.0.45 tokenizers-0.10.3 transformers-4.6.1\n" ], [ "# install dependencies for TPU\n#!curl https://raw.githubusercontent.com/pytorch/xla/master/contrib/scripts/env-setup.py -o pytorch-xla-env-setup.py\n#!python pytorch-xla-env-setup.py --apt-packages libomp5 libopenblas-dev", "_____no_output_____" ] ], [ [ "# Download data", "_____no_output_____" ] ], [ [ "# https://drive.google.com/file/d/1jfkX_NXF8shxyWZCxJkzsLPDr4ebvdOP/view?usp=sharing\n!pip install gdown\n!gdown https://drive.google.com/uc?id=1jfkX_NXF8shxyWZCxJkzsLPDr4ebvdOP\n!unzip -q plant-pathology-2020-fgvc7.zip -d /content/plant-pathology-2020-fgvc7\n!rm plant-pathology-2020-fgvc7.zip", "Requirement already satisfied: gdown in /usr/local/lib/python3.7/dist-packages (3.6.4)\nRequirement already satisfied: requests in /usr/local/lib/python3.7/dist-packages (from gdown) (2.23.0)\nRequirement already satisfied: tqdm in /usr/local/lib/python3.7/dist-packages (from gdown) (4.41.1)\nRequirement already satisfied: six in /usr/local/lib/python3.7/dist-packages (from gdown) (1.15.0)\nRequirement already satisfied: certifi>=2017.4.17 in /usr/local/lib/python3.7/dist-packages (from requests->gdown) (2020.12.5)\nRequirement already satisfied: idna<3,>=2.5 in /usr/local/lib/python3.7/dist-packages (from requests->gdown) (2.10)\nRequirement already satisfied: chardet<4,>=3.0.2 in /usr/local/lib/python3.7/dist-packages (from requests->gdown) (3.0.4)\nRequirement already satisfied: urllib3!=1.25.0,!=1.25.1,<1.26,>=1.21.1 in /usr/local/lib/python3.7/dist-packages (from requests->gdown) (1.24.3)\nDownloading...\nFrom: https://drive.google.com/uc?id=1jfkX_NXF8shxyWZCxJkzsLPDr4ebvdOP\nTo: /content/plant-pathology-2020-fgvc7.zip\n817MB [00:04, 199MB/s]\n" ] ], [ [ "# Import libraries", "_____no_output_____" ] ], [ [ "# Import os\nimport os\n\n# Import libraries for data manipulation\nimport numpy as np\nimport pandas as pd\n\n# Import libries for data agumentations: albumentations\nimport albumentations as A\nfrom albumentations.pytorch.transforms import ToTensor\nfrom albumentations import Rotate \nimport cv2 as cv\n\n# Import Pytorch\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torch.optim as optim\nimport torchvision\nfrom torch.utils.data import Dataset, DataLoader\n\n# Import pretrainmodels\nimport pretrainedmodels\n\n# Import transformers\nfrom transformers import get_cosine_schedule_with_warmup\nfrom transformers import AdamW\n\n# Import metrics for model evaluation\nfrom sklearn.metrics import roc_auc_score\nfrom sklearn.model_selection import StratifiedKFold\n\n# Import libraries for data visualization\nimport matplotlib.pyplot as plt\n\n# Import tqdm.notebook for loading visualization\nfrom tqdm.notebook import tqdm\n\n# Ignore warnings\nimport warnings \nwarnings.filterwarnings('ignore')", "_____no_output_____" ], [ "# Import for TPU configuration\n#import torch_xla\n#import torch_xla.core.xla_model as xm", "_____no_output_____" ] ], [ [ "# Settings", "_____no_output_____" ] ], [ [ "# Configuration path\n\n# Data folder\nIMAGES_PATH = '/content/plant-pathology-2020-fgvc7/images/'\n\n# Sample submission csv\nSAMPLE_SUBMISSION = '/content/plant-pathology-2020-fgvc7/sample_submission.csv'\n\n# Train, test data path\nTRAIN_DATA = '/content/plant-pathology-2020-fgvc7/train.csv'\nTEST_DATA = '/content/plant-pathology-2020-fgvc7/test.csv'", "_____no_output_____" ], [ "# Configuration for training workflow\nSEED = 1234\nN_FOLDS = 5\nN_EPOCHS = 20\nBATCH_SIZE = 2\nSIZE = 512\nIMG_SHAPE = (1365, 2048, 3)\nlr = 8e-4", "_____no_output_____" ], [ "submission_df = pd.read_csv(SAMPLE_SUBMISSION)\ndf_train = pd.read_csv(TRAIN_DATA)\ndf_test = pd.read_csv(TEST_DATA)\n\ndef get_image_path(filename):\n return (IMAGES_PATH + filename + '.jpg')\n\n#df_train['image_path'] = df_train['image_id'].apply(get_image_path)\n#df_test['image_path'] = df_test['image_id'].apply(get_image_path)\n#rain_labels = df_train.loc[:, 'healthy':'scab']\n#train_paths = df_train.image_path\n#test_paths = df_test.image_path", "_____no_output_____" ], [ "df_train.head()", "_____no_output_____" ], [ "df_test.head()", "_____no_output_____" ], [ "submission_df.head()", "_____no_output_____" ], [ "# for GPU\ndevice = torch.device(\"cuda:0\" if torch.cuda.is_available() else \"cpu\")\ndevice\n\n\n# for TPU\n#device = xm.xla_device()\n#torch.set_default_tensor_type('torch.FloatTensor')\n#device", "_____no_output_____" ] ], [ [ "# Define Dataset", "_____no_output_____" ] ], [ [ "class PlantDataset(Dataset):\n \n def __init__(self, df, transforms=None):\n self.df = df\n self.transforms=transforms\n \n def __len__(self):\n return self.df.shape[0]\n \n def __getitem__(self, idx):\n # Solution 01: Read from raw image\n image_src = IMAGES_PATH + self.df.loc[idx, 'image_id'] + '.jpg'\n\n # Solution 02: Read from npy file, we convert all images in images folder from .jpg to .npy\n # image_src = np.load(IMAGES_PATH + self.df.loc[idx, 'image_id'] + '.npy')\n\n # print(image_src)\n image = cv.imread(image_src, cv.IMREAD_COLOR)\n if image.shape != IMG_SHAPE:\n image = image.transpose(1, 0, 2)\n image = cv.cvtColor(image, cv.COLOR_BGR2RGB)\n labels = self.df.loc[idx, ['healthy', 'multiple_diseases', 'rust', 'scab']].values\n labels = torch.from_numpy(labels.astype(np.int8))\n labels = labels.unsqueeze(-1)\n \n if self.transforms:\n transformed = self.transforms(image=image)\n image = transformed['image']\n\n return image, labels", "_____no_output_____" ] ], [ [ "# Data Agumentations", "_____no_output_____" ] ], [ [ "# Train transformation\ntransforms_train = A.Compose([\n A.RandomResizedCrop(height=SIZE, width=SIZE, p=1.0),\n A.OneOf([A.RandomBrightness(limit=0.1, p=1), A.RandomContrast(limit=0.1, p=1)]),\n A.OneOf([A.MotionBlur(blur_limit=3), A.MedianBlur(blur_limit=3), A.GaussianBlur(blur_limit=3)], p=0.5),\n A.VerticalFlip(p=0.5),\n A.HorizontalFlip(p=0.5),\n A.ShiftScaleRotate(\n shift_limit=0.2,\n scale_limit=0.2,\n rotate_limit=20,\n interpolation=cv.INTER_LINEAR,\n border_mode=cv.BORDER_REFLECT_101,\n p=1,\n ),\n A.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225), max_pixel_value=255.0, p=1.0),\n A.pytorch.ToTensorV2(p=1.0),\n ], p=1.0)\n\n# Validation transformation\ntransforms_valid = A.Compose([\n A.Resize(height=SIZE, width=SIZE, p=1.0),\n A.Normalize(mean=(0.485, 0.456, 0.406), std=(0.229, 0.224, 0.225), max_pixel_value=255.0, p=1.0),\n A.pytorch.ToTensorV2(p=1.0),\n])", "_____no_output_____" ] ], [ [ "# StratifiedKFold", "_____no_output_____" ] ], [ [ "# Get label from df_train\ntrain_labels = df_train.iloc[:, 1:].values\ntrain_y = train_labels[:, 2] + train_labels[:, 3] * 2 + train_labels[:, 1] * 3", "_____no_output_____" ], [ "folds = StratifiedKFold(n_splits=N_FOLDS, shuffle=True, random_state=SEED)\noof_preds = np.zeros((df_train.shape[0], 4))", "_____no_output_____" ] ], [ [ "# PretrainedModels", "_____no_output_____" ], [ "## Define cross entropy loss one hot", "_____no_output_____" ] ], [ [ "# define cross entropy loss one hot\nclass CrossEntropyLossOneHot(nn.Module):\n def __init__(self):\n super(CrossEntropyLossOneHot, self).__init__()\n self.log_softmax = nn.LogSoftmax(dim=-1)\n\n def forward(self, preds, labels):\n return torch.mean(torch.sum(-labels * self.log_softmax(preds), -1))", "_____no_output_____" ] ], [ [ "## Define dense cross entropy", "_____no_output_____" ] ], [ [ "# define dense cross entropy\nclass DenseCrossEntropy(nn.Module):\n\n def __init__(self):\n super(DenseCrossEntropy, self).__init__()\n \n \n def forward(self, logits, labels):\n logits = logits.float()\n labels = labels.float()\n \n logprobs = F.log_softmax(logits, dim=-1)\n \n loss = -labels * logprobs\n loss = loss.sum(-1)\n\n return loss.mean()", "_____no_output_____" ] ], [ [ "## Define plant model with ResNet34", "_____no_output_____" ] ], [ [ "# define plant model with ResNet\nclass PlantModel(nn.Module):\n # define init function\n def __init__(self, num_classes=4):\n super().__init__()\n \n self.backbone = torchvision.models.resnet34(pretrained=True)\n \n in_features = self.backbone.fc.in_features\n\n self.logit = nn.Linear(in_features, num_classes)\n # define forward function\n def forward(self, x):\n batch_size, C, H, W = x.shape\n \n x = self.backbone.conv1(x)\n x = self.backbone.bn1(x)\n x = self.backbone.relu(x)\n x = self.backbone.maxpool(x)\n\n x = self.backbone.layer1(x)\n x = self.backbone.layer2(x)\n x = self.backbone.layer3(x)\n x = self.backbone.layer4(x)\n \n x = F.adaptive_avg_pool2d(x,1).reshape(batch_size,-1)\n x = F.dropout(x, 0.25, self.training)\n\n x = self.logit(x)\n\n return x", "_____no_output_____" ] ], [ [ "# Train for StratifiedKFold", "_____no_output_____" ] ], [ [ "def train_one_fold(i_fold, model, criterion, optimizer, lr_scheduler, dataloader_train, dataloader_valid):\n \n train_fold_results = []\n\n for epoch in range(N_EPOCHS):\n # print information\n print(' Epoch {}/{}'.format(epoch + 1, N_EPOCHS))\n print(' ' + ('-' * 20))\n os.system(f'echo \\\" Epoch {epoch}\\\"')\n\n # call model\n model.train()\n tr_loss = 0\n\n # looping\n for step, batch in enumerate(dataloader_train):\n # data preparation\n images = batch[0].to(device)\n labels = batch[1].to(device)\n \n # forward pass and calculate loss\n outputs = model(images)\n loss = criterion(outputs, labels.squeeze(-1))\n\n # backward pass \n loss.backward()\n\n tr_loss += loss.item()\n \n # updates\n # for TPU\n #xm.optimizer_step(optimizer, barrier=True)\n\n # for GPU\n optimizer.step()\n \n # empty gradient\n optimizer.zero_grad()\n\n # Validate\n model.eval()\n \n # init validation loss, predicted and labels\n val_loss = 0\n val_preds = None\n val_labels = None\n\n for step, batch in enumerate(dataloader_valid):\n # data preparation\n images = batch[0].to(device)\n labels = batch[1].to(device)\n\n # labels preparation\n if val_labels is None:\n val_labels = labels.clone().squeeze(-1)\n else:\n val_labels = torch.cat((val_labels, labels.squeeze(-1)), dim=0)\n\n \n # disable torch grad to calculating normally\n with torch.no_grad():\n # calculate the output\n outputs = model(batch[0].to(device))\n\n # calculate the loss value\n loss = criterion(outputs, labels.squeeze(-1))\n val_loss += loss.item()\n\n # predict with softmax activation function\n preds = torch.softmax(outputs, dim=1).data.cpu()\n #preds = torch.softmax(outputs, dim=1).detach().cpu().numpy()\n\n \n if val_preds is None:\n val_preds = preds\n else:\n val_preds = torch.cat((val_preds, preds), dim=0)\n \n # if train mode\n lr_scheduler.step(tr_loss)\n\n with torch.no_grad():\n train_loss = tr_loss / len(dataloader_train)\n valid_loss = val_loss / len(dataloader_valid)\n valid_score = roc_auc_score(val_labels.view(-1).cpu(), val_preds.view(-1).cpu(), average='macro')\n # print information\n if epoch % 2 == 0:\n print(f'Fold {i_fold} Epoch {epoch}: train_loss={train_loss:.4f}, valid_loss={valid_loss:.4f}, acc={valid_score:.4f}')\n train_fold_results.append({\n 'fold': i_fold,\n 'epoch': epoch,\n 'train_loss': train_loss,\n 'valid_loss': valid_loss,\n 'valid_score': valid_score,\n })\n\n return val_preds, train_fold_results", "_____no_output_____" ] ], [ [ "## Prepare submission file", "_____no_output_____" ] ], [ [ "submission_df.iloc[:, 1:] = 0", "_____no_output_____" ] ], [ [ "## Dataset test", "_____no_output_____" ] ], [ [ "dataset_test = PlantDataset(df=submission_df, transforms=transforms_valid)\ndataloader_test = DataLoader(dataset_test, batch_size=BATCH_SIZE, num_workers=4, shuffle=False)", "_____no_output_____" ] ], [ [ "# Init model: EfficientNetB5", "_____no_output_____" ] ], [ [ "# EfficientNetB5\n# B5 is the largest EfficientNet variant that fits in GPU memory with batch size 8.\n!pip install efficientnet_pytorch\n\nfrom efficientnet_pytorch import EfficientNet", "Collecting efficientnet_pytorch\n Downloading https://files.pythonhosted.org/packages/2e/a0/dd40b50aebf0028054b6b35062948da01123d7be38d08b6b1e5435df6363/efficientnet_pytorch-0.7.1.tar.gz\nRequirement already satisfied: torch in /usr/local/lib/python3.7/dist-packages (from efficientnet_pytorch) (1.8.1+cu101)\nRequirement already satisfied: numpy in /usr/local/lib/python3.7/dist-packages (from torch->efficientnet_pytorch) (1.19.5)\nRequirement already satisfied: typing-extensions in /usr/local/lib/python3.7/dist-packages (from torch->efficientnet_pytorch) (3.7.4.3)\nBuilding wheels for collected packages: efficientnet-pytorch\n Building wheel for efficientnet-pytorch (setup.py) ... \u001b[?25l\u001b[?25hdone\n Created wheel for efficientnet-pytorch: filename=efficientnet_pytorch-0.7.1-cp37-none-any.whl size=16443 sha256=a892a26c440e5bd908d1011a93001b287ff579cc0b95da55096c809dfe4560c2\n Stored in directory: /root/.cache/pip/wheels/84/27/aa/c46d23c4e8cc72d41283862b1437e0b3ad318417e8ed7d5921\nSuccessfully built efficientnet-pytorch\nInstalling collected packages: efficientnet-pytorch\nSuccessfully installed efficientnet-pytorch-0.7.1\n" ], [ "model = EfficientNet.from_pretrained('efficientnet-b5')\n\nnum_ftrs = model._fc.in_features\nmodel._fc = nn.Sequential(nn.Linear(num_ftrs,1000,bias=True),\n nn.ReLU(),\n nn.Dropout(p=0.5),\n nn.Linear(1000,4, bias = True))\n\nmodel.to(device)\nmodel", "Downloading: \"https://github.com/lukemelas/EfficientNet-PyTorch/releases/download/1.0/efficientnet-b5-b6417697.pth\" to /root/.cache/torch/hub/checkpoints/efficientnet-b5-b6417697.pth\n" ] ], [ [ "# Init model: ResNet", "_____no_output_____" ] ], [ [ "\"\"\"\n# Download pretrained weights.\n# model = PlantModel(num_classes=4)\nmodel = torchvision.models.resnet18(pretrained=True)\n\n# print number of features\nnum_features = model.fc.in_features\nprint(num_features)\n\n\n# custome layers\nmodel.fc = nn.Sequential(\n nn.Linear(num_features, 512),\n nn.ReLU(),\n nn.BatchNorm1d(512),\n nn.Dropout(0.5),\n \n nn.Linear(512, 256),\n nn.ReLU(),\n nn.BatchNorm1d(256),\n nn.Dropout(0.5),\n \n nn.Linear(256, 4))\n\n# initialize weights function\ndef init_weights(m):\n if type(m) == nn.Linear:\n torch.nn.init.xavier_uniform_(m.weight)\n m.bias.data.fill_(0.01)\n\n# apply model with init weights\nmodel.apply(init_weights)\n\n# transfer model to device (cuda:0 mean using GPU, xla mean using TPU, otherwise using CPU)\nmodel = model.to(device)\n\n# Model details\nmodel\n\"\"\"", "_____no_output_____" ], [ "print(torch.cuda.memory_summary(device=None, abbreviated=False))", "|===========================================================================|\n| PyTorch CUDA memory summary, device ID 0 |\n|---------------------------------------------------------------------------|\n| CUDA OOMs: 0 | cudaMalloc retries: 0 |\n|===========================================================================|\n| Metric | Cur Usage | Peak Usage | Tot Alloc | Tot Freed |\n|---------------------------------------------------------------------------|\n| Allocated memory | 120937 KB | 120937 KB | 120937 KB | 0 B |\n| from large pool | 85888 KB | 85888 KB | 85888 KB | 0 B |\n| from small pool | 35049 KB | 35049 KB | 35049 KB | 0 B |\n|---------------------------------------------------------------------------|\n| Active memory | 120937 KB | 120937 KB | 120937 KB | 0 B |\n| from large pool | 85888 KB | 85888 KB | 85888 KB | 0 B |\n| from small pool | 35049 KB | 35049 KB | 35049 KB | 0 B |\n|---------------------------------------------------------------------------|\n| GPU reserved memory | 139264 KB | 139264 KB | 139264 KB | 0 B |\n| from large pool | 102400 KB | 102400 KB | 102400 KB | 0 B |\n| from small pool | 36864 KB | 36864 KB | 36864 KB | 0 B |\n|---------------------------------------------------------------------------|\n| Non-releasable memory | 18327 KB | 22180 KB | 108539 KB | 90212 KB |\n| from large pool | 16512 KB | 20212 KB | 81188 KB | 64676 KB |\n| from small pool | 1815 KB | 2468 KB | 27351 KB | 25536 KB |\n|---------------------------------------------------------------------------|\n| Allocations | 856 | 856 | 856 | 0 |\n| from large pool | 29 | 29 | 29 | 0 |\n| from small pool | 827 | 827 | 827 | 0 |\n|---------------------------------------------------------------------------|\n| Active allocs | 856 | 856 | 856 | 0 |\n| from large pool | 29 | 29 | 29 | 0 |\n| from small pool | 827 | 827 | 827 | 0 |\n|---------------------------------------------------------------------------|\n| GPU reserved segments | 23 | 23 | 23 | 0 |\n| from large pool | 5 | 5 | 5 | 0 |\n| from small pool | 18 | 18 | 18 | 0 |\n|---------------------------------------------------------------------------|\n| Non-releasable allocs | 5 | 6 | 23 | 18 |\n| from large pool | 3 | 3 | 5 | 2 |\n| from small pool | 2 | 5 | 18 | 16 |\n|===========================================================================|\n\n" ] ], [ [ "# Training model", "_____no_output_____" ] ], [ [ "submissions = None\ntrain_results = []\n\nfor i_fold, (train_idx, valid_idx) in enumerate(folds.split(df_train, train_y)):\n # data preparation phase\n print(\"Fold {}/{}\".format(i_fold + 1, N_FOLDS))\n\n valid = df_train.iloc[valid_idx]\n valid.reset_index(drop=True, inplace=True)\n\n train = df_train.iloc[train_idx]\n train.reset_index(drop=True, inplace=True) \n\n # data transformation phase\n dataset_train = PlantDataset(df=train, transforms=transforms_train)\n dataset_valid = PlantDataset(df=valid, transforms=transforms_valid)\n\n # data loader phase\n dataloader_train = DataLoader(dataset_train, batch_size=BATCH_SIZE, num_workers=4, shuffle=True, pin_memory=True, drop_last=True)\n dataloader_valid = DataLoader(dataset_valid, batch_size=BATCH_SIZE, num_workers=4, shuffle=False, pin_memory=True, drop_last=False)\n\n\n # device = torch.device(\"cuda:0\") \n model = model.to(device)\n\n # optimization phase\n criterion = DenseCrossEntropy()\n\n # optimizer = optim.Adam(model.parameters(), lr=0.001)\n optimizer = AdamW(model.parameters(), lr = lr, weight_decay = 1e-3)\n\n # lr_scheduler = optim.lr_scheduler.MultiStepLR(optimizer=optimizer, milestones=[int(N_EPOCHS * 0.5), int(N_EPOCHS * 0.75)], gamma=0.1, last_epoch=-1)\n num_train_steps = int(len(dataset_train) / BATCH_SIZE * N_EPOCHS)\n lr_scheduler = get_cosine_schedule_with_warmup(optimizer, num_warmup_steps=len(dataset_train)/BATCH_SIZE*5, num_training_steps=num_train_steps)\n \n # training in one fold\n val_preds, train_fold_results = train_one_fold(i_fold, model, criterion, optimizer, lr_scheduler, dataloader_train, dataloader_valid)\n oof_preds[valid_idx, :] = val_preds\n \n # calculate the results phase\n train_results = train_results + train_fold_results\n\n # model evaluation phase\n model.eval()\n test_preds = None\n\n # looping test dataloader\n for step, batch in enumerate(dataloader_test):\n\n images = batch[0].to(device, dtype=torch.float)\n\n # empty torch gradient\n with torch.no_grad():\n outputs = model(images)\n\n if test_preds is None:\n test_preds = outputs.data.cpu()\n else:\n test_preds = torch.cat((test_preds, outputs.data.cpu()), dim=0)\n \n \n # Save predictions per fold\n submission_df[['healthy', 'multiple_diseases', 'rust', 'scab']] = torch.softmax(test_preds, dim=1)\n submission_df.to_csv('submission_fold_{}.csv'.format(i_fold), index=False)\n\n # logits avg\n if submissions is None:\n submissions = test_preds / N_FOLDS\n else:\n submissions += test_preds / N_FOLDS\n\nprint(\"5-Folds CV score: {:.4f}\".format(roc_auc_score(train_labels, oof_preds, average='macro')))\n\ntorch.save(model.state_dict(), '5-folds_rnn34.pth')", "Fold 1/5\n Epoch 1/20\n --------------------\nFold 0 Epoch 0: train_loss=1.3960, valid_loss=1.3966, acc=0.4153\n Epoch 2/20\n --------------------\n Epoch 3/20\n --------------------\nFold 0 Epoch 2: train_loss=0.6221, valid_loss=0.5551, acc=0.9469\n Epoch 4/20\n --------------------\n Epoch 5/20\n --------------------\nFold 0 Epoch 4: train_loss=0.4295, valid_loss=0.2986, acc=0.9803\n Epoch 6/20\n --------------------\n Epoch 7/20\n --------------------\nFold 0 Epoch 6: train_loss=0.4015, valid_loss=0.2740, acc=0.9827\n Epoch 8/20\n --------------------\n Epoch 9/20\n --------------------\nFold 0 Epoch 8: train_loss=0.3564, valid_loss=0.3162, acc=0.9792\n Epoch 10/20\n --------------------\n Epoch 11/20\n --------------------\nFold 0 Epoch 10: train_loss=0.3358, valid_loss=0.2532, acc=0.9865\n Epoch 12/20\n --------------------\n Epoch 13/20\n --------------------\nFold 0 Epoch 12: train_loss=0.3264, valid_loss=0.3874, acc=0.9745\n Epoch 14/20\n --------------------\n Epoch 15/20\n --------------------\nFold 0 Epoch 14: train_loss=0.3065, valid_loss=0.2948, acc=0.9848\n Epoch 16/20\n --------------------\n Epoch 17/20\n --------------------\nFold 0 Epoch 16: train_loss=0.3086, valid_loss=0.1817, acc=0.9924\n Epoch 18/20\n --------------------\n Epoch 19/20\n --------------------\nFold 0 Epoch 18: train_loss=0.2818, valid_loss=0.2035, acc=0.9908\n Epoch 20/20\n --------------------\nFold 2/5\n Epoch 1/20\n --------------------\nFold 1 Epoch 0: train_loss=0.2868, valid_loss=0.1407, acc=0.9951\n Epoch 2/20\n --------------------\n Epoch 3/20\n --------------------\nFold 1 Epoch 2: train_loss=0.2925, valid_loss=0.1157, acc=0.9971\n Epoch 4/20\n --------------------\n Epoch 5/20\n --------------------\nFold 1 Epoch 4: train_loss=0.2729, valid_loss=0.1732, acc=0.9940\n Epoch 6/20\n --------------------\n Epoch 7/20\n --------------------\nFold 1 Epoch 6: train_loss=0.2638, valid_loss=0.1526, acc=0.9946\n Epoch 8/20\n --------------------\n Epoch 9/20\n --------------------\nFold 1 Epoch 8: train_loss=0.2541, valid_loss=0.1448, acc=0.9962\n Epoch 10/20\n --------------------\n Epoch 11/20\n --------------------\nFold 1 Epoch 10: train_loss=0.2565, valid_loss=0.1886, acc=0.9926\n Epoch 12/20\n --------------------\n Epoch 13/20\n --------------------\nFold 1 Epoch 12: train_loss=0.2478, valid_loss=0.1564, acc=0.9940\n Epoch 14/20\n --------------------\n Epoch 15/20\n --------------------\nFold 1 Epoch 14: train_loss=0.2373, valid_loss=0.1130, acc=0.9972\n Epoch 16/20\n --------------------\n Epoch 17/20\n --------------------\nFold 1 Epoch 16: train_loss=0.2498, valid_loss=0.0987, acc=0.9980\n Epoch 18/20\n --------------------\n Epoch 19/20\n --------------------\nFold 1 Epoch 18: train_loss=0.2236, valid_loss=0.0953, acc=0.9970\n Epoch 20/20\n --------------------\nFold 3/5\n Epoch 1/20\n --------------------\nFold 2 Epoch 0: train_loss=0.2501, valid_loss=0.0587, acc=0.9994\n Epoch 2/20\n --------------------\n Epoch 3/20\n --------------------\nFold 2 Epoch 2: train_loss=0.2408, valid_loss=0.0460, acc=0.9996\n Epoch 4/20\n --------------------\n Epoch 5/20\n --------------------\nFold 2 Epoch 4: train_loss=0.1916, valid_loss=0.1718, acc=0.9949\n Epoch 6/20\n --------------------\n Epoch 7/20\n --------------------\nFold 2 Epoch 6: train_loss=0.2098, valid_loss=0.1658, acc=0.9948\n Epoch 8/20\n --------------------\n Epoch 9/20\n --------------------\nFold 2 Epoch 8: train_loss=0.2234, valid_loss=0.1086, acc=0.9979\n Epoch 10/20\n --------------------\n Epoch 11/20\n --------------------\nFold 2 Epoch 10: train_loss=0.2169, valid_loss=0.1105, acc=0.9980\n Epoch 12/20\n --------------------\n Epoch 13/20\n --------------------\nFold 2 Epoch 12: train_loss=0.2227, valid_loss=0.1422, acc=0.9944\n Epoch 14/20\n --------------------\n Epoch 15/20\n --------------------\nFold 2 Epoch 14: train_loss=0.1909, valid_loss=0.2184, acc=0.9904\n Epoch 16/20\n --------------------\n Epoch 17/20\n --------------------\nFold 2 Epoch 16: train_loss=0.2014, valid_loss=0.1684, acc=0.9932\n Epoch 18/20\n --------------------\n Epoch 19/20\n --------------------\nFold 2 Epoch 18: train_loss=0.1995, valid_loss=0.1158, acc=0.9969\n Epoch 20/20\n --------------------\nFold 4/5\n Epoch 1/20\n --------------------\nFold 3 Epoch 0: train_loss=0.2056, valid_loss=0.0563, acc=0.9995\n Epoch 2/20\n --------------------\n Epoch 3/20\n --------------------\nFold 3 Epoch 2: train_loss=0.1968, valid_loss=0.0775, acc=0.9990\n Epoch 4/20\n --------------------\n Epoch 5/20\n --------------------\nFold 3 Epoch 4: train_loss=0.1755, valid_loss=0.0208, acc=0.9999\n Epoch 6/20\n --------------------\n Epoch 7/20\n --------------------\nFold 3 Epoch 6: train_loss=0.1935, valid_loss=0.1251, acc=0.9973\n Epoch 8/20\n --------------------\n Epoch 9/20\n --------------------\nFold 3 Epoch 8: train_loss=0.1805, valid_loss=0.0968, acc=0.9978\n Epoch 10/20\n --------------------\n Epoch 11/20\n --------------------\nFold 3 Epoch 10: train_loss=0.1551, valid_loss=0.1007, acc=0.9984\n Epoch 12/20\n --------------------\n Epoch 13/20\n --------------------\nFold 3 Epoch 12: train_loss=0.1879, valid_loss=0.0521, acc=0.9995\n Epoch 14/20\n --------------------\n Epoch 15/20\n --------------------\nFold 3 Epoch 14: train_loss=0.1836, valid_loss=0.0457, acc=0.9996\n Epoch 16/20\n --------------------\n Epoch 17/20\n --------------------\nFold 3 Epoch 16: train_loss=0.1792, valid_loss=0.0370, acc=0.9996\n Epoch 18/20\n --------------------\n Epoch 19/20\n --------------------\nFold 3 Epoch 18: train_loss=0.1578, valid_loss=0.0396, acc=0.9995\n Epoch 20/20\n --------------------\nFold 5/5\n Epoch 1/20\n --------------------\nFold 4 Epoch 0: train_loss=0.1843, valid_loss=0.0132, acc=1.0000\n Epoch 2/20\n --------------------\n Epoch 3/20\n --------------------\nFold 4 Epoch 2: train_loss=0.1788, valid_loss=0.0543, acc=0.9980\n Epoch 4/20\n --------------------\n Epoch 5/20\n --------------------\nFold 4 Epoch 4: train_loss=0.1738, valid_loss=0.0494, acc=0.9985\n Epoch 6/20\n --------------------\n Epoch 7/20\n --------------------\nFold 4 Epoch 6: train_loss=0.1609, valid_loss=0.0091, acc=1.0000\n Epoch 8/20\n --------------------\n Epoch 9/20\n --------------------\nFold 4 Epoch 8: train_loss=0.1665, valid_loss=0.0816, acc=0.9988\n Epoch 10/20\n --------------------\n Epoch 11/20\n --------------------\nFold 4 Epoch 10: train_loss=0.1449, valid_loss=0.0572, acc=0.9995\n Epoch 12/20\n --------------------\n Epoch 13/20\n --------------------\nFold 4 Epoch 12: train_loss=0.1529, valid_loss=0.0107, acc=1.0000\n Epoch 14/20\n --------------------\n Epoch 15/20\n --------------------\nFold 4 Epoch 14: train_loss=0.1838, valid_loss=0.0219, acc=0.9999\n Epoch 16/20\n --------------------\n Epoch 17/20\n --------------------\nFold 4 Epoch 16: train_loss=0.1652, valid_loss=0.0569, acc=0.9970\n Epoch 18/20\n --------------------\n Epoch 19/20\n --------------------\nFold 4 Epoch 18: train_loss=0.1600, valid_loss=0.0187, acc=0.9999\n Epoch 20/20\n --------------------\n5-Folds CV score: 0.9922\n" ] ], [ [ "# Generate training results", "_____no_output_____" ] ], [ [ "train_results = pd.DataFrame(train_results)\ntrain_results.head(10)", "_____no_output_____" ], [ "train_results.to_csv('train_result.csv')", "_____no_output_____" ] ], [ [ "# Plotting results", "_____no_output_____" ], [ "## Training loss", "_____no_output_____" ] ], [ [ "def show_training_loss(train_result):\n plt.figure(figsize=(15,10))\n plt.subplot(3,1,1)\n train_loss = train_result['train_loss']\n plt.plot(train_loss.index, train_loss, label = 'train_loss')\n plt.legend()\n\n val_loss = train_result['valid_loss']\n plt.plot(val_loss.index, val_loss, label = 'val_loss')\n plt.legend()", "_____no_output_____" ], [ "show_training_loss(train_results)", "_____no_output_____" ] ], [ [ "## Validation score", "_____no_output_____" ] ], [ [ "def show_valid_score(train_result):\n plt.figure(figsize=(15,10))\n plt.subplot(3,1,1)\n valid_score = train_result['valid_score']\n plt.plot(valid_score.index, valid_score, label = 'valid_score')\n plt.legend()", "_____no_output_____" ], [ "show_valid_score(train_results)", "_____no_output_____" ], [ "submission_df[['healthy', 'multiple_diseases', 'rust', 'scab']] = torch.softmax(submissions, dim=1)\nsubmission_df.to_csv('submission.csv', index=False)\nsubmission_df.head()", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code" ] ]

4ad7ae2d215eac5337584db95f321ae6a471ca91

4,213

ipynb

Jupyter Notebook

Machine Learning/Problem3/1_Simple_Neural_Network.ipynb

bayeslabs/AiGym

30c126fc2e140f9f164ff3f20638242b230e7e52

[ "MIT" ]

22

2019-07-15T08:26:31.000Z

2022-01-17T06:29:17.000Z

Machine Learning/Problem3/1_Simple_Neural_Network.ipynb

bayeslabs/AiGym

30c126fc2e140f9f164ff3f20638242b230e7e52

[ "MIT" ]

26

2020-03-24T17:18:21.000Z

2022-03-11T23:54:37.000Z

Machine Learning/Problem3/1_Simple_Neural_Network.ipynb

bayeslabs/AiGym

30c126fc2e140f9f164ff3f20638242b230e7e52

[ "MIT" ]

8

2019-07-17T09:13:11.000Z

2021-04-16T11:20:51.000Z

34.252033

449

0.57465

[ [ [ "# CS229: Problem Set 3\n## Problem 1: A Simple Neural Network\n\n\n**C. Combier**\n\nThis iPython Notebook provides solutions to Stanford's CS229 (Machine Learning, Fall 2017) graduate course problem set 3, taught by Andrew Ng.\n\nThe problem set can be found here: [./ps3.pdf](ps3.pdf)\n\nI chose to write the solutions to the coding questions in Python, whereas the Stanford class is taught with Matlab/Octave.\n\n## Notation\n\n- $x^i$ is the $i^{th}$ feature vector\n- $y^i$ is the expected outcome for the $i^{th}$ training example\n- $m$ is the number of training examples\n- $n$ is the number of features", "_____no_output_____" ], [ "### Question 1.b)\n\n![triangle separation](data/triangle_pb3_1.jpg)\n\nIt seems that a triangle can separate the data.\n\nWe can construct a weight matrix by using a combination of linear classifiers, where each side of the triangle represents a decision boundary.\n\nEach side of the triangle can be represented by an equation of the form $w_0 +w_1 x_1 + w_2 x_2 = 0$. If we transform this equality into an inequality, then the output represents on which side of the decision boundary a given data point $(x_1,x_2)$ belongs. The intersection of the outputs for each of these decision boundaries tells us whether $(x_1,x_2)$ lies within the triangle, in which case we will classify it $0$, and if not as $1$.\n\nThe first weight matrix can be written as:\n\n", "_____no_output_____" ], [ "$$\nW^{[1]} = \\left ( \\begin{array}{ccc}\n-1 & 4 & 0 \\\\\n-1 & 0 & 4 \\\\\n4.5 & -1 & -1\n\\end{array} \\right )\n$$\n\nThe input vector is:\n$$\nX = (\\begin{array}{ccc}\n1 & x_1 & x_2\n\\end{array})^T\n$$\n\n- The first line of $W^{[1]}$ is the equation for the vertical side of the triangle, $x_1 = 0.25$\n- The second line of $W^{[1]}$ is the equation for the horizontal side of the triangle, $x_2 = 0.25$\n- The third line of $W^{[1]}$ is the equation for the oblique side of the triangle, $x_2 = -x_1 + 4.5$\n\nConsequently, with the given activation function, if the training example given by ($x_1$, $x_2$) lies within the triangle, then:\n\n$$\nf(W^{[1]}X) = (\\begin{array}{ccc}\n1 & 1 & 1\n\\end{array})^T\n$$\n\nIn all other cases, at least one element of the output vector $f(W^{[1]}X)$ is not equal to $1$.\n\nWe can use this observation to find weights for the ouput layer. We take the sum of the components of $f(W^{[1]}X)$, and compare the value to 2.5 to check if all elements are equal to $1$ or not. This gives the weight matrix:\n\n$$\nW^{[2]} =(\\begin{array}{cccc}\n2.5 & -1 & -1 & -1\n\\end{array})\n$$\n\nThe additional term 2.5 is the zero intercept. With this weight matrix, the ouput of the final layer will be $0$ if the training example is within the triangle, and $1$ if it is outside of the triangle.\n\nThe ", "_____no_output_____" ], [ "### Question 1.c)\n\nA linear activation function does not work, because the problem is not linearly separable, i.e. there is no hyperplane that perfectly separates the data.", "_____no_output_____" ] ] ]

[ "markdown" ]

[ [ "markdown", "markdown", "markdown", "markdown" ] ]

4ad7b54308cdd33272dd3de1ff18efd56ceaf349

16,650

ipynb

Jupyter Notebook

01-Lesson-Plans/23-Project-4/1/Activities/01-Evr_MNIST/Unsolved/MNIST.ipynb

anirudhmungre/sneaky-lessons

8e48015c50865059db96f8cd369bcc15365d66c7

[ "ADSL" ]

null

01-Lesson-Plans/23-Project-4/1/Activities/01-Evr_MNIST/Unsolved/MNIST.ipynb

anirudhmungre/sneaky-lessons

8e48015c50865059db96f8cd369bcc15365d66c7

[ "ADSL" ]

null

01-Lesson-Plans/23-Project-4/1/Activities/01-Evr_MNIST/Unsolved/MNIST.ipynb

anirudhmungre/sneaky-lessons

8e48015c50865059db96f8cd369bcc15365d66c7

[ "ADSL" ]

null

20.734745

183

0.54012

[ [ [ "## Dependencies", "_____no_output_____" ] ], [ [ "# Dependencies to Visualize the model\n%matplotlib inline\nfrom IPython.display import Image, SVG\nimport matplotlib.pyplot as plt\nimport numpy as np\nnp.random.seed(0)", "_____no_output_____" ], [ "# Filepaths, numpy, and Tensorflow\nimport os\nimport numpy as np\nimport tensorflow as tf", "_____no_output_____" ], [ "# Sklearn scaling\nfrom sklearn.preprocessing import MinMaxScaler", "_____no_output_____" ] ], [ [ "### Keras Specific Dependencies", "_____no_output_____" ] ], [ [ "# Keras\nfrom tensorflow import keras\nfrom tensorflow.keras.models import Sequential\nfrom tensorflow.keras.utils import to_categorical\nfrom tensorflow.keras.layers import Dense\nfrom tensorflow.keras.datasets import mnist", "_____no_output_____" ] ], [ [ "## Loading and Preprocessing our Data", "_____no_output_____" ], [ "### Load the MNIST Handwriting Dataset from Keras", "_____no_output_____" ] ], [ [ "(X_train, y_train), (X_test, y_test) = mnist.load_data()\nprint(\"Training Data Info\")\nprint(\"Training Data Shape:\", X_train.shape)\nprint(\"Training Data Labels Shape:\", y_train.shape)", "_____no_output_____" ] ], [ [ "### Plot the first digit", "_____no_output_____" ] ], [ [ "# Plot the first image from the dataset\nplt.imshow(X_train[0,:,:], cmap=plt.cm.Greys)", "_____no_output_____" ] ], [ [ "### Each Image is a 28x28 Pixel greyscale image with values from 0 to 255", "_____no_output_____" ] ], [ [ "# Our image is an array of pixels ranging from 0 to 255\nX_train[0, :, :]", "_____no_output_____" ] ], [ [ "### For training a model, we want to flatten our data into rows of 1D image arrays", "_____no_output_____" ] ], [ [ "# We want to flatten our image of 28x28 pixels to a 1D array of 784 pixels\nndims = X_train.shape[1] * X_train.shape[2]\nX_train = X_train.reshape(X_train.shape[0], ndims)\nX_test = X_test.reshape(X_test.shape[0], ndims)\nprint(\"Training Shape:\", X_train.shape)\nprint(\"Testing Shape:\", X_test.shape)", "_____no_output_____" ] ], [ [ "## Scaling and Normalization\n\nWe use Sklearn's MinMaxScaler to normalize our data between 0 and 1", "_____no_output_____" ] ], [ [ "# Next, we normalize our training data to be between 0 and 1\nscaler = MinMaxScaler().fit(X_train)\n\nX_train = scaler.transform(X_train)\nX_test = scaler.transform(X_test)", "_____no_output_____" ], [ "# Alternative way to normalize this dataset since we know that the max pixel value is 255\n# X_train = X_train.astype(\"float32\")\n# X_test = X_test.astype(\"float32\")\n# X_train /= 255.0\n# X_test /= 255.0", "_____no_output_____" ] ], [ [ "## One-Hot Encoding\n\nWe need to one-hot encode our integer labels using the `to_categorical` helper function", "_____no_output_____" ] ], [ [ "# Our Training and Testing labels are integer encoded from 0 to 9\ny_train[:20]", "_____no_output_____" ], [ "# We need to convert our target labels (expected values) to categorical data\nnum_classes = 10\ny_train = to_categorical(y_train, num_classes)\ny_test = to_categorical(y_test, num_classes)\n# Original label of `5` is one-hot encoded as `0000010000`\ny_train[0]", "_____no_output_____" ] ], [ [ "## Building our Model\n\nIn this example, we are going to build a Deep Multi-Layer Perceptron model with 2 hidden layers.", "_____no_output_____" ], [ "## Our first step is to create an empty sequential model", "_____no_output_____" ] ], [ [ "# Create an empty sequential model\n", "_____no_output_____" ] ], [ [ "## Next, we add our first hidden layer\n\nIn the first hidden layer, we must also specify the dimension of our input layer. This will simply be the number of elements (pixels) in each image.", "_____no_output_____" ] ], [ [ "# Add the first layer where the input dimensions are the 784 pixel values\n# We can also choose our activation function. `relu` is a common\n", "_____no_output_____" ] ], [ [ "## We then add a second hidden layer with 100 densely connected nodes\n\nA dense layer is when every node from the previous layer is connected to each node in the current layer.", "_____no_output_____" ] ], [ [ "# Add a second hidden layer\n", "_____no_output_____" ] ], [ [ "## Our final output layer uses a `softmax` activation function for logistic regression.\n\nWe also need to specify the number of output classes. In this case, the number of digits that we wish to classify.", "_____no_output_____" ] ], [ [ "# Add our final output layer where the number of nodes \n# corresponds to the number of y labels\n", "_____no_output_____" ] ], [ [ "## Model Summary", "_____no_output_____" ] ], [ [ "# We can summarize our model\nmodel.summary()", "_____no_output_____" ] ], [ [ "## Compile and Train our Model\n\nNow that we have our model architecture defined, we must compile the model using a loss function and optimizer. We can also specify additional training metrics such as accuracy.", "_____no_output_____" ] ], [ [ "# Compile the model\n", "_____no_output_____" ] ], [ [ "## Finally, we train our model using our training data", "_____no_output_____" ], [ "Training consists of updating our weights using our optimizer and loss function. In this example, we choose 10 iterations (loops) of training that are called epochs.\n\nWe also choose to shuffle our training data and increase the detail printed out during each training cycle.", "_____no_output_____" ] ], [ [ "# Fit (train) the model\n", "_____no_output_____" ] ], [ [ "## Saving and Loading models\n\nWe can save our trained models using the HDF5 binary format with the extension `.h5`", "_____no_output_____" ] ], [ [ "# Save the model\n", "_____no_output_____" ], [ "# Load the model\n", "_____no_output_____" ] ], [ [ "## Evaluating the Model\n\nWe use our testing data to validate our model. This is how we determine the validity of our model (i.e. the ability to predict new and previously unseen data points)", "_____no_output_____" ] ], [ [ "# Evaluate the model using the training data \nmodel_loss, model_accuracy = model.evaluate(X_test, y_test, verbose=2)\nprint(f\"Loss: {model_loss}, Accuracy: {model_accuracy}\")", "_____no_output_____" ] ], [ [ "## Making Predictions\n\nWe can use our trained model to make predictions using `model.predict`", "_____no_output_____" ] ], [ [ "# Grab just one data point to test with\ntest = np.expand_dims(X_train[0], axis=0)\ntest.shape", "_____no_output_____" ], [ "plt.imshow(scaler.inverse_transform(test).reshape(28, 28), cmap=plt.cm.Greys)", "_____no_output_____" ], [ "# Make a prediction. The result should be 0000010000000 for a 5\nmodel.predict(test).round()", "_____no_output_____" ], [ "# Grab just one data point to test with\ntest = np.expand_dims(X_train[2], axis=0)\ntest.shape", "_____no_output_____" ], [ "plt.imshow(scaler.inverse_transform(test).reshape(28, 28), cmap=plt.cm.Greys)", "_____no_output_____" ], [ "# Make a prediction. The resulting class should match the digit\nprint(f\"One-Hot-Encoded Prediction: {model.predict(test).round()}\")\nprint(f\"Predicted class: {model.predict_classes(test)}\")", "_____no_output_____" ] ], [ [ "# Import a Custom Image", "_____no_output_____" ] ], [ [ "filepath = \"../Images/test8.png\"", "_____no_output_____" ], [ "# Import the image using the `load_img` function in keras preprocessing\n", "_____no_output_____" ], [ "# Convert the image to a numpy array \n", "_____no_output_____" ], [ "# Scale the image pixels by 255 (or use a scaler from sklearn here)", "_____no_output_____" ], [ "# Flatten into a 1x28*28 array \n", "_____no_output_____" ], [ "plt.imshow(img.reshape(28, 28), cmap=plt.cm.Greys)", "_____no_output_____" ], [ "# Invert the pixel values to match the original data\n", "_____no_output_____" ], [ "# Make predictions\nmodel.predict_classes(img)", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad7cc2b42f69bc76b541432a0376f2ae4c98a28

262,137

ipynb

Jupyter Notebook

P1.ipynb

Emad2018/CarND-LaneLines-P1

c4eb2d0c8f12144879bb7d4910b8a7cc8324c778

[ "MIT" ]

null

P1.ipynb

Emad2018/CarND-LaneLines-P1

c4eb2d0c8f12144879bb7d4910b8a7cc8324c778

[ "MIT" ]

null

P1.ipynb

Emad2018/CarND-LaneLines-P1

c4eb2d0c8f12144879bb7d4910b8a7cc8324c778

[ "MIT" ]

null

348.123506

117,828

0.926687

[ [ [ "# Self-Driving Car Engineer Nanodegree\n\n\n## Project: **Finding Lane Lines on the Road** \n***\nIn this project, you will use the tools you learned about in the lesson to identify lane lines on the road. You can develop your pipeline on a series of individual images, and later apply the result to a video stream (really just a series of images). Check out the video clip \"raw-lines-example.mp4\" (also contained in this repository) to see what the output should look like after using the helper functions below. \n\nOnce you have a result that looks roughly like \"raw-lines-example.mp4\", you'll need to get creative and try to average and/or extrapolate the line segments you've detected to map out the full extent of the lane lines. You can see an example of the result you're going for in the video \"P1_example.mp4\". Ultimately, you would like to draw just one line for the left side of the lane, and one for the right.\n\nIn addition to implementing code, there is a brief writeup to complete. The writeup should be completed in a separate file, which can be either a markdown file or a pdf document. There is a [write up template](https://github.com/udacity/CarND-LaneLines-P1/blob/master/writeup_template.md) that can be used to guide the writing process. Completing both the code in the Ipython notebook and the writeup template will cover all of the [rubric points](https://review.udacity.com/#!/rubrics/322/view) for this project.\n\n---\nLet's have a look at our first image called 'test_images/solidWhiteRight.jpg'. Run the 2 cells below (hit Shift-Enter or the \"play\" button above) to display the image.\n\n**Note: If, at any point, you encounter frozen display windows or other confounding issues, you can always start again with a clean slate by going to the \"Kernel\" menu above and selecting \"Restart & Clear Output\".**\n\n---", "_____no_output_____" ], [ "**The tools you have are color selection, region of interest selection, grayscaling, Gaussian smoothing, Canny Edge Detection and Hough Tranform line detection. You are also free to explore and try other techniques that were not presented in the lesson. Your goal is piece together a pipeline to detect the line segments in the image, then average/extrapolate them and draw them onto the image for display (as below). Once you have a working pipeline, try it out on the video stream below.**\n\n---\n\n<figure>\n <img src=\"examples/line-segments-example.jpg\" width=\"380\" alt=\"Combined Image\" />\n <figcaption>\n \n Your output should look something like this (above) after detecting line segments using the helper functions below \n </figcaption>\n</figure>\n \n<figure>\n <img src=\"examples/laneLines_thirdPass.jpg\" width=\"380\" alt=\"Combined Image\" />\n <figcaption>\n \n Your goal is to connect/average/extrapolate line segments to get output like this \n </figcaption>\n</figure>", "_____no_output_____" ], [ "**Run the cell below to import some packages. If you get an `import error` for a package you've already installed, try changing your kernel (select the Kernel menu above --> Change Kernel). Still have problems? Try relaunching Jupyter Notebook from the terminal prompt. Also, consult the forums for more troubleshooting tips.** ", "_____no_output_____" ], [ "## Import Packages", "_____no_output_____" ] ], [ [ "#importing some useful packages\nimport matplotlib.pyplot as plt\nimport matplotlib.image as mpimg\nimport numpy as np\nimport cv2\n%matplotlib inline", "_____no_output_____" ] ], [ [ "## Read in an Image", "_____no_output_____" ] ], [ [ "#reading in an image\nimage = mpimg.imread('test_images/solidWhiteRight.jpg')\n\n#printing out some stats and plotting\nprint('This image is:', type(image), 'with dimensions:', image.shape)\nplt.imshow(image) # if you wanted to show a single color channel image called 'gray', for example, call as plt.imshow(gray, cmap='gray')", "This image is: <class 'numpy.ndarray'> with dimensions: (540, 960, 3)\n" ] ], [ [ "## Ideas for Lane Detection Pipeline", "_____no_output_____" ], [ "**Some OpenCV functions (beyond those introduced in the lesson) that might be useful for this project are:**\n\n`cv2.inRange()` for color selection \n`cv2.fillPoly()` for regions selection \n`cv2.line()` to draw lines on an image given endpoints \n`cv2.addWeighted()` to coadd / overlay two images\n`cv2.cvtColor()` to grayscale or change color\n`cv2.imwrite()` to output images to file \n`cv2.bitwise_and()` to apply a mask to an image\n\n**Check out the OpenCV documentation to learn about these and discover even more awesome functionality!**", "_____no_output_____" ], [ "## Helper Functions", "_____no_output_____" ], [ "Below are some helper functions to help get you started. They should look familiar from the lesson!", "_____no_output_____" ] ], [ [ "import math\n\ndef grayscale(img):\n \"\"\"Applies the Grayscale transform\n This will return an image with only one color channel\n but NOTE: to see the returned image as grayscale\n (assuming your grayscaled image is called 'gray')\n you should call plt.imshow(gray, cmap='gray')\"\"\"\n return cv2.cvtColor(img, cv2.COLOR_RGB2GRAY)\n # Or use BGR2GRAY if you read an image with cv2.imread()\n # return cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)\n \ndef canny(img, low_threshold, high_threshold):\n \"\"\"Applies the Canny transform\"\"\"\n return cv2.Canny(img, low_threshold, high_threshold)\n\ndef gaussian_blur(img, kernel_size):\n \"\"\"Applies a Gaussian Noise kernel\"\"\"\n return cv2.GaussianBlur(img, (kernel_size, kernel_size), 0)\n\ndef region_of_interest(img, vertices):\n \"\"\"\n Applies an image mask.\n \n Only keeps the region of the image defined by the polygon\n formed from `vertices`. The rest of the image is set to black.\n `vertices` should be a numpy array of integer points.\n \"\"\"\n #defining a blank mask to start with\n mask = np.zeros_like(img) \n \n #defining a 3 channel or 1 channel color to fill the mask with depending on the input image\n if len(img.shape) > 2:\n channel_count = img.shape[2] # i.e. 3 or 4 depending on your image\n ignore_mask_color = (255,) * channel_count\n else:\n ignore_mask_color = 255\n \n #filling pixels inside the polygon defined by \"vertices\" with the fill color \n cv2.fillPoly(mask, vertices, ignore_mask_color)\n \n #returning the image only where mask pixels are nonzero\n masked_image = cv2.bitwise_and(img, mask)\n return masked_image\n\n\ndef draw_lines(img, lines, color=[255, 0, 0], thickness=8):\n \"\"\"\n NOTE: this is the function you might want to use as a starting point once you want to \n average/extrapolate the line segments you detect to map out the full\n extent of the lane (going from the result shown in raw-lines-example.mp4\n to that shown in P1_example.mp4). \n \n Think about things like separating line segments by their \n slope ((y2-y1)/(x2-x1)) to decide which segments are part of the left\n line vs. the right line. Then, you can average the position of each of \n the lines and extrapolate to the top and bottom of the lane.\n \n This function draws `lines` with `color` and `thickness`. \n Lines are drawn on the image inplace (mutates the image).\n If you want to make the lines semi-transparent, think about combining\n this function with the weighted_img() function below\n \"\"\"\n \n left={\"x\":[],\"y\":[]}\n right={\"x\":[],\"y\":[]}\n midpoint=img.shape[1]//2\n for line in lines:\n for x1,y1,x2,y2 in line:\n slope=((y2-y1)/(x2-x1))\n if ( x1>midpoint and x2>midpoint):\n right[\"x\"].append(x1)\n right[\"x\"].append(x2)\n right[\"y\"].append(y1)\n right[\"y\"].append(y2)\n #cv2.line(img, (x1, y1), (x2, y2), color, thickness)\n elif( x1<midpoint and x2<midpoint):\n left[\"x\"].append(x1)\n left[\"x\"].append(x2)\n left[\"y\"].append(y1)\n left[\"y\"].append(y2)\n #cv2.line(img, (x1, y1), (x2, y2), [0, 255, 0], thickness)\n if(len(right[\"x\"])>5 and len(right[\"y\"])>5 ):\n right_fit = np.polyfit(right[\"x\"],right[\"y\"], 1)\n for i in range(len(right[\"x\"])-1):\n x1=image.shape[1]\n y1=int(right_fit[0]*x1+right_fit[1])\n x2=right[\"x\"][i+1]\n y2=int(right_fit[0]*x2+right_fit[1])\n cv2.line(img, (x1,y1), (x2, y2), color, thickness)\n if(len(left[\"x\"])>5 and len(left[\"y\"])>5 ): \n left_fit = np.polyfit(left[\"x\"],left[\"y\"], 1)\n for i in range(len(left[\"x\"])-1): \n x1=0\n y1=int(left_fit[0]*x1+left_fit[1])\n x2=left[\"x\"][i+1]\n y2=int(left_fit[0]*x2+left_fit[1])\n cv2.line(img, (x1,y1), (x2, y2), color, thickness)\n\n \n \ndef hough_lines(img, rho, theta, threshold, min_line_len, max_line_gap):\n \"\"\"\n `img` should be the output of a Canny transform.\n \n Returns an image with hough lines drawn.\n \"\"\"\n lines = cv2.HoughLinesP(img, rho, theta, threshold, np.array([]), minLineLength=min_line_len, maxLineGap=max_line_gap)\n line_img = np.zeros((img.shape[0], img.shape[1], 3), dtype=np.uint8)\n draw_lines(line_img, lines)\n return line_img\n\n# Python 3 has support for cool math symbols.\n\ndef weighted_img(img, initial_img, α=0.8, β=1., γ=0.):\n \"\"\"\n `img` is the output of the hough_lines(), An image with lines drawn on it.\n Should be a blank image (all black) with lines drawn on it.\n \n `initial_img` should be the image before any processing.\n \n The result image is computed as follows:\n \n initial_img * α + img * β + γ\n NOTE: initial_img and img must be the same shape!\n \"\"\"\n return cv2.addWeighted(initial_img, α, img, β, γ)", "_____no_output_____" ] ], [ [ "## Test Images\n\nBuild your pipeline to work on the images in the directory \"test_images\" \n**You should make sure your pipeline works well on these images before you try the videos.**", "_____no_output_____" ] ], [ [ "import os\nos.listdir(\"test_images/\")\ntest_dir=\"test_images/\"\ntest_images=os.listdir(test_dir)", "_____no_output_____" ] ], [ [ "## Build a Lane Finding Pipeline\n\n", "_____no_output_____" ], [ "Build the pipeline and run your solution on all test_images. Make copies into the `test_images_output` directory, and you can use the images in your writeup report.\n\nTry tuning the various parameters, especially the low and high Canny thresholds as well as the Hough lines parameters.", "_____no_output_____" ] ], [ [ "# TODO: Build your pipeline that will draw lane lines on the test_images\n# then save them to the test_images_output directory.\ntest_img_index=2\nimage = mpimg.imread(test_dir+test_images[test_img_index])\ngray_img=grayscale(image)\nblured=gaussian_blur(gray_img, 5)\ncanny_img=canny(blured, 70, 140)\nimshape=canny_img.shape\nvertices = np.array([[(110,imshape[0]),(950,imshape[0] ), (530, 320), (440,320)]], dtype=np.int32)\nROI_img=region_of_interest(canny_img,vertices)\nlines=hough_lines(ROI_img, 1, np.pi/180, 1, 5, 1)\noutput=weighted_img(image, lines, α=0.8, β=1., γ=0.)\nvertices = vertices.reshape((-1,1,2))\ncv2.polylines(output,[vertices],True,(0,255,255))\nplt.imshow(ROI_img,cmap=\"gray\")\nplt.imshow(output)", "_____no_output_____" ] ], [ [ "## Test on Videos\n\nYou know what's cooler than drawing lanes over images? Drawing lanes over video!\n\nWe can test our solution on two provided videos:\n\n`solidWhiteRight.mp4`\n\n`solidYellowLeft.mp4`\n\n**Note: if you get an import error when you run the next cell, try changing your kernel (select the Kernel menu above --> Change Kernel). Still have problems? Try relaunching Jupyter Notebook from the terminal prompt. Also, consult the forums for more troubleshooting tips.**\n\n**If you get an error that looks like this:**\n```\nNeedDownloadError: Need ffmpeg exe. \nYou can download it by calling: \nimageio.plugins.ffmpeg.download()\n```\n**Follow the instructions in the error message and check out [this forum post](https://discussions.udacity.com/t/project-error-of-test-on-videos/274082) for more troubleshooting tips across operating systems.**", "_____no_output_____" ] ], [ [ "# Import everything needed to edit/save/watch video clips\nfrom moviepy.editor import VideoFileClip\nfrom IPython.display import HTML", "_____no_output_____" ], [ "def process_image(image):\n # NOTE: The output you return should be a color image (3 channel) for processing video below\n # TODO: put your pipeline here,\n # you should return the final output (image where lines are drawn on lanes)\n gray_img=grayscale(image)\n blured=gaussian_blur(gray_img, 5)\n canny_img=canny(blured, 70, 140)\n imshape=canny_img.shape\n vertices = np.array([[(110,imshape[0]),(950,imshape[0] ), (530, 320), (440,320)]], dtype=np.int32)\n ROI_img=region_of_interest(canny_img,vertices)\n lines=hough_lines(ROI_img, 1, np.pi/180, 1, 5, 1)\n result=weighted_img(image, lines, α=0.8, β=1., γ=0.)\n return result", "_____no_output_____" ] ], [ [ "Let's try the one with the solid white lane on the right first ...", "_____no_output_____" ] ], [ [ "white_output = 'test_videos_output/solidWhiteRight.mp4'\n## To speed up the testing process you may want to try your pipeline on a shorter subclip of the video\n## To do so add .subclip(start_second,end_second) to the end of the line below\n## Where start_second and end_second are integer values representing the start and end of the subclip\n## You may also uncomment the following line for a subclip of the first 5 seconds\n##clip1 = VideoFileClip(\"test_videos/solidWhiteRight.mp4\").subclip(0,5)\nclip1 = VideoFileClip(\"test_videos/solidWhiteRight.mp4\")\nwhite_clip = clip1.fl_image(process_image) #NOTE: this function expects color images!!\n%time white_clip.write_videofile(white_output, audio=False)", "[MoviePy] >>>> Building video test_videos_output/solidWhiteRight.mp4\n[MoviePy] Writing video test_videos_output/solidWhiteRight.mp4\n" ] ], [ [ "Play the video inline, or if you prefer find the video in your filesystem (should be in the same directory) and play it in your video player of choice.", "_____no_output_____" ] ], [ [ "HTML(\"\"\"\n<video width=\"960\" height=\"540\" controls>\n <source src=\"{0}\">\n</video>\n\"\"\".format(white_output))", "_____no_output_____" ] ], [ [ "\n\n## Improve the draw_lines() function\n\n**At this point, if you were successful with making the pipeline and tuning parameters, you probably have the Hough line segments drawn onto the road, but what about identifying the full extent of the lane and marking it clearly as in the example video (P1_example.mp4)? Think about defining a line to run the full length of the visible lane based on the line segments you identified with the Hough Transform. As mentioned previously, try to average and/or extrapolate the line segments you've detected to map out the full extent of the lane lines. You can see an example of the result you're going for in the video \"P1_example.mp4\".**\n\n**Go back and modify your draw_lines function accordingly and try re-running your pipeline. The new output should draw a single, solid line over the left lane line and a single, solid line over the right lane line. The lines should start from the bottom of the image and extend out to the top of the region of interest.**", "_____no_output_____" ], [ "Now for the one with the solid yellow lane on the left. This one's more tricky!", "_____no_output_____" ] ], [ [ "yellow_output = 'test_videos_output/solidYellowLeft.mp4'\n## To speed up the testing process you may want to try your pipeline on a shorter subclip of the video\n## To do so add .subclip(start_second,end_second) to the end of the line below\n## Where start_second and end_second are integer values representing the start and end of the subclip\n## You may also uncomment the following line for a subclip of the first 5 seconds\n##clip2 = VideoFileClip('test_videos/solidYellowLeft.mp4').subclip(0,5)\nclip2 = VideoFileClip('test_videos/solidYellowLeft.mp4')\nyellow_clip = clip2.fl_image(process_image)\n%time yellow_clip.write_videofile(yellow_output, audio=False)", "[MoviePy] >>>> Building video test_videos_output/solidYellowLeft.mp4\n[MoviePy] Writing video test_videos_output/solidYellowLeft.mp4\n" ], [ "HTML(\"\"\"\n<video width=\"960\" height=\"540\" controls>\n <source src=\"{0}\">\n</video>\n\"\"\".format(yellow_output))", "_____no_output_____" ] ], [ [ "## Writeup and Submission\n\nIf you're satisfied with your video outputs, it's time to make the report writeup in a pdf or markdown file. Once you have this Ipython notebook ready along with the writeup, it's time to submit for review! Here is a [link](https://github.com/udacity/CarND-LaneLines-P1/blob/master/writeup_template.md) to the writeup template file.\n", "_____no_output_____" ], [ "## Optional Challenge\n\nTry your lane finding pipeline on the video below. Does it still work? Can you figure out a way to make it more robust? If you're up for the challenge, modify your pipeline so it works with this video and submit it along with the rest of your project!", "_____no_output_____" ] ], [ [ "challenge_output = 'test_videos_output/challenge.mp4'\n## To speed up the testing process you may want to try your pipeline on a shorter subclip of the video\n## To do so add .subclip(start_second,end_second) to the end of the line below\n## Where start_second and end_second are integer values representing the start and end of the subclip\n## You may also uncomment the following line for a subclip of the first 5 seconds\n##clip3 = VideoFileClip('test_videos/challenge.mp4').subclip(0,5)\nclip3 = VideoFileClip('test_videos/challenge.mp4')\nchallenge_clip = clip3.fl_image(process_image)\n%time challenge_clip.write_videofile(challenge_output, audio=False)", "[MoviePy] >>>> Building video test_videos_output/challenge.mp4\n[MoviePy] Writing video test_videos_output/challenge.mp4\n" ], [ "HTML(\"\"\"\n<video width=\"960\" height=\"540\" controls>\n <source src=\"{0}\">\n</video>\n\"\"\".format(challenge_output))", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code", "code" ] ]

4ad7e73bb1fec38b5d64f252b1e66824bec93950

49,529

ipynb

Jupyter Notebook

notebooks/08-Summarizing-Data.ipynb

janetzahner/intro-python-datasci

54d4e686d873b4e8f8c251846a1132aad7b52d1e

[ "MIT" ]

1

2021-11-01T05:14:45.000Z

notebooks/08-Summarizing-Data.ipynb

janetzahner/intro-python-datasci

54d4e686d873b4e8f8c251846a1132aad7b52d1e

[ "MIT" ]

null

notebooks/08-Summarizing-Data.ipynb

janetzahner/intro-python-datasci

54d4e686d873b4e8f8c251846a1132aad7b52d1e

[ "MIT" ]

1

2021-10-30T06:48:48.000Z

26.10912

196

0.422036

[ [ [ "# Summarizing Data\n\n> What we have is a data glut.\n>\n> \\- Vernor Vinge, Professor Emeritus of Mathematics, San Diego State University", "_____no_output_____" ], [ "## Applied Review", "_____no_output_____" ], [ "### Dictionaries", "_____no_output_____" ], [ "* The `dict` structure is used to represent **key-value pairs**", "_____no_output_____" ], [ "* Like a real dictionary, you look up a word (**key**) and get its definition (**value**)", "_____no_output_____" ], [ "* Below is an example:\n\n```python\nethan = {\n 'first_name': 'Ethan',\n 'last_name': 'Swan',\n 'alma_mater': 'Notre Dame',\n 'employer': '84.51˚',\n 'zip_code': 45208\n}\n```", "_____no_output_____" ], [ "### DataFrame Structure", "_____no_output_____" ], [ "* We will start by importing the `flights` data set as a DataFrame:", "_____no_output_____" ] ], [ [ "import pandas as pd\nflights_df = pd.read_csv('../data/flights.csv')", "_____no_output_____" ] ], [ [ "* Each DataFrame variable is a **Series** and can be accessed with bracket subsetting notation: `DataFrame['SeriesName']`", "_____no_output_____" ], [ "* The DataFrame has an **Index** that is visible the far left side and can be used to slide the DataFrame", "_____no_output_____" ], [ "### Methods", "_____no_output_____" ], [ "* Methods are *operations* that are specific to Python classes", "_____no_output_____" ], [ "* These operations end in parentheses and *make something happen*", "_____no_output_____" ], [ "* An example of a method is `DataFrame.head()`", "_____no_output_____" ], [ "## General Model", "_____no_output_____" ], [ "### Window Operations", "_____no_output_____" ], [ "Yesterday we learned how to manipulate data across one or more variables within the row(s):\n\n![series-plus-series.png](images/series-plus-series.png)", "_____no_output_____" ], [ "Note that we return the same number of elements that we started with. This is known as a **window function**, but you can also think of it as summarizing at the row-level.", "_____no_output_____" ], [ "We could achieve this result with the following code:\n\n```python\nDataFrame['A'] + DataFrame['B']\n```", "_____no_output_____" ], [ "We subset the two Series and then add them together using the `+` operator to achieve the sum.", "_____no_output_____" ], [ "Note that we could also use some other operation on `DataFrame['B']` as long as it returns the same number of elements.", "_____no_output_____" ], [ "### Summary Operations", "_____no_output_____" ], [ "However, sometimes we want to work with data across rows within a variable -- that is, aggregate/summarize values rowwise rather than columnwise.", "_____no_output_____" ], [ "<img src=\"images/aggregate-series.png\" alt=\"aggregate-series.png\" width=\"500\" height=\"500\">", "_____no_output_____" ], [ "Note that we return a single value representing some aggregation of the elements we started with. This is known as a **summary function**, but you can think of it as summarizing across rows.", "_____no_output_____" ], [ "This is what we are going to talk about next.", "_____no_output_____" ], [ "## Summarizing a Series", "_____no_output_____" ], [ "### Summary Methods", "_____no_output_____" ], [ "The easiest way to summarize a specific series is by using bracket subsetting notation and the built-in Series methods:", "_____no_output_____" ] ], [ [ "flights_df['distance'].sum()", "_____no_output_____" ] ], [ [ "Note that a *single value* was returned because this is a **summary operation** -- we are summing the `distance` variable across all rows.", "_____no_output_____" ], [ "There are other summary methods with a series:", "_____no_output_____" ] ], [ [ "flights_df['distance'].mean()", "_____no_output_____" ], [ "flights_df['distance'].median()", "_____no_output_____" ], [ "flights_df['distance'].mode()", "_____no_output_____" ] ], [ [ "All of the above methods work on quantitative variables, but we also have methods for character variables:", "_____no_output_____" ] ], [ [ "flights_df['carrier'].value_counts()", "_____no_output_____" ] ], [ [ "While the above isn't *technically* returning a single value, it's still a useful Series method summarizing our data.", "_____no_output_____" ], [ "\n Your Turn\n\n\n1\\. What is the difference between a window operation and a summary operation?\n\n2\\. Fill in the blanks in the following code to calculate the mean delay in departure:\n\n ```python\n flights_df['_____']._____()\n ```\n\n\n3\\. Find the distinct number of carriers. Hint: look for a method to find the number of unique values in a Series.", "_____no_output_____" ], [ "### Describe Method", "_____no_output_____" ], [ "There is also a method `describe()` that provides a lot of this information -- this is especially useful in exploratory data analysis.", "_____no_output_____" ] ], [ [ "flights_df['distance'].describe()", "_____no_output_____" ] ], [ [ "Note that `describe()` will return different results depending on the `type` of the Series:", "_____no_output_____" ] ], [ [ "flights_df['carrier'].describe()", "_____no_output_____" ] ], [ [ "## Summarizing a DataFrame", "_____no_output_____" ], [ "The above methods and operations are nice, but sometimes we want to work with multiple variables rather than just one.", "_____no_output_____" ], [ "### Extending Summary Methods to DataFrames", "_____no_output_____" ], [ "Recall how we select variables from a DataFrame:", "_____no_output_____" ], [ "* Single-bracket subset notation", "_____no_output_____" ], [ "* Pass a list of quoted variable names into the list", "_____no_output_____" ], [ "```python\nflights_df[['sched_dep_time', 'dep_time']]\n```", "_____no_output_____" ], [ "We can use *the same summary methods from the Series on the DataFrame* to summarize data:", "_____no_output_____" ] ], [ [ "flights_df[['sched_dep_time', 'dep_time']].mean()", "_____no_output_____" ], [ "flights_df[['sched_dep_time', 'dep_time']].median()", "_____no_output_____" ] ], [ [ "\n Question: What is the class of <code>flights_df[['sched_dep_time', 'dep_time']].median()</code>?\n", "_____no_output_____" ], [ "This returns a `pandas.core.series.Series` object -- the Index is the variable name and the values are the summarieze values.", "_____no_output_____" ], [ "### The Aggregation Method", "_____no_output_____" ], [ "While summary methods can be convenient, there are a few drawbacks to using them on DataFrames:", "_____no_output_____" ], [ "1. You have to lookup or remember the method names each time\n2. You can only apply one summary method at a time\n3. You have to apply the same summary method to all variables\n4. A Series is returned rather than a DataFrame -- this makes it difficult to use the values in our analysis later", "_____no_output_____" ], [ "In order to get around these problems, the DataFrame has a powerful method `agg()`:", "_____no_output_____" ] ], [ [ "flights_df.agg({\n 'sched_dep_time': ['mean']\n})", "_____no_output_____" ] ], [ [ "There are a few things to notice about the `agg()` method:", "_____no_output_____" ], [ "1. A `dict` is passed to the method with variable names as keys and a list of quoted summaries as values", "_____no_output_____" ], [ "2. *A DataFrame is returned* with variable names as variables and summaries as rows", "_____no_output_____" ], [ "We can extend this to multiple variables by adding elements to the `dict`:", "_____no_output_____" ] ], [ [ "flights_df.agg({\n 'sched_dep_time': ['mean'],\n 'dep_time': ['mean']\n})", "_____no_output_____" ] ], [ [ "And because the values of the `dict` are lists, we can do additional aggregations at the same time:", "_____no_output_____" ] ], [ [ "flights_df.agg({\n 'sched_dep_time': ['mean', 'median'],\n 'dep_time': ['mean', 'min']\n})", "_____no_output_____" ] ], [ [ "And notice that not all variables have to have the same list of summaries.", "_____no_output_____" ], [ "\n Your Turn\n", "_____no_output_____" ], [ "1\\. What class of object is the returned by the below code?\n\n```python\nflights_df[['air_time', 'distance']].mean()\n```", "_____no_output_____" ], [ "2\\. What class of object is returned by the below code?\n\n```python\nflights_df.agg({\n 'air_time': ['mean'],\n 'distance': ['mean']\n})\n```", "_____no_output_____" ], [ "3\\. Fill in the blanks in the below code to calculate the minimum and maximum distances traveled and the mean and median arrival delay:\n\n```python\nflights_df.agg({\n '_____': ['min', '_____'],\n '_____': ['_____', 'median']\n})\n```", "_____no_output_____" ], [ "### Describe Method", "_____no_output_____" ], [ "While `agg()` is a powerful method, the `describe()` method -- similar to the Series `describe()` method -- is a great choice during exploratory data analysis:", "_____no_output_____" ] ], [ [ "flights_df.describe()", "_____no_output_____" ] ], [ [ "\n Question: What is missing from the above result?\n", "_____no_output_____" ], [ "The string variables are missing!", "_____no_output_____" ], [ "We can make `describe()` compute on all variable types using the `include` parameter and passing a list of data types to include:", "_____no_output_____" ] ], [ [ "flights_df.describe(include = ['int', 'float', 'object'])", "_____no_output_____" ] ], [ [ "## Questions\n\nAre there any questions before we move on?", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown" ], [ "code", "code" ], [ "markdown", "markdown", "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ] ]

4ad7e74e3e976d1b7c7448fb1fc844c41fbec26a

1,969

ipynb

Jupyter Notebook

104_Maximum_Depth_of_Binary_Tree.ipynb

NikhilAshodariya/LeetCode

58584be92a23eb01153801560867d3076fbf72b6

[ "MIT" ]

null

104_Maximum_Depth_of_Binary_Tree.ipynb

NikhilAshodariya/LeetCode

58584be92a23eb01153801560867d3076fbf72b6

[ "MIT" ]

null

104_Maximum_Depth_of_Binary_Tree.ipynb

NikhilAshodariya/LeetCode

58584be92a23eb01153801560867d3076fbf72b6

[ "MIT" ]

null

26.972603

80

0.472829

[ [ [ "class TreeNode:\n def __init__(self, x):\n self.val = x\n self.left = None\n self.right = None", "_____no_output_____" ], [ "def maxDepth(root):\n def countHeight(node,nodeCount):\n if node.left == None and node.right == None:\n return nodeCount\n \n elif node.left == None and node.right != None:\n count = countHeight(node.right, nodeCount + 1)\n return count if count > nodeCount else nodeCount\n \n elif node.left != None and node.right == None:\n count = countHeight(node.left, nodeCount + 1)\n return count if count > nodeCount else nodeCount\n \n elif node.left != None and node.right != None:\n lc = nodeCount + countHeight(node.left, nodeCount + 1)\n rc = nodeCount + countHeight(node.right, nodeCount + 1)\n \n return lc if lc > rc else rc\n return countHeight(root, 1)", "_____no_output_____" ] ] ]

[ "code" ]

[ [ "code", "code" ] ]

4ad7ed023657c68b1444d146d279baf3a7a2408d

641,649

ipynb

Jupyter Notebook

analysis/.ipynb_checkpoints/Philly Trace Analysis-checkpoint.ipynb

boringlee24/philly-traces

653e107e69c22b7216b2bc502786b13d95883968

[ "CC-BY-4.0" ]

null

analysis/.ipynb_checkpoints/Philly Trace Analysis-checkpoint.ipynb

boringlee24/philly-traces

653e107e69c22b7216b2bc502786b13d95883968

[ "CC-BY-4.0" ]

null

analysis/.ipynb_checkpoints/Philly Trace Analysis-checkpoint.ipynb

boringlee24/philly-traces

653e107e69c22b7216b2bc502786b13d95883968

[ "CC-BY-4.0" ]

null

658.101538

34,916

0.946128

[ [ [ "import csv\nimport datetime\nimport json\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport os", "_____no_output_____" ] ], [ [ "## Constants", "_____no_output_____" ] ], [ [ "LOGDIR = '../trace-data'\nDATE_FORMAT_STR = '%Y-%m-%d %H:%M:%S'\nMINUTES_PER_DAY = (24 * 60)\nMICROSECONDS_PER_MINUTE = (60 * 1000)", "_____no_output_____" ] ], [ [ "## Utility code", "_____no_output_____" ] ], [ [ "def parse_date(date_str):\n \"\"\"Parses a date string and returns a datetime object if possible.\n \n Args:\n date_str: A string representing a date.\n \n Returns:\n A datetime object if the input string could be successfully\n parsed, None otherwise.\n \"\"\"\n if date_str is None or date_str == '' or date_str == 'None':\n return None\n return datetime.datetime.strptime(date_str, DATE_FORMAT_STR)\n\ndef timedelta_to_minutes(timedelta):\n \"\"\"Converts a datetime timedelta object to minutes.\n \n Args:\n timedelta: The timedelta to convert.\n \n Returns:\n The number of minutes captured in the timedelta.\n \"\"\"\n minutes = 0.0\n minutes += timedelta.days * MINUTES_PER_DAY\n minutes += timedelta.seconds / 60.0\n minutes += timedelta.microseconds / MICROSECONDS_PER_MINUTE\n return minutes\n\ndef round_to_nearest_minute(t):\n \"\"\"Rounds a datetime object down to the nearest minute.\n \n Args:\n t: A datetime object.\n \n Returns:\n A new rounded down datetime object.\n \"\"\"\n return t - datetime.timedelta(seconds=t.second, microseconds=t.microsecond)\n\ndef add_minute(t):\n \"\"\"Adds a single minute to a datetime object.\n \n Args:\n t: A datetime object.\n \n Returns:\n A new datetime object with an additional minute.\n \"\"\"\n return t + datetime.timedelta(seconds=60)", "_____no_output_____" ], [ "def get_cdf(data):\n \"\"\"Returns the CDF of the given data.\n \n Args:\n data: A list of numerical values.\n \n Returns:\n An pair of lists (x, y) for plotting the CDF.\n \"\"\"\n sorted_data = sorted(data)\n p = 100. * np.arange(len(sorted_data)) / (len(sorted_data) - 1)\n return sorted_data, p", "_____no_output_____" ], [ "class Job:\n \"\"\"Encapsulates a job.\"\"\"\n \n def __init__(self, status, vc, jobid, attempts, submitted_time, user):\n \"\"\"Records job parameters and computes key metrics.\n \n Stores the passed in arguments as well as the number of GPUs\n requested by the job. In addition, computes the queueing delay\n as defined as the delta between the submission time and the start\n time of the first attempt. Finally, computes run time as defined\n as the delta between the initial attempt's start time and the last\n attempt's finish time.\n \n NOTE: Some jobs do not have any recorded attempts, and some attempts\n have missing start and/or end times. A job's latest attempt having no\n end time indicates that the job was still running when the log data\n was collected.\n \n Args:\n status: One of 'Pass', 'Killed', 'Failed'.\n vc: The hash of the virtual cluster id the job was run in.\n jobid: The hash of the job id.\n attempts: A list of dicts, where each dict contains the following keys:\n 'start_time': The start time of the attempt.\n 'end_time': The end time of the attempt.\n 'detail': A list of nested dicts where each dict contains \n the following keys:\n 'ip': The server id.\n 'gpus': A list of the GPU ids allotted for this attempt.\n submitted_time: The time the job was submitted to the queue.\n user: The user's id. \n \"\"\"\n self._status = status\n self._vc = vc\n self._jobid = jobid\n for attempt in attempts:\n attempt['start_time'] = parse_date(attempt['start_time'])\n attempt['end_time'] = parse_date(attempt['end_time'])\n self._attempts = attempts\n self._submitted_time = parse_date(submitted_time)\n self._user = user\n \n if len(self._attempts) == 0:\n self._num_gpus = None\n self._run_time = None\n self._queueing_delay = None\n else:\n self._num_gpus = sum([len(detail['gpus']) for detail in self._attempts[0]['detail']])\n if self._attempts[0]['start_time'] is None:\n self._run_time = None\n self._queueing_delay = None\n else:\n if self._attempts[-1]['end_time'] is None:\n self._run_time = None\n else:\n self._run_time = \\\n timedelta_to_minutes(self._attempts[-1]['end_time'] -\n self._attempts[0]['start_time'])\n self._queueing_delay = \\\n timedelta_to_minutes(self._attempts[0]['start_time'] -\n self._submitted_time)\n \n @property\n def status(self):\n return self._status\n \n @property\n def vc(self):\n return self._vc\n \n @property\n def jobid(self):\n return self._jobid\n \n @property\n def attempts(self):\n return self._attempts\n \n @property\n def submitted_time(self):\n return self._submitted_time\n \n @property\n def user(self):\n return self._user\n \n @property\n def num_gpus(self):\n return self._num_gpus\n \n @property\n def queueing_delay(self):\n return self._queueing_delay\n \n @property\n def run_time(self):\n return self._run_time", "_____no_output_____" ], [ "def get_bucket_from_num_gpus(num_gpus):\n \"\"\"Maps GPU count to a bucket for plotting purposes.\"\"\"\n if num_gpus is None:\n return None\n if num_gpus == 1:\n return 0\n elif num_gpus >= 2 and num_gpus <= 4:\n return 1\n elif num_gpus >= 5 and num_gpus <= 8:\n return 2\n elif num_gpus > 8:\n return 3\n else:\n return None\n \ndef get_plot_config_from_bucket(bucket):\n \"\"\"Returns plotting configuration information.\"\"\"\n if bucket == 0:\n return ('1', 'green', '-')\n elif bucket == 1:\n return ('2-4', 'blue', '-.')\n elif bucket == 2:\n return ('5-8', 'red', '--')\n elif bucket == 3:\n return ('>8', 'purple', ':')", "_____no_output_____" ] ], [ [ "## Load the cluster log", "_____no_output_____" ] ], [ [ "cluster_job_log_path = os.path.join(LOGDIR, 'cluster_job_log')\nwith open(cluster_job_log_path, 'r') as f:\n cluster_job_log = json.load(f)\njobs = [Job(**job) for job in cluster_job_log]", "_____no_output_____" ], [ "len(jobs)", "_____no_output_____" ] ], [ [ "# Job Runtimes (Figure 2)", "_____no_output_____" ] ], [ [ "run_times = {}\nfor job in jobs:\n num_gpus = job.num_gpus\n bucket = get_bucket_from_num_gpus(num_gpus)\n if bucket is None:\n continue\n if bucket not in run_times:\n run_times[bucket] = []\n run_time = job.run_time\n if run_time is not None:\n run_times[bucket].append(run_time)", "_____no_output_____" ], [ "buckets = sorted([bucket for bucket in run_times])\nfor bucket in buckets:\n num_gpus, color, linestyle = get_plot_config_from_bucket(bucket)\n x, y = get_cdf(run_times[bucket])\n plt.plot(x, y, label='%s GPU' % (num_gpus), color=color, linestyle=linestyle)\nplt.legend(loc='lower right')\nplt.xscale('log')\nplt.xlim(10 ** -1, 10 ** 4)\nplt.ylim(0, 100)\nplt.xlabel('Time (min)')\nplt.ylabel('CDF')\nplt.grid(alpha=.3, linestyle='--')\nplt.show()", "_____no_output_____" ] ], [ [ "# Queueing Delay (Figure 3)", "_____no_output_____" ] ], [ [ "queueing_delays = {}\nfor job in jobs:\n vc = job.vc\n if vc not in queueing_delays:\n queueing_delays[vc] = {}\n bucket = get_bucket_from_num_gpus(job.num_gpus)\n if bucket is None:\n continue\n if bucket not in queueing_delays[vc]:\n queueing_delays[vc][bucket] = []\n # NOTE: Each period between the job being placed on the queue\n # and being scheduled on a machine is recorded as an individual\n # queueing delay.\n queueing_delay = 0.0\n queue_time = job.submitted_time\n for attempt in job.attempts:\n start_time = attempt['start_time']\n if queue_time is not None and start_time is not None:\n queueing_delay = timedelta_to_minutes(start_time - queue_time)\n queue_time = attempt['end_time']\n queueing_delays[vc][bucket].append(queueing_delay)\nfor vc in queueing_delays:\n for bucket in queueing_delays[vc]:\n queueing_delays[vc][bucket] = filter(None, queueing_delays[vc][bucket])", "_____no_output_____" ], [ "vcs = queueing_delays.keys()\nfor i, vc in enumerate(vcs):\n for bucket in queueing_delays[vc]:\n num_gpus, color, linestyle = get_plot_config_from_bucket(bucket)\n x, y = get_cdf(queueing_delays[vc][bucket])\n plt.plot(x, y, label='%s GPU' % (num_gpus), color=color, linestyle=linestyle)\n plt.title('VC %s' % (vc))\n plt.legend(loc='lower right')\n plt.xscale('log')\n plt.ylim(0, 100)\n plt.xlim(10 ** -1, 10 ** 4)\n plt.xlabel('Time (min)')\n plt.ylabel('CDF')\n plt.grid(alpha=.3, linestyle='--')\n if i < len(vcs) - 1:\n plt.figure()\nplt.show()", "_____no_output_____" ] ], [ [ "# Locality Constraints (Figure 4)", "_____no_output_____" ] ], [ [ "data = {}\nfor i, job in enumerate(jobs):\n if len(job.attempts) == 0:\n continue\n num_gpus = job.num_gpus\n if num_gpus < 5:\n continue\n bucket = get_bucket_from_num_gpus(num_gpus)\n if bucket not in data:\n data[bucket] = {\n 'x': [],\n 'y': []\n }\n queueing_delay = job.queueing_delay\n num_servers = len(job.attempts[0]['detail'])\n data[bucket]['x'].append(queueing_delay)\n data[bucket]['y'].append(num_servers)\nfor bucket in data:\n num_gpus, _, _ = get_plot_config_from_bucket(bucket)\n if bucket == 2:\n marker = '+'\n facecolors = 'black'\n edgecolors = 'none'\n else:\n marker = 'o'\n facecolors = 'none'\n edgecolors = 'red'\n plt.scatter(data[bucket]['x'], data[bucket]['y'], label='%s GPU' % (num_gpus),\n marker=marker, facecolors=facecolors, edgecolors=edgecolors)\n plt.legend()\nplt.xscale('log')\nplt.xlabel('Time (min)')\nplt.ylabel('Num. Servers')\nplt.show()", "_____no_output_____" ] ], [ [ "# GPU Utilization (Figures 5, 6)", "_____no_output_____" ] ], [ [ "gpu_util_path = os.path.join(LOGDIR, 'cluster_gpu_util')\ngpu_util = {}\nwith open(gpu_util_path, 'r') as f:\n reader = csv.reader(f)\n next(reader)\n for row in reader:\n time = row[0][:-4] # Remove the timezone\n machineId = row[1]\n if machineId not in gpu_util:\n gpu_util[machineId] = {}\n gpu_util[machineId][time] = row[2:-1] # Ignore extra empty string at the end", "_____no_output_____" ], [ "def get_utilization_data(jobs, only_large_jobs=False, only_dedicated_servers=False):\n \"\"\"Aggregates GPU utilization data for a set of jobs.\n \n Args:\n jobs: A list of Jobs.\n only_large_jobs: If True, only considers jobs of size 8 or 16 GPUs.\n Otherwise, considers jobs of size 1, 4, 8, or 16 GPUs. \n only_dedicated_servers: If True, only considers jobs that use all GPUs\n available on a server(s).\n \n Returns:\n A dict indexed by 1) job completion status, 2) number of GPUs requested\n by the job, and 3) timestamp. The value of each nested dict is a list of\n percentages indicating the utilization of each individual GPU on the\n servers used by the job at the particular time requested.\n \"\"\"\n data = {}\n for job in jobs:\n num_gpus = job.num_gpus\n if (len(job.attempts) == 0 or\n (num_gpus != 1 and num_gpus != 4 and num_gpus != 8 and num_gpus != 16)):\n continue\n if only_large_jobs and num_gpus < 8:\n continue\n status = job.status\n if status not in data:\n data[status] = {}\n if num_gpus not in data[status]:\n data[status][num_gpus] = []\n for attempt in job.attempts:\n if only_dedicated_servers and len(attempt['detail']) > (num_gpus / 8):\n continue\n current_time = attempt['start_time']\n if current_time is None or attempt['end_time'] is None:\n continue\n current_minute = round_to_nearest_minute(current_time)\n while current_minute < attempt['end_time']:\n current_minute_str = str(current_minute)\n for detail in attempt['detail']:\n machineId = detail['ip']\n if current_minute_str in gpu_util[machineId]:\n for gpu_id in detail['gpus']:\n gpu_num = int(gpu_id[3:]) # Remove the 'gpu' prefix\n try:\n u = gpu_util[machineId][current_minute_str][gpu_num]\n if u != 'NA':\n data[status][num_gpus].append(float(u))\n except Exception as e:\n print(gpu_util[machineId][current_minute_str])\n print(gpu_num)\n raise ValueError(e)\n current_minute = add_minute(current_minute)\n return data", "_____no_output_____" ], [ "data = get_utilization_data(jobs)\nstatuses = data.keys()\nfor i, status in enumerate(statuses):\n all_num_gpus = sorted(data[status].keys())\n for num_gpus in all_num_gpus:\n if num_gpus == 1:\n color = 'green'\n linestyle = '-'\n elif num_gpus == 4:\n color = 'blue'\n linestyle = '-.'\n elif num_gpus == 8:\n color = 'red'\n linestyle = '--'\n elif num_gpus == 16:\n color = 'cyan'\n linestyle = ':'\n x, y = get_cdf(data[status][num_gpus])\n plt.plot(x, y, label='%s GPU' % (num_gpus), color=color, linestyle=linestyle)\n plt.title(status)\n plt.xlim(0, 100)\n plt.ylim(0, 100)\n plt.legend(loc='lower right')\n plt.xlabel('Utilization (%)')\n plt.ylabel('CDF')\n plt.grid(alpha=.3, linestyle='--')\n if i < len(statuses) - 1:\n plt.figure()\nplt.show()", "_____no_output_____" ], [ "data = get_utilization_data(jobs, only_large_jobs=True, only_dedicated_servers=True)\naggregate_data = {}\nfor status in data:\n for num_gpus in data[status]:\n if num_gpus not in aggregate_data:\n aggregate_data[num_gpus] = []\n aggregate_data[num_gpus] += data[status][num_gpus]\nall_num_gpus = sorted(aggregate_data.keys())\nfor num_gpus in all_num_gpus:\n if num_gpus == 8:\n linestyle = '-'\n elif num_gpus == 16:\n linestyle = '-.'\n x, y = get_cdf(aggregate_data[num_gpus])\n plt.plot(x, y, label='%s GPU' % (num_gpus), color='black', linestyle=linestyle)\nplt.xlim(0, 100)\nplt.ylim(0, 100)\nplt.legend(loc='lower right')\nplt.xlabel('Utilization (%)')\nplt.ylabel('CDF')\nplt.grid(alpha=.3, linestyle='--')\nplt.show()", "_____no_output_____" ] ], [ [ "# Host Resource Utilization (Figure 7)", "_____no_output_____" ] ], [ [ "mem_util_path = os.path.join(LOGDIR, 'cluster_mem_util')\nmem_util = []\nwith open(mem_util_path, 'r') as f:\n reader = csv.reader(f)\n next(reader)\n for row in reader:\n if row[2] == 'NA':\n continue\n mem_total = float(row[2])\n mem_free = float(row[3])\n if mem_total == 0:\n continue\n mem_util.append(100.0 * (mem_total - mem_free) / mem_total)", "_____no_output_____" ], [ "cpu_util_path = os.path.join(LOGDIR, 'cluster_cpu_util')\ncpu_util = []\nwith open(cpu_util_path, 'r') as f:\n reader = csv.reader(f)\n next(reader)\n for row in reader:\n if row[2] == 'NA':\n continue\n cpu_util.append(float(row[2]))", "_____no_output_____" ], [ "x, y = get_cdf(cpu_util)\nplt.plot(x, y, label='CPU', color='black', linestyle='-')\nx, y = get_cdf(mem_util)\nplt.plot(x, y, label='Memory', color='black', linestyle='-.')\nplt.xlim(0, 100)\nplt.ylim(0, 100)\nplt.legend(loc='lower right')\nplt.xlabel('Utilization (%)')\nplt.ylabel('CDF')\nplt.show()", "_____no_output_____" ] ] ]

[ "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ] ]

4ad80180bb06c416af73f6b743e05a8060cde597

51,919

ipynb

Jupyter Notebook

scikit-learn-exercises.ipynb

KyrylR/Machine-Learning-and-Data-Science-intoduction

cfe42a4a8ac69e20ad65b56ce53683e9c69c152d

[ "MIT" ]

null

scikit-learn-exercises.ipynb

KyrylR/Machine-Learning-and-Data-Science-intoduction

cfe42a4a8ac69e20ad65b56ce53683e9c69c152d

[ "MIT" ]

null

scikit-learn-exercises.ipynb

KyrylR/Machine-Learning-and-Data-Science-intoduction

cfe42a4a8ac69e20ad65b56ce53683e9c69c152d

[ "MIT" ]

null

37.005702

486

0.619311

[ [ [ "# Scikit-Learn Practice Exercises\n\nThis notebook offers a set of excercises for different tasks with Scikit-Learn.\n\nNotes:\n* There may be more than one different way to answer a question or complete an exercise. \n* Some skeleton code has been implemented for you.\n* Exercises are based off (and directly taken from) the quick [introduction to Scikit-Learn notebook](https://github.com/mrdbourke/zero-to-mastery-ml/blob/master/section-2-data-science-and-ml-tools/introduction-to-scikit-learn.ipynb).\n* Different tasks will be detailed by comments or text. Places to put your own code are defined by `###` (don't remove anything other than `###`).\n\nFor further reference and resources, it's advised to check out the [Scikit-Learn documnetation](https://scikit-learn.org/stable/user_guide.html).\n\nAnd if you get stuck, try searching for a question in the following format: \"how to do XYZ with Scikit-Learn\", where XYZ is the function you want to leverage from Scikit-Learn.\n\nSince we'll be working with data, we'll import Scikit-Learn's counterparts, Matplotlib, NumPy and pandas.\n\nLet's get started.", "_____no_output_____" ] ], [ [ "# Setup matplotlib to plot inline (within the notebook)\n###\n\n# Import the pyplot module of Matplotlib as plt\n###\n\n# Import pandas under the abbreviation 'pd'\n###\n\n# Import NumPy under the abbreviation 'np'\n###", "_____no_output_____" ] ], [ [ "## End-to-end Scikit-Learn classification workflow\n\nLet's start with an end to end Scikit-Learn workflow.\n\nMore specifically, we'll:\n1. Get a dataset ready\n2. Prepare a machine learning model to make predictions\n3. Fit the model to the data and make a prediction\n4. Evaluate the model's predictions \n\nThe data we'll be using is [stored on GitHub](https://github.com/mrdbourke/zero-to-mastery-ml/tree/master/data). We'll start with [`heart-disease.csv`](https://raw.githubusercontent.com/mrdbourke/zero-to-mastery-ml/master/data/heart-disease.csv), a dataset which contains anonymous patient data and whether or not they have heart disease.\n\n**Note:** When viewing a `.csv` on GitHub, make sure it's in the raw format. For example, the URL should look like: https://raw.githubusercontent.com/mrdbourke/zero-to-mastery-ml/master/data/heart-disease.csv\n\n### 1. Getting a dataset ready", "_____no_output_____" ] ], [ [ "# Import the heart disease dataset and save it to a variable\n# using pandas and read_csv()\n# Hint: You can directly pass the URL of a csv to read_csv()\nheart_disease = ###\n\n# Check the first 5 rows of the data\n###", "_____no_output_____" ] ], [ [ "Our goal here is to build a machine learning model on all of the columns except `target` to predict `target`.\n\nIn essence, the `target` column is our **target variable** (also called `y` or `labels`) and the rest of the other columns are our independent variables (also called `data` or `X`).\n\nAnd since our target variable is one thing or another (heart disease or not), we know our problem is a classification problem (classifying whether something is one thing or another).\n\nKnowing this, let's create `X` and `y` by splitting our dataframe up.", "_____no_output_____" ] ], [ [ "# Create X (all columns except target)\nX = ###\n\n# Create y (only the target column)\ny = ###", "_____no_output_____" ] ], [ [ "Now we've split our data into `X` and `y`, we'll use Scikit-Learn to split it into training and test sets.", "_____no_output_____" ] ], [ [ "# Import train_test_split from sklearn's model_selection module\n###\n\n# Use train_test_split to split X & y into training and test sets\nX_train, X_test, y_train, y_test = ###", "_____no_output_____" ], [ "# View the different shapes of the training and test datasets\n###", "_____no_output_____" ] ], [ [ "What do you notice about the different shapes of the data?\n\nSince our data is now in training and test sets, we'll build a machine learning model to fit patterns in the training data and then make predictions on the test data.\n\nTo figure out which machine learning model we should use, you can refer to [Scikit-Learn's machine learning map](https://scikit-learn.org/stable/tutorial/machine_learning_map/index.html).\n\nAfter following the map, you decide to use the [`RandomForestClassifier`](https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html).\n\n### 2. Preparing a machine learning model", "_____no_output_____" ] ], [ [ "# Import the RandomForestClassifier from sklearn's ensemble module\n###\n\n# Instantiate an instance of RandomForestClassifier as clf\nclf = ", "_____no_output_____" ] ], [ [ "Now you've got a `RandomForestClassifier` instance, let's fit it to the training data.\n\nOnce it's fit, we'll make predictions on the test data.\n\n### 3. Fitting a model and making predictions", "_____no_output_____" ] ], [ [ "# Fit the RandomForestClassifier to the training data\nclf.fit(###, ###)", "_____no_output_____" ], [ "# Use the fitted model to make predictions on the test data and\n# save the predictions to a variable called y_preds\ny_preds = clf.predict(###)", "_____no_output_____" ] ], [ [ "### 4. Evaluating a model's predictions\n\nEvaluating predictions is as important making them. Let's check how our model did by calling the `score()` method on it and passing it the training (`X_train, y_train`) and testing data (`X_test, y_test`).", "_____no_output_____" ] ], [ [ "# Evaluate the fitted model on the training set using the score() function\n###", "_____no_output_____" ], [ "# Evaluate the fitted model on the test set using the score() function\n###", "_____no_output_____" ] ], [ [ "* How did you model go? \n* What metric does `score()` return for classifiers? \n* Did your model do better on the training dataset or test dataset?", "_____no_output_____" ], [ "## Experimenting with different classification models\n\nNow we've quickly covered an end-to-end Scikit-Learn workflow and since experimenting is a large part of machine learning, we'll now try a series of different machine learning models and see which gets the best results on our dataset.\n\nGoing through the [Scikit-Learn machine learning map](https://scikit-learn.org/stable/tutorial/machine_learning_map/index.html), we see there are a number of different classification models we can try (different models are in the green boxes).\n\nFor this exercise, the models we're going to try and compare are:\n* [LinearSVC](https://scikit-learn.org/stable/modules/svm.html#classification)\n* [KNeighborsClassifier](https://scikit-learn.org/stable/modules/neighbors.html) (also known as K-Nearest Neighbors or KNN)\n* [SVC](https://scikit-learn.org/stable/modules/svm.html#classification) (also known as support vector classifier, a form of [support vector machine](https://en.wikipedia.org/wiki/Support-vector_machine))\n* [LogisticRegression](https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html) (despite the name, this is actually a classifier)\n* [RandomForestClassifier](https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestClassifier.html) (an ensemble method and what we used above)\n\nWe'll follow the same workflow we used above (except this time for multiple models):\n1. Import a machine learning model\n2. Get it ready\n3. Fit it to the data and make predictions\n4. Evaluate the fitted model\n\n**Note:** Since we've already got the data ready, we can reuse it in this section.", "_____no_output_____" ] ], [ [ "# Import LinearSVC from sklearn's svm module\n###\n\n# Import KNeighborsClassifier from sklearn's neighbors module\n###\n\n# Import SVC from sklearn's svm module\n###\n\n# Import LogisticRegression from sklearn's linear_model module\n###\n\n# Note: we don't have to import RandomForestClassifier, since we already have", "_____no_output_____" ] ], [ [ "Thanks to the consistency of Scikit-Learn's API design, we can use virtually the same code to fit, score and make predictions with each of our models.\n\nTo see which model performs best, we'll do the following:\n1. Instantiate each model in a dictionary\n2. Create an empty results dictionary\n3. Fit each model on the training data\n4. Score each model on the test data\n5. Check the results\n\nIf you're wondering what it means to instantiate each model in a dictionary, see the example below.", "_____no_output_____" ] ], [ [ "# EXAMPLE: Instantiating a RandomForestClassifier() in a dictionary\nexample_dict = {\"RandomForestClassifier\": RandomForestClassifier()}\n\n# Create a dictionary called models which contains all of the classification models we've imported\n# Make sure the dictionary is in the same format as example_dict\n# The models dictionary should contain 5 models\nmodels = {\"LinearSVC\": ###,\n \"KNN\": ###,\n \"SVC\": ###,\n \"LogisticRegression\": ###,\n \"RandomForestClassifier\": ###}\n\n# Create an empty dictionary called results\nresults = ###", "_____no_output_____" ] ], [ [ "Since each model we're using has the same `fit()` and `score()` functions, we can loop through our models dictionary and, call `fit()` on the training data and then call `score()` with the test data.", "_____no_output_____" ] ], [ [ "# EXAMPLE: Looping through example_dict fitting and scoring the model\nexample_results = {}\nfor model_name, model in example_dict.items():\n model.fit(X_train, y_train)\n example_results[model_name] = model.score(X_test, y_test)\n\n# EXAMPLE: View the results\nexample_results ", "_____no_output_____" ], [ "# Loop through the models dictionary items, fitting the model on the training data\n# and appending the model name and model score on the test data to the results dictionary\nfor model_name, model in ###:\n model.fit(###)\n results[model_name] = model.score(###)\n\n# View the results\nresults", "_____no_output_____" ] ], [ [ "* Which model performed the best? \n* Do the results change each time you run the cell? \n* Why do you think this is?\n\nDue to the randomness of how each model finds patterns in the data, you might notice different results each time.\n\nWithout manually setting the random state using the `random_state` parameter of some models or using a NumPy random seed, every time you run the cell, you'll get slightly different results.\n\nLet's see this in effect by running the same code as the cell above, except this time setting a [NumPy random seed equal to 42](https://docs.scipy.org/doc/numpy-1.15.1/reference/generated/numpy.random.seed.html).", "_____no_output_____" ] ], [ [ "# Run the same code as the cell above, except this time set a NumPy random seed\n# equal to 42\nnp.random.seed(###)\n\nfor model_name, model in models.items():\n model.fit(X_train, y_train)\n results[model_name] = model.score(X_test, y_test)\n \nresults", "_____no_output_____" ] ], [ [ "* Run the cell above a few times, what do you notice about the results? \n* Which model performs the best this time?\n* What happens if you add a NumPy random seed to the cell where you called `train_test_split()` (towards the top of the notebook) and then rerun the cell above?\n\nLet's make our results a little more visual.", "_____no_output_____" ] ], [ [ "# Create a pandas dataframe with the data as the values of the results dictionary,\n# the index as the keys of the results dictionary and a single column called accuracy.\n# Be sure to save the dataframe to a variable.\nresults_df = pd.DataFrame(results.###(), \n results.###(), \n columns=[####])\n\n# Create a bar plot of the results dataframe using plot.bar()\n###", "_____no_output_____" ] ], [ [ "Using `np.random.seed(42)` results in the `LogisticRegression` model perfoming the best (at least on my computer).\n\nLet's tune its hyperparameters and see if we can improve it.\n\n### Hyperparameter Tuning\n\nRemember, if you're ever trying to tune a machine learning models hyperparameters and you're not sure where to start, you can always search something like \"MODEL_NAME hyperparameter tuning\".\n\nIn the case of LogisticRegression, you might come across articles, such as [Hyperparameter Tuning Using Grid Search by Chris Albon](https://chrisalbon.com/machine_learning/model_selection/hyperparameter_tuning_using_grid_search/).\n\nThe article uses [`GridSearchCV`](https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.GridSearchCV.html) but we're going to be using [`RandomizedSearchCV`](https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.RandomizedSearchCV.html).\n\nThe different hyperparameters to search over have been setup for you in `log_reg_grid` but feel free to change them.", "_____no_output_____" ] ], [ [ "# Different LogisticRegression hyperparameters\nlog_reg_grid = {\"C\": np.logspace(-4, 4, 20),\n \"solver\": [\"liblinear\"]}", "_____no_output_____" ] ], [ [ "Since we've got a set of hyperparameters we can import `RandomizedSearchCV`, pass it our dictionary of hyperparameters and let it search for the best combination.", "_____no_output_____" ] ], [ [ "# Setup np random seed of 42\nnp.random.seed(###)\n\n# Import RandomizedSearchCV from sklearn's model_selection module\n\n\n# Setup an instance of RandomizedSearchCV with a LogisticRegression() estimator,\n# our log_reg_grid as the param_distributions, a cv of 5 and n_iter of 5.\nrs_log_reg = RandomizedSearchCV(estimator=###,\n param_distributions=###,\n cv=###,\n n_iter=###,\n verbose=###)\n\n# Fit the instance of RandomizedSearchCV\n###", "_____no_output_____" ] ], [ [ "Once `RandomizedSearchCV` has finished, we can find the best hyperparmeters it found using the `best_params_` attributes.", "_____no_output_____" ] ], [ [ "# Find the best parameters of the RandomizedSearchCV instance using the best_params_ attribute\n###", "_____no_output_____" ], [ "# Score the instance of RandomizedSearchCV using the test data\n###", "_____no_output_____" ] ], [ [ "After hyperparameter tuning, did the models score improve? What else could you try to improve it? Are there any other methods of hyperparameter tuning you can find for `LogisticRegression`?\n\n### Classifier Model Evaluation\n\nWe've tried to find the best hyperparameters on our model using `RandomizedSearchCV` and so far we've only been evaluating our model using the `score()` function which returns accuracy. \n\nBut when it comes to classification, you'll likely want to use a few more evaluation metrics, including:\n* [**Confusion matrix**](https://www.dataschool.io/simple-guide-to-confusion-matrix-terminology/) - Compares the predicted values with the true values in a tabular way, if 100% correct, all values in the matrix will be top left to bottom right (diagnol line).\n* [**Cross-validation**](https://scikit-learn.org/stable/modules/cross_validation.html) - Splits your dataset into multiple parts and train and tests your model on each part and evaluates performance as an average. \n* [**Precision**](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.precision_score.html#sklearn.metrics.precision_score) - Proportion of true positives over total number of samples. Higher precision leads to less false positives.\n* [**Recall**](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.recall_score.html#sklearn.metrics.recall_score) - Proportion of true positives over total number of true positives and false positives. Higher recall leads to less false negatives.\n* [**F1 score**](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.f1_score.html#sklearn.metrics.f1_score) - Combines precision and recall into one metric. 1 is best, 0 is worst.\n* [**Classification report**](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.classification_report.html) - Sklearn has a built-in function called `classification_report()` which returns some of the main classification metrics such as precision, recall and f1-score.\n* [**ROC Curve**](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.roc_score.html) - [Receiver Operating Characterisitc](https://en.wikipedia.org/wiki/Receiver_operating_characteristic) is a plot of true positive rate versus false positive rate.\n* [**Area Under Curve (AUC)**](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.roc_auc_score.html) - The area underneath the ROC curve. A perfect model achieves a score of 1.0.\n\nBefore we get to these, we'll instantiate a new instance of our model using the best hyerparameters found by `RandomizedSearchCV`. ", "_____no_output_____" ] ], [ [ "# Instantiate a LogisticRegression classifier using the best hyperparameters from RandomizedSearchCV\nclf = LogisticRegression(###)\n\n# Fit the new instance of LogisticRegression with the best hyperparameters on the training data \n###", "_____no_output_____" ] ], [ [ "Now it's to import the relative Scikit-Learn methods for each of the classification evaluation metrics we're after.", "_____no_output_____" ] ], [ [ "# Import confusion_matrix and classification_report from sklearn's metrics module\n###\n\n# Import precision_score, recall_score and f1_score from sklearn's metrics module\n###\n\n# Import plot_roc_curve from sklearn's metrics module\n###", "_____no_output_____" ] ], [ [ "Evaluation metrics are very often comparing a model's predictions to some ground truth labels.\n\nLet's make some predictions on the test data using our latest model and save them to `y_preds`.", "_____no_output_____" ] ], [ [ "# Make predictions on test data and save them\n###", "_____no_output_____" ] ], [ [ "Time to use the predictions our model has made to evaluate it beyond accuracy.", "_____no_output_____" ] ], [ [ "# Create a confusion matrix using the confusion_matrix function\n###", "_____no_output_____" ] ], [ [ "**Challenge:** The in-built `confusion_matrix` function in Scikit-Learn produces something not too visual, how could you make your confusion matrix more visual?\n\nYou might want to search something like \"how to plot a confusion matrix\". Note: There may be more than one way to do this.", "_____no_output_____" ] ], [ [ "# Create a more visual confusion matrix\n###", "_____no_output_____" ] ], [ [ "How about a classification report?", "_____no_output_____" ] ], [ [ "# Create a classification report using the classification_report function\n###", "_____no_output_____" ] ], [ [ "**Challenge:** Write down what each of the columns in this classification report are.\n\n* **Precision** - Indicates the proportion of positive identifications (model predicted class 1) which were actually correct. A model which produces no false positives has a precision of 1.0.\n* **Recall** - Indicates the proportion of actual positives which were correctly classified. A model which produces no false negatives has a recall of 1.0.\n* **F1 score** - A combination of precision and recall. A perfect model achieves an F1 score of 1.0.\n* **Support** - The number of samples each metric was calculated on.\n* **Accuracy** - The accuracy of the model in decimal form. Perfect accuracy is equal to 1.0.\n* **Macro avg** - Short for macro average, the average precision, recall and F1 score between classes. Macro avg doesn’t class imbalance into effort, so if you do have class imbalances, pay attention to this metric.\n* **Weighted avg** - Short for weighted average, the weighted average precision, recall and F1 score between classes. Weighted means each metric is calculated with respect to how many samples there are in each class. This metric will favour the majority class (e.g. will give a high value when one class out performs another due to having more samples).\n\nThe classification report gives us a range of values for precision, recall and F1 score, time to find these metrics using Scikit-Learn functions.", "_____no_output_____" ] ], [ [ "# Find the precision score of the model using precision_score()\n###", "_____no_output_____" ], [ "# Find the recall score\n###", "_____no_output_____" ], [ "# Find the F1 score\n###", "_____no_output_____" ] ], [ [ "Confusion matrix: done.\nClassification report: done.\nROC (receiver operator characteristic) curve & AUC (area under curve) score: not done.\n\nLet's fix this.\n\nIf you're unfamiliar with what a ROC curve, that's your first challenge, to read up on what one is.\n\nIn a sentence, a [ROC curve](https://en.wikipedia.org/wiki/Receiver_operating_characteristic) is a plot of the true positive rate versus the false positive rate.\n\nAnd the AUC score is the area behind the ROC curve.\n\nScikit-Learn provides a handy function for creating both of these called [`plot_roc_curve()`](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.plot_roc_curve.html).", "_____no_output_____" ] ], [ [ "# Plot a ROC curve using our current machine learning model using plot_roc_curve\n###", "_____no_output_____" ] ], [ [ "Beautiful! We've gone far beyond accuracy with a plethora extra classification evaluation metrics.\n\nIf you're not sure about any of these, don't worry, they can take a while to understand. That could be an optional extension, reading up on a classification metric you're not sure of.\n\nThe thing to note here is all of these metrics have been calculated using a single training set and a single test set. Whilst this is okay, a more robust way is to calculate them using [cross-validation](https://scikit-learn.org/stable/modules/cross_validation.html).\n\nWe can calculate various evaluation metrics using cross-validation using Scikit-Learn's [`cross_val_score()`](https://scikit-learn.org/stable/modules/generated/sklearn.model_selection.cross_val_score.html) function along with the `scoring` parameter.", "_____no_output_____" ] ], [ [ "# Import cross_val_score from sklearn's model_selection module\n###", "_____no_output_____" ], [ "# EXAMPLE: By default cross_val_score returns 5 values (cv=5).\ncross_val_score(clf, \n X, \n y, \n scoring=\"accuracy\",\n cv=5)", "_____no_output_____" ], [ "# EXAMPLE: Taking the mean of the returned values from cross_val_score \n# gives a cross-validated version of the scoring metric.\ncross_val_acc = np.mean(cross_val_score(clf,\n X,\n y,\n scoring=\"accuracy\",\n cv=5))\n\ncross_val_acc", "_____no_output_____" ] ], [ [ "In the examples, the cross-validated accuracy is found by taking the mean of the array returned by `cross_val_score()`.\n\nNow it's time to find the same for precision, recall and F1 score.", "_____no_output_____" ] ], [ [ "# Find the cross-validated precision\n###", "_____no_output_____" ], [ "# Find the cross-validated recall\n###", "_____no_output_____" ], [ "# Find the cross-validated F1 score\n###", "_____no_output_____" ] ], [ [ "### Exporting and importing a trained model\n\nOnce you've trained a model, you may want to export it and save it to file so you can share it or use it elsewhere.\n\nOne method of exporting and importing models is using the joblib library.\n\nIn Scikit-Learn, exporting and importing a trained model is known as [model persistence](https://scikit-learn.org/stable/modules/model_persistence.html).", "_____no_output_____" ] ], [ [ "# Import the dump and load functions from the joblib library\n###", "_____no_output_____" ], [ "# Use the dump function to export the trained model to file\n###", "_____no_output_____" ], [ "# Use the load function to import the trained model you just exported\n# Save it to a different variable name to the origial trained model\n###\n\n# Evaluate the loaded trained model on the test data\n###", "_____no_output_____" ] ], [ [ "What do you notice about the loaded trained model results versus the original (pre-exported) model results?\n\n\n## Scikit-Learn Regression Practice\n\nFor the next few exercises, we're going to be working on a regression problem, in other words, using some data to predict a number.\n\nOur dataset is a [table of car sales](https://docs.google.com/spreadsheets/d/1LPEIWJdSSJYrfn-P3UQDIXbEn5gg-o6I7ExLrWTTBWs/edit?usp=sharing), containing different car characteristics as well as a sale price.\n\nWe'll use Scikit-Learn's built-in regression machine learning models to try and learn the patterns in the car characteristics and their prices on a certain group of the dataset before trying to predict the sale price of a group of cars the model has never seen before.\n\nTo begin, we'll [import the data from GitHub](https://raw.githubusercontent.com/mrdbourke/zero-to-mastery-ml/master/data/car-sales-extended-missing-data.csv) into a pandas DataFrame, check out some details about it and try to build a model as soon as possible.", "_____no_output_____" ] ], [ [ "# Read in the car sales data\ncar_sales = pd.read_csv(\"https://raw.githubusercontent.com/mrdbourke/zero-to-mastery-ml/master/data/car-sales-extended-missing-data.csv\")\n\n# View the first 5 rows of the car sales data\n###", "_____no_output_____" ], [ "# Get information about the car sales DataFrame\n###", "_____no_output_____" ] ], [ [ "Looking at the output of `info()`,\n* How many rows are there total?\n* What datatypes are in each column?\n* How many missing values are there in each column?", "_____no_output_____" ] ], [ [ "# Find number of missing values in each column\n###", "_____no_output_____" ], [ "# Find the datatypes of each column of car_sales\n###", "_____no_output_____" ] ], [ [ "Knowing this information, what would happen if we tried to model our data as it is?\n\nLet's see.", "_____no_output_____" ] ], [ [ "# EXAMPLE: This doesn't work because our car_sales data isn't all numerical\nfrom sklearn.ensemble import RandomForestRegressor\ncar_sales_X, car_sales_y = car_sales.drop(\"Price\", axis=1), car_sales.Price\nrf_regressor = RandomForestRegressor().fit(car_sales_X, car_sales_y)", "_____no_output_____" ] ], [ [ "As we see, the cell above breaks because our data contains non-numerical values as well as missing data.\n\nTo take care of some of the missing data, we'll remove the rows which have no labels (all the rows with missing values in the `Price` column).", "_____no_output_____" ] ], [ [ "# Remove rows with no labels (NaN's in the Price column)\n###", "_____no_output_____" ] ], [ [ "### Building a pipeline\nSince our `car_sales` data has missing numerical values as well as the data isn't all numerical, we'll have to fix these things before we can fit a machine learning model on it.\n\nThere are ways we could do this with pandas but since we're practicing Scikit-Learn, we'll see how we might do it with the [`Pipeline`](https://scikit-learn.org/stable/modules/generated/sklearn.pipeline.Pipeline.html) class. \n\nBecause we're modifying columns in our dataframe (filling missing values, converting non-numerical data to numbers) we'll need the [`ColumnTransformer`](https://scikit-learn.org/stable/modules/generated/sklearn.compose.ColumnTransformer.html), [`SimpleImputer`](https://scikit-learn.org/stable/modules/generated/sklearn.impute.SimpleImputer.html) and [`OneHotEncoder`](https://scikit-learn.org/stable/modules/generated/sklearn.preprocessing.OneHotEncoder.html) classes as well.\n\nFinally, because we'll need to split our data into training and test sets, we'll import `train_test_split` as well.", "_____no_output_____" ] ], [ [ "# Import Pipeline from sklearn's pipeline module\n###\n\n# Import ColumnTransformer from sklearn's compose module\n###\n\n# Import SimpleImputer from sklearn's impute module\n###\n\n# Import OneHotEncoder from sklearn's preprocessing module\n###\n\n# Import train_test_split from sklearn's model_selection module\n###", "_____no_output_____" ] ], [ [ "Now we've got the necessary tools we need to create our preprocessing `Pipeline` which fills missing values along with turning all non-numerical data into numbers.\n\nLet's start with the categorical features.", "_____no_output_____" ] ], [ [ "# Define different categorical features \ncategorical_features = [\"Make\", \"Colour\"]\n\n# Create categorical transformer Pipeline\ncategorical_transformer = Pipeline(steps=[\n # Set SimpleImputer strategy to \"constant\" and fill value to \"missing\"\n (\"imputer\", SimpleImputer(strategy=###, fill_value=###)),\n # Set OneHotEncoder to ignore the unknowns\n (\"onehot\", OneHotEncoder(handle_unknown=###))])", "_____no_output_____" ] ], [ [ "It would be safe to treat `Doors` as a categorical feature as well, however since we know the vast majority of cars have 4 doors, we'll impute the missing `Doors` values as 4.", "_____no_output_____" ] ], [ [ "# Define Doors features\ndoor_feature = [\"Doors\"]\n\n# Create Doors transformer Pipeline\ndoor_transformer = Pipeline(steps=[\n # Set SimpleImputer strategy to \"constant\" and fill value to 4\n (\"imputer\", SimpleImputer(strategy=###, fill_value=###))])", "_____no_output_____" ] ], [ [ "Now onto the numeric features. In this case, the only numeric feature is the `Odometer (KM)` column. Let's fill its missing values with the median.", "_____no_output_____" ] ], [ [ "# Define numeric features (only the Odometer (KM) column)\nnumeric_features = [\"Odometer (KM)\"]\n\n# Crearte numeric transformer Pipeline\nnumeric_transformer = ###(steps=[\n # Set SimpleImputer strategy to fill missing values with the \"Median\"\n (\"imputer\", ###(strategy=###))])", "_____no_output_____" ] ], [ [ "Time to put all of our individual transformer `Pipeline`'s into a single `ColumnTransformer` instance.", "_____no_output_____" ] ], [ [ "# Setup preprocessing steps (fill missing values, then convert to numbers)\npreprocessor = ColumnTransformer(\n transformers=[\n # Use the categorical_transformer to transform the categorical_features\n (\"cat\", categorical_transformer, ###),\n # Use the door_transformer to transform the door_feature\n (\"door\", ###, door_feature),\n # Use the numeric_transformer to transform the numeric_features\n (\"num\", ###, ###)])", "_____no_output_____" ] ], [ [ "Boom! Now our `preprocessor` is ready, time to import some regression models to try out.\n\nComparing our data to the [Scikit-Learn machine learning map](https://scikit-learn.org/stable/tutorial/machine_learning_map/index.html), we can see there's a handful of different regression models we can try.\n\n* [RidgeRegression](https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.Ridge.html)\n* [SVR(kernel=\"linear\")](https://scikit-learn.org/stable/modules/generated/sklearn.svm.SVR.html) - short for Support Vector Regressor, a form form of support vector machine.\n* [SVR(kernel=\"rbf\")](https://scikit-learn.org/stable/modules/generated/sklearn.svm.SVR.html) - short for Support Vector Regressor, a form of support vector machine.\n* [RandomForestRegressor](https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestRegressor.html) - the regression version of RandomForestClassifier.", "_____no_output_____" ] ], [ [ "# Import Ridge from sklearn's linear_model module\n\n\n# Import SVR from sklearn's svm module\n\n\n# Import RandomForestRegressor from sklearn's ensemble module\n", "_____no_output_____" ] ], [ [ "Again, thanks to the design of the Scikit-Learn library, we're able to use very similar code for each of these models.\n\nTo test them all, we'll create a dictionary of regression models and an empty dictionary for regression model results.", "_____no_output_____" ] ], [ [ "# Create dictionary of model instances, there should be 4 total key, value pairs\n# in the form {\"model_name\": model_instance}.\n# Don't forget there's two versions of SVR, one with a \"linear\" kernel and the\n# other with kernel set to \"rbf\".\nregression_models = {\"Ridge\": ###,\n \"SVR_linear\": ###,\n \"SVR_rbf\": ###,\n \"RandomForestRegressor\": ###}\n\n# Create an empty dictionary for the regression results\nregression_results = ###", "_____no_output_____" ] ], [ [ "Our regression model dictionary is prepared as well as an empty dictionary to append results to, time to get the data split into `X` (feature variables) and `y` (target variable) as well as training and test sets.\n\nIn our car sales problem, we're trying to use the different characteristics of a car (`X`) to predict its sale price (`y`).", "_____no_output_____" ] ], [ [ "# Create car sales X data (every column of car_sales except Price)\ncar_sales_X = ###\n\n# Create car sales y data (the Price column of car_sales)\ncar_sales_y = ###", "_____no_output_____" ], [ "# Use train_test_split to split the car_sales_X and car_sales_y data into \n# training and test sets.\n# Give the test set 20% of the data using the test_size parameter.\n# For reproducibility set the random_state parameter to 42.\ncar_X_train, car_X_test, car_y_train, car_y_test = train_test_split(###,\n ###,\n test_size=###,\n random_state=###)\n\n# Check the shapes of the training and test datasets\n###", "_____no_output_____" ] ], [ [ "* How many rows are in each set?\n* How many columns are in each set?\n\nAlright, our data is split into training and test sets, time to build a small loop which is going to:\n1. Go through our `regression_models` dictionary\n2. Create a `Pipeline` which contains our `preprocessor` as well as one of the models in the dictionary\n3. Fits the `Pipeline` to the car sales training data\n4. Evaluates the target model on the car sales test data and appends the results to our `regression_results` dictionary", "_____no_output_____" ] ], [ [ "# Loop through the items in the regression_models dictionary\nfor model_name, model in regression_models.items():\n \n # Create a model Pipeline with a preprocessor step and model step\n model_pipeline = Pipeline(steps=[(\"preprocessor\", ###),\n (\"model\", ###)])\n \n # Fit the model Pipeline to the car sales training data\n print(f\"Fitting {model_name}...\")\n model_pipeline.###(###, ###)\n \n # Score the model Pipeline on the test data appending the model_name to the \n # results dictionary\n print(f\"Scoring {model_name}...\")\n regression_results[model_name] = model_pipeline.score(###, \n ###)", "_____no_output_____" ] ], [ [ "Our regression models have been fit, let's see how they did!", "_____no_output_____" ] ], [ [ "# Check the results of each regression model by printing the regression_results\n# dictionary\n###", "_____no_output_____" ] ], [ [ "* Which model did the best?\n* How could you improve its results?\n* What metric does the `score()` method of a regression model return by default?\n\nSince we've fitted some models but only compared them via the default metric contained in the `score()` method (R^2 score or coefficient of determination), let's take the `RidgeRegression` model and evaluate it with a few other [regression metrics](https://scikit-learn.org/stable/modules/model_evaluation.html#regression-metrics).\n\nSpecifically, let's find:\n1. **R^2 (pronounced r-squared) or coefficient of determination** - Compares your models predictions to the mean of the targets. Values can range from negative infinity (a very poor model) to 1. For example, if all your model does is predict the mean of the targets, its R^2 value would be 0. And if your model perfectly predicts a range of numbers it's R^2 value would be 1. \n2. **Mean absolute error (MAE)** - The average of the absolute differences between predictions and actual values. It gives you an idea of how wrong your predictions were.\n3. **Mean squared error (MSE)** - The average squared differences between predictions and actual values. Squaring the errors removes negative errors. It also amplifies outliers (samples which have larger errors).\n\nScikit-Learn has a few classes built-in which are going to help us with these, namely, [`mean_absolute_error`](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.mean_absolute_error.html), [`mean_squared_error`](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.mean_squared_error.html) and [`r2_score`](https://scikit-learn.org/stable/modules/generated/sklearn.metrics.r2_score.html).", "_____no_output_____" ] ], [ [ "# Import mean_absolute_error from sklearn's metrics module\n###\n\n# Import mean_squared_error from sklearn's metrics module\n###\n\n# Import r2_score from sklearn's metrics module\n###", "_____no_output_____" ] ], [ [ "All the evaluation metrics we're concerned with compare a model's predictions with the ground truth labels. Knowing this, we'll have to make some predictions.\n\nLet's create a `Pipeline` with the `preprocessor` and a `Ridge()` model, fit it on the car sales training data and then make predictions on the car sales test data.", "_____no_output_____" ] ], [ [ "# Create RidgeRegression Pipeline with preprocessor as the \"preprocessor\" and\n# Ridge() as the \"model\".\nridge_pipeline = ###(steps=[(\"preprocessor\", ###),\n (\"model\", Ridge())])\n\n# Fit the RidgeRegression Pipeline to the car sales training data\nridge_pipeline.fit(###, ###)\n\n# Make predictions on the car sales test data using the RidgeRegression Pipeline\ncar_y_preds = ridge_pipeline.###(###)\n\n# View the first 50 predictions\n###", "_____no_output_____" ] ], [ [ "Nice! Now we've got some predictions, time to evaluate them. We'll find the mean squared error (MSE), mean absolute error (MAE) and R^2 score (coefficient of determination) of our model.", "_____no_output_____" ] ], [ [ "# EXAMPLE: Find the MSE by comparing the car sales test labels to the car sales predictions\nmse = mean_squared_error(car_y_test, car_y_preds)\n# Return the MSE\nmse", "_____no_output_____" ], [ "# Find the MAE by comparing the car sales test labels to the car sales predictions\n###\n# Return the MAE\n###", "_____no_output_____" ], [ "# Find the R^2 score by comparing the car sales test labels to the car sales predictions\n###\n# Return the R^2 score\n###", "_____no_output_____" ] ], [ [ "Boom! Our model could potentially do with some hyperparameter tuning (this would be a great extension). And we could probably do with finding some more data on our problem, 1000 rows doesn't seem to be sufficient.\n\n* How would you export the trained regression model?", "_____no_output_____" ], [ "## Extensions\n\nYou should be proud. Getting this far means you've worked through a classification problem and regression problem using pure (mostly) Scikit-Learn (no easy feat!).\n\nFor more exercises, check out the [Scikit-Learn getting started documentation](https://scikit-learn.org/stable/getting_started.html). A good practice would be to read through it and for the parts you find interesting, add them into the end of this notebook.\n\nFinally, as always, remember, the best way to learn something new is to try it. And try it relentlessly. If you're unsure of how to do something, never be afraid to ask a question or search for something such as, \"how to tune the hyperparmaters of a scikit-learn ridge regression model\".", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown", "markdown" ] ]

4ad80270a0390e60b951c5e97b24ca348942b070

16,746

ipynb

Jupyter Notebook

03_classification_ex2.ipynb

sferencik/handson-ml

a4011d88616392ef8e8e633fdee512e4e41a7a90

[ "Apache-2.0" ]

null

03_classification_ex2.ipynb

sferencik/handson-ml

a4011d88616392ef8e8e633fdee512e4e41a7a90

[ "Apache-2.0" ]

null

03_classification_ex2.ipynb

sferencik/handson-ml

a4011d88616392ef8e8e633fdee512e4e41a7a90

[ "Apache-2.0" ]

null

65.929134

5,860

0.828735

[ [ [ "# get data", "_____no_output_____" ] ], [ [ "import scipy.io as sio\nmnist_raw = sio.loadmat('datasets/mnist/mnist-original.mat')\nX, y = mnist_raw['data'].transpose(), mnist_raw['label'].transpose().ravel()\nX_train, X_test, y_train, y_test = X[:60000], X[60000:], y[:60000], y[60000:]", "_____no_output_____" ] ], [ [ "# plotting", "_____no_output_____" ] ], [ [ "%matplotlib inline\nimport matplotlib\nimport matplotlib.pyplot as plt\nplt.rcParams['axes.labelsize'] = 14\nplt.rcParams['xtick.labelsize'] = 12\nplt.rcParams['ytick.labelsize'] = 12", "_____no_output_____" ], [ "def plot_digit(data):\n image = data.reshape(28, 28)\n plt.imshow(image, cmap = matplotlib.cm.binary,\n interpolation=\"nearest\")\n plt.axis(\"on\")", "_____no_output_____" ], [ "plot_digit(X_train[0])", "_____no_output_____" ] ], [ [ "# shifting", "_____no_output_____" ] ], [ [ "from scipy.ndimage.interpolation import shift\ndef shift_digit(digit, delta):\n return shift(digit.reshape(28, 28), delta).reshape(28 * 28)", "_____no_output_____" ], [ "plot_digit(shift_digit(X_train[0], [3, 4]))", "_____no_output_____" ] ], [ [ "# enhance data set", "_____no_output_____" ] ], [ [ "X_train_expanded = []\ny_train_expanded = []\n\nfor (X, y) in zip(X_train, y_train):\n for delta in ([0, 0], [-1, 0], [1, 0], [0, -1], [0, 1]):\n X_train_expanded.append(shift_digit(X, delta))\n y_train_expanded.append(y)", "_____no_output_____" ] ], [ [ "# train & eval", "_____no_output_____" ] ], [ [ "from sklearn.neighbors import KNeighborsClassifier", "_____no_output_____" ], [ "knn_clf = KNeighborsClassifier(weights='distance', n_neighbors=4)", "_____no_output_____" ], [ "knn_clf.fit(X_train_expanded, y_train_expanded)", "_____no_output_____" ], [ "y_test_pred = knn_clf.predict(X_test)", "_____no_output_____" ], [ "from sklearn.metrics import accuracy_score\naccuracy_score(y_test_pred, y_test)", "_____no_output_____" ] ], [ [ "# result", "_____no_output_____" ], [ "Accuracy has increased from .9714 to .9763.", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ] ]

4ad804422238b890ae9ed68f1b2689ce02172c53

27,576

ipynb

Jupyter Notebook

notebooks/Exploratory-data-analysis-with-pandas.ipynb

Melbourne-BMDS/mimic34md2020_materials

45ff27874d211795c4cf525f85201692aec13809

[ "CC0-1.0" ]

8

2020-12-20T02:59:59.000Z

2021-09-23T06:04:01.000Z

notebooks/Exploratory-data-analysis-with-pandas.ipynb

abchapman93/mimic34md2020_materials

a348de4a7cd26bbcc21249f57819d5466ad823fe

[ "CC0-1.0" ]

5

2021-06-08T21:54:49.000Z

2022-03-12T00:38:51.000Z

notebooks/Exploratory-data-analysis-with-pandas.ipynb

philasaro/SpaCy_Melbourne_tutorial

2f49cb6d7b81a089575bcd5e41d96e8a1a129001

[ "CC0-1.0" ]

9

2020-06-26T06:00:15.000Z

2022-01-06T04:07:38.000Z

31.265306

561

0.626886

[ [ [ "This material has been adapted by @dcapurro from the Jupyter Notebook developed by:\n\nAuthor: [Yury Kashnitsky](https://yorko.github.io). Translated and edited by [Christina Butsko](https://www.linkedin.com/in/christinabutsko/), [Yuanyuan Pao](https://www.linkedin.com/in/yuanyuanpao/), [Anastasia Manokhina](https://www.linkedin.com/in/anastasiamanokhina), Sergey Isaev and [Artem Trunov](https://www.linkedin.com/in/datamove/). This material is subject to the terms and conditions of the [Creative Commons CC BY-NC-SA 4.0](https://creativecommons.org/licenses/by-nc-sa/4.0/) license. Free use is permitted for any non-commercial purpose.\n", "_____no_output_____" ], [ "## 1. Demonstration of main Pandas methods\nWell... There are dozens of cool tutorials on Pandas and visual data analysis. This one will guide us through the basic tasks when you are exploring your data (how deos the data look like?) \n\n**[Pandas](http://pandas.pydata.org)** is a Python library that provides extensive means for data analysis. Data scientists often work with data stored in table formats like `.csv`, `.tsv`, or `.xlsx`. Pandas makes it very convenient to load, process, and analyze such tabular data using SQL-like queries. In conjunction with `Matplotlib` and `Seaborn`, `Pandas` provides a wide range of opportunities for visual analysis of tabular data.\n\nThe main data structures in `Pandas` are implemented with **Series** and **DataFrame** classes. The former is a one-dimensional indexed array of some fixed data type. The latter is a two-dimensional data structure - a table - where each column contains data of the same type. You can see it as a dictionary of `Series` instances. `DataFrames` are great for representing real data: rows correspond to instances (examples, observations, etc.), and columns correspond to features of these instances.", "_____no_output_____" ] ], [ [ "import numpy as np\nimport pandas as pd\npd.set_option(\"display.precision\", 2)", "_____no_output_____" ] ], [ [ "We'll demonstrate the main methods in action by analyzing a dataset that is an extract of the MIMIC III Database.\n\nLet's read the data (using `read_csv`), and take a look at the first 5 lines using the `head` method:", "_____no_output_____" ] ], [ [ "df = pd.read_csv('/home/shared/icu_2012.txt')\ndf.head()", "_____no_output_____" ] ], [ [ "<details>\n<summary>About printing DataFrames in Jupyter notebooks</summary>\n\nIn Jupyter notebooks, Pandas DataFrames are printed as these pretty tables seen above while `print(df.head())` is less nicely formatted.\nBy default, Pandas displays 20 columns and 60 rows, so, if your DataFrame is bigger, use the `set_option` function as shown in the example below:\n\n```python\npd.set_option('display.max_columns', 100)\npd.set_option('display.max_rows', 100)\n```\n\n</details>\n\nRecall that each row corresponds to one patient, an **instance**, and columns are **features** of this instance.", "_____no_output_____" ], [ "Let’s have a look at data dimensionality, feature names, and feature types.", "_____no_output_____" ] ], [ [ "print(df.shape)", "_____no_output_____" ] ], [ [ "From the output, we can see that the table contains 4000 rows and 79 columns.\n\nNow let's try printing out column names using `columns`:", "_____no_output_____" ] ], [ [ "print(df.columns)", "_____no_output_____" ] ], [ [ "We can use the `info()` method to output some general information about the dataframe: ", "_____no_output_____" ] ], [ [ "print(df.info())", "_____no_output_____" ] ], [ [ "`bool`, `int64`, `float64` and `object` are the data types of our features. We see that one feature is logical (`bool`), 3 features are of type `object`, and 16 features are numeric. With this same method, we can easily see if there are any missing values. Here, we can see that there are columns with missing variables because some columns contain less than the 4000 number of instances (or rows) we saw before with `shape`.", "_____no_output_____" ], [ "The `describe` method shows basic statistical characteristics of each numerical feature (`int64` and `float64` types): number of non-missing values, mean, standard deviation, range, median, 0.25 and 0.75 quartiles.", "_____no_output_____" ] ], [ [ "df.describe()", "_____no_output_____" ] ], [ [ "The `describe` methods only gives us information about numerical variables. Some of these don't really make sense, like the `subject_id` or `gender` but since they are numbers, we are getting summary statistics anyways.\n\nIn order to see statistics on non-numerical features, one has to explicitly indicate data types of interest in the `include` parameter. We would use `df.describe(include=['object', 'bool'])` but in this case, the dataset only has variables of type `int` and `float`.", "_____no_output_____" ], [ "For categorical (type `object`) and boolean (type `bool`) features we can use the `value_counts` method. This also woeks for variables that have been encoded into integers like Gender. Let's have a look at the distribution of `Gender`:", "_____no_output_____" ] ], [ [ "df['Gender'].value_counts()", "_____no_output_____" ] ], [ [ "Since Gender is encoded in the following way: (0: female, or 1: male)\n\n2246 intances are male patients", "_____no_output_____" ], [ "\n### Sorting\n\nA DataFrame can be sorted by the value of one of the variables (i.e columns). For example, we can sort by *Age* (use `ascending=False` to sort in descending order):\n", "_____no_output_____" ] ], [ [ "pd.set_option('display.max_columns', 100)\npd.set_option('display.max_rows', 100)\n\ndf.sort_values(by='Age', ascending=False).head()", "_____no_output_____" ] ], [ [ "We can also sort by multiple columns:", "_____no_output_____" ] ], [ [ "df.sort_values(by=['Age', 'Height'],\n ascending=[False, False]).head()", "_____no_output_____" ] ], [ [ "### Indexing and retrieving data\n\nA DataFrame can be indexed in a few different ways. \n\nTo get a single column, you can use a `DataFrame['Name']` construction. Let's use this to answer a question about that column alone: **what is the average maximum heart rate of admitted patients in our dataframe?**", "_____no_output_____" ] ], [ [ "df['HR_max'].mean()", "_____no_output_____" ] ], [ [ "106 bpm is slightly elevated, but it seems reasonable for an ICU population\n\n**Boolean indexing** with one column is also very convenient. The syntax is `df[P(df['Name'])]`, where `P` is some logical condition that is checked for each element of the `Name` column. The result of such indexing is the DataFrame consisting only of rows that satisfy the `P` condition on the `Name` column. \n\nLet's use it to answer the question:\n\n**What are average values of numerical features for male patients?**", "_____no_output_____" ] ], [ [ "df[df['Gender'] == 1].mean()", "_____no_output_____" ] ], [ [ "**What is the average Max Creatinine for patients female patients?**", "_____no_output_____" ] ], [ [ "df[df['Gender'] == 0]['Creatinine_max'].mean()", "_____no_output_____" ] ], [ [ "DataFrames can be indexed by column name (label) or row name (index) or by the serial number of a row. The `loc` method is used for **indexing by name**, while `iloc()` is used for **indexing by number**.\n\nIn the first case below, we say *\"give us the values of the rows with index from 0 to 5 (inclusive) and columns labeled from State to Area code (inclusive)\"*. In the second case, we say *\"give us the values of the first five rows in the first three columns\"* (as in a typical Python slice: the maximal value is not included).", "_____no_output_____" ] ], [ [ "df.loc[0:5, 'RecordID':'ICUType']", "_____no_output_____" ], [ "df.iloc[0:5, 0:3]", "_____no_output_____" ] ], [ [ "If we need the first or the last line of the data frame, we can use the `df[:1]` or `df[-1:]` construct:", "_____no_output_____" ] ], [ [ "df[-1:]", "_____no_output_____" ] ], [ [ "\n### Applying Functions to Cells, Columns and Rows\n\n**To apply functions to each column, use `apply()`:**\nIn this example, we will obtain the max value for each feature.\n", "_____no_output_____" ] ], [ [ "df.apply(np.max)", "_____no_output_____" ] ], [ [ "The `map` method can be used to **replace values in a column** by passing a dictionary of the form `{old_value: new_value}` as its argument. Let's change the values of female and male for the corresponding `strings`", "_____no_output_____" ] ], [ [ "d = {0 : 'Female', 1 : 'Male'}\ndf['Gender'] = df['Gender'].map(d)\ndf.head()", "_____no_output_____" ] ], [ [ "The same thing can be done with the `replace` method:", "_____no_output_____" ] ], [ [ "d2 = {1: 'Coronary Care Unit', 2: 'Cardiac Surgery Recovery Unit', 3: 'Medical ICU', 4: 'Surgical ICU'}\ndf = df.replace({'ICUType': d2})\ndf.head()", "_____no_output_____" ] ], [ [ "We can also replace missing values when it is necessary. For that we use the `filna()` methohd. In this case, we will replace them in the Mechanical Ventilation column. ", "_____no_output_____" ] ], [ [ "df['MechVent_min'].fillna(0, inplace=True)\ndf.head()", "_____no_output_____" ] ], [ [ "### Histograms\n\nHistograms are an important tool to understand the distribution of your variables. It can help you detect errors in the data, like extreme or unplausible values. ", "_____no_output_____" ] ], [ [ "df['Age'].hist()", "_____no_output_____" ] ], [ [ "We can quickly see that the distribution of age is not normal. Let's look at Na", "_____no_output_____" ] ], [ [ "df['Na_max'].hist()", "_____no_output_____" ] ], [ [ "Not a lot of resolution here. Let's increase the number of bins to 30", "_____no_output_____" ] ], [ [ "df['Na_max'].hist(bins=30)", "_____no_output_____" ] ], [ [ "Much better! It is easy to see that this is approximately a normal distribution.", "_____no_output_____" ], [ "\n### Grouping\n\nIn general, grouping data in Pandas works as follows:\n", "_____no_output_____" ], [ "\n```python\ndf.groupby(by=grouping_columns)[columns_to_show].function()\n```", "_____no_output_____" ], [ "\n1. First, the `groupby` method divides the `grouping_columns` by their values. They become a new index in the resulting dataframe.\n2. Then, columns of interest are selected (`columns_to_show`). If `columns_to_show` is not included, all non groupby clauses will be included.\n3. Finally, one or several functions are applied to the obtained groups per selected columns.\n\nHere is an example where we group the data according to `Gender` variable and display statistics of three columns in each group:", "_____no_output_____" ] ], [ [ "columns_to_show = ['Na_max', 'K_max', \n 'HCO3_max']\n\ndf.groupby(['Gender'])[columns_to_show].describe(percentiles=[])", "_____no_output_____" ] ], [ [ "Let’s do the same thing, but slightly differently by passing a list of functions to `agg()`:", "_____no_output_____" ] ], [ [ "columns_to_show = ['Na_max', 'K_max', \n 'HCO3_max']\n\ndf.groupby(['Gender'])[columns_to_show].agg([np.mean, np.std, np.min, \n np.max])", "_____no_output_____" ] ], [ [ "\n### Summary tables\n\nSuppose we want to see how the observations in our sample are distributed in the context of two variables - `Gender` and `ICUType`. To do so, we can build a **contingency table** using the `crosstab` method:\n\n", "_____no_output_____" ] ], [ [ "pd.crosstab(df['Gender'], df['ICUType'])", "_____no_output_____" ] ], [ [ "This will resemble **pivot tables** to those familiar with Excel. And, of course, pivot tables are implemented in Pandas: the `pivot_table` method takes the following parameters:\n\n* `values` – a list of variables to calculate statistics for,\n* `index` – a list of variables to group data by,\n* `aggfunc` – what statistics we need to calculate for groups, ex. sum, mean, maximum, minimum or something else.\n\nLet's take a look at the average number of day, evening, and night calls by area code:", "_____no_output_____" ] ], [ [ "df.pivot_table(['TroponinI_max', 'TroponinT_max'],\n ['ICUType'], aggfunc='mean')", "_____no_output_____" ] ], [ [ "Nothing surprising here, patients in the coronary/cardiac units have higher values of Troponins.\n\n### DataFrame transformations\n\nLike many other things in Pandas, adding columns to a DataFrame is doable in many ways.\n\nFor example, if we want to calculate the change in creatinine, let's create the `Delta_creatinine` Series and paste it into the DataFrame:\n\n", "_____no_output_____" ] ], [ [ "Delta_creatinine = df['Creatinine_max'] - df['Creatinine_min']\n\ndf.insert(loc=len(df.columns), column='Delta_creatinine', value=Delta_creatinine) \n# loc parameter is the number of columns after which to insert the Series object\n# we set it to len(df.columns) to paste it at the very end of the dataframe\ndf.head()", "_____no_output_____" ] ], [ [ "It is possible to add a column more easily without creating an intermediate Series instance:", "_____no_output_____" ] ], [ [ "df['Delta_BUN'] = df['BUN_max'] - df['BUN_min']\ndf.head()", "_____no_output_____" ] ], [ [ "To delete columns or rows, use the `drop` method, passing the required indexes and the `axis` parameter (`1` if you delete columns, and nothing or `0` if you delete rows). The `inplace` argument tells whether to change the original DataFrame. With `inplace=False`, the `drop` method doesn't change the existing DataFrame and returns a new one with dropped rows or columns. With `inplace=True`, it alters the DataFrame.", "_____no_output_____" ] ], [ [ "# get rid of just created columns\ndf.drop(['Delta_creatinine', 'Delta_BUN'], axis=1, inplace=True) \n# and here’s how you can delete rows\ndf.drop([1, 2]).head()", "_____no_output_____" ] ], [ [ "## 2. Exploring some associations\n\n\nLet's see how mechanical ventilation is related to Gender. We'll do this using a `crosstab` contingency table and also through visual analysis with `Seaborn`.\n", "_____no_output_____" ] ], [ [ "pd.crosstab(df['MechVent_min'], df['Gender'], margins=True)", "_____no_output_____" ], [ "# some imports to set up plotting \nimport matplotlib.pyplot as plt\n# pip install seaborn \nimport seaborn as sns\n# Graphics in retina format are more sharp and legible\n%config InlineBackend.figure_format = 'retina'", "_____no_output_____" ] ], [ [ "Now we create the plot that will show us the counts of mechanically ventilated patients by gender.", "_____no_output_____" ] ], [ [ "sns.countplot(x='Gender', hue='MechVent_min', data=df);", "_____no_output_____" ] ], [ [ "We see that th number (and probably the proportion) of mechanically ventilated patients is greater among males. \n\nNext, let's look at the same distribution but comparing the different ICU types: Let's also make a summary table and a picture.", "_____no_output_____" ] ], [ [ "pd.crosstab(df['ICUType'], df['MechVent_min'], margins=True)", "_____no_output_____" ], [ "sns.countplot(x='ICUType', hue='MechVent_min', data=df);", "_____no_output_____" ] ], [ [ "As you can see, the proportion of patients ventilated and not ventilated is very different across the different types of ICUs. That is particularly true in the cardiac surgery recovery unit. Can you think of a reason why that might be?", "_____no_output_____" ], [ "## 3. Some useful resources\n\n* [\"Merging DataFrames with pandas\"](https://nbviewer.jupyter.org/github/Yorko/mlcourse.ai/blob/master/jupyter_english/tutorials/merging_dataframes_tutorial_max_palko.ipynb) - a tutorial by Max Plako within mlcourse.ai (full list of tutorials is [here](https://mlcourse.ai/tutorials))\n* [\"Handle different dataset with dask and trying a little dask ML\"](https://nbviewer.jupyter.org/github/Yorko/mlcourse.ai/blob/master/jupyter_english/tutorials/dask_objects_and_little_dask_ml_tutorial_iknyazeva.ipynb) - a tutorial by Irina Knyazeva within mlcourse.ai\n* Main course [site](https://mlcourse.ai), [course repo](https://github.com/Yorko/mlcourse.ai), and YouTube [channel](https://www.youtube.com/watch?v=QKTuw4PNOsU&list=PLVlY_7IJCMJeRfZ68eVfEcu-UcN9BbwiX)\n* Official Pandas [documentation](http://pandas.pydata.org/pandas-docs/stable/index.html)\n* Course materials as a [Kaggle Dataset](https://www.kaggle.com/kashnitsky/mlcourse)\n* Medium [\"story\"](https://medium.com/open-machine-learning-course/open-machine-learning-course-topic-1-exploratory-data-analysis-with-pandas-de57880f1a68) based on this notebook\n* If you read Russian: an [article](https://habrahabr.ru/company/ods/blog/322626/) on Habr.com with ~ the same material. And a [lecture](https://youtu.be/dEFxoyJhm3Y) on YouTube\n* [10 minutes to pandas](http://pandas.pydata.org/pandas-docs/stable/10min.html)\n* [Pandas cheatsheet PDF](https://github.com/pandas-dev/pandas/blob/master/doc/cheatsheet/Pandas_Cheat_Sheet.pdf)\n* GitHub repos: [Pandas exercises](https://github.com/guipsamora/pandas_exercises/) and [\"Effective Pandas\"](https://github.com/TomAugspurger/effective-pandas)\n* [scipy-lectures.org](http://www.scipy-lectures.org/index.html) — tutorials on pandas, numpy, matplotlib and scikit-learn", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ] ]

4ad80fb3f44394403bcf56a2247843164e21b371

20,921

ipynb

Jupyter Notebook

PrimalCore/doc/homogeneous_table/mldataset_userguide.ipynb

imariasantos/primal

31443e0e7cc0b94e875e4da2a489053b7b93e554

[ "BSD-3-Clause" ]

2

2020-09-25T07:38:08.000Z

2020-11-22T06:22:30.000Z

PrimalCore/doc/homogeneous_table/mldataset_userguide.ipynb

imariasantos/primal

31443e0e7cc0b94e875e4da2a489053b7b93e554

[ "BSD-3-Clause" ]

8

2020-05-14T09:11:05.000Z

2021-01-28T10:02:07.000Z

PrimalCore/doc/homogeneous_table/mldataset_userguide.ipynb

imariasantos/primal

31443e0e7cc0b94e875e4da2a489053b7b93e554

[ "BSD-3-Clause" ]

2

2020-06-03T14:11:57.000Z

2021-01-14T16:17:52.000Z

26.684949

680

0.552555

[ [ [ ".. _MLDataSet_user_guide:\n", "_____no_output_____" ] ], [ [ "# Example of building a MLDataSet\n", "_____no_output_____" ] ], [ [ "We provide a simple workflow to build a fetures dataset using the class :class:`dataset.MLDataSet` (see for full API the module :mod:`~PrimalCore.homogeneous_table.dataset.MLDataSet`)\n\n.. currentmodule:: PrimalCore.homogeneous_table.dataset\n\n.. contents:: :local:\n\n.. toctree::\n", "_____no_output_____" ] ], [ [ "## Building a Features MLDataSet from a Table", "_____no_output_____" ] ], [ [ "we follow the approach in `Table_user_guide`_ to build a catalog using the :class:`PrimalCore.heterogeneous_table.table.Table`", "_____no_output_____" ] ], [ [ "from PrimalCore.heterogeneous_table.table import Table\nfrom ElementsKernel.Path import getPathFromEnvVariable", "_____no_output_____" ], [ "ph_catalog=getPathFromEnvVariable('PrimalCore/test_table.fits','ELEMENTS_AUX_PATH')", "_____no_output_____" ], [ "catalog=Table.from_fits_file(ph_catalog,fits_ext=0)", "| input data built\n| data Rows,Cols 1000 124\n" ], [ "catalog.keep_columns(['FLUX*','reliable_S15','STAR','AGN','MASKED','FLAG_PHOT'],regex=True)", "_____no_output_____" ] ], [ [ "We now build a Features dataset using the :class:`dataset.MLDataSet` class from the :mod:`PrimalCore.homogeneous_table.dataset` python module.", "_____no_output_____" ] ], [ [ "First we import the classes and the functions we need", "_____no_output_____" ] ], [ [ "from PrimalCore.homogeneous_table.dataset import MLDataSet", "_____no_output_____" ] ], [ [ ".. note::\n \n It is worth noting that for the :class:`MLDataSet` class, most of the functions to modify the dataset \n content have been implemented as functions in separate python modules. This is made on purpose, and shows a \n different approach compared to that used for the :class:`PrimalCore.heterogeneous_table.table.Table` class, where \n the functions are implemented as methods of the same class.", "_____no_output_____" ], [ "To build a MLDataSet directly from a Table we can use the classmethod :func:`MLDataSet.new_from_table`", "_____no_output_____" ] ], [ [ "dataset=MLDataSet.new_from_table(catalog)", "| building features\n| features built\n| Rows,Cols 1000 50\n" ], [ "print dataset.features_names", "['FLUX_G_1', 'FLUX_G_2', 'FLUX_G_3', 'FLUX_R_1', 'FLUX_R_2', 'FLUX_R_3', 'FLUX_I_1', 'FLUX_I_2', 'FLUX_I_3', 'FLUX_VIS', 'FLUX_Z_1', 'FLUX_Z_2', 'FLUX_Z_3', 'FLUX_Y_1', 'FLUX_Y_2', 'FLUX_Y_3', 'FLUX_J_1', 'FLUX_J_2', 'FLUX_J_3', 'FLUX_H_1', 'FLUX_H_2', 'FLUX_H_3', 'FLUXERR_G_1', 'FLUXERR_G_2', 'FLUXERR_G_3', 'FLUXERR_R_1', 'FLUXERR_R_2', 'FLUXERR_R_3', 'FLUXERR_I_1', 'FLUXERR_I_2', 'FLUXERR_I_3', 'FLUXERR_VIS', 'FLUXERR_Z_1', 'FLUXERR_Z_2', 'FLUXERR_Z_3', 'FLUXERR_Y_1', 'FLUXERR_Y_2', 'FLUXERR_Y_3', 'FLUXERR_J_1', 'FLUXERR_J_2', 'FLUXERR_J_3', 'FLUXERR_H_1', 'FLUXERR_H_2', 'FLUXERR_H_3', 'FLAG_PHOT', 'FLUX_RADIUS_DETECT', 'MASKED', 'reliable_S15', 'STAR', 'AGN']\n" ] ], [ [ ".. note::\n as you can see, the **__original_entry_ID__** is not present among the features, indeed it is used only to track \n the original catalog IDs, but it is present as separate member of the dataset (we print onlty the 10 first\n elements)", "_____no_output_____" ] ], [ [ "print dataset.features_original_entry_ID[1:10]", "[1 2 3 4 5 6 7 8 9]\n" ] ], [ [ "and in this way it **safely** can not be used as a feature.\n", "_____no_output_____" ], [ "## Building a Features MLDataSet from a FITS file", "_____no_output_____" ] ], [ [ "To build a MLDataSet directly from a FITS file we can use the classmethod :func:`MLDataSet.new_from_fits_file`", "_____no_output_____" ] ], [ [ "dataset_from_file=MLDataSet.new_from_fits_file(ph_catalog,fits_ext=0,\\\n use_col_names_list=['FLUX*','reliable_S15','STAR','AGN','MASKED','FLAG_PHOT'],\\\n regex=True)", "| input data built\n| data Rows,Cols 1000 124\n| building features\n| features built\n| Rows,Cols 1000 50\n" ], [ "print dataset_from_file.features_names", "['FLUX_G_1', 'FLUX_G_2', 'FLUX_G_3', 'FLUX_R_1', 'FLUX_R_2', 'FLUX_R_3', 'FLUX_I_1', 'FLUX_I_2', 'FLUX_I_3', 'FLUX_VIS', 'FLUX_Z_1', 'FLUX_Z_2', 'FLUX_Z_3', 'FLUX_Y_1', 'FLUX_Y_2', 'FLUX_Y_3', 'FLUX_J_1', 'FLUX_J_2', 'FLUX_J_3', 'FLUX_H_1', 'FLUX_H_2', 'FLUX_H_3', 'FLUXERR_G_1', 'FLUXERR_G_2', 'FLUXERR_G_3', 'FLUXERR_R_1', 'FLUXERR_R_2', 'FLUXERR_R_3', 'FLUXERR_I_1', 'FLUXERR_I_2', 'FLUXERR_I_3', 'FLUXERR_VIS', 'FLUXERR_Z_1', 'FLUXERR_Z_2', 'FLUXERR_Z_3', 'FLUXERR_Y_1', 'FLUXERR_Y_2', 'FLUXERR_Y_3', 'FLUXERR_J_1', 'FLUXERR_J_2', 'FLUXERR_J_3', 'FLUXERR_H_1', 'FLUXERR_H_2', 'FLUXERR_H_3', 'FLUX_RADIUS_DETECT', 'reliable_S15', 'STAR', 'AGN', 'MASKED', 'FLAG_PHOT']\n" ] ], [ [ "## Columns selection\n### using `use_col_names_list` in the factories", "_____no_output_____" ] ], [ [ "Columns can be selected using the `use_col_names_list` parameter in the classmethod factories :func:`MLDataSet.new_from_table` and :func:`MLDataSet.new_from_fits_file`", "_____no_output_____" ] ], [ [ "dataset=MLDataSet.new_from_table(catalog,use_col_names_list=['FLUX*','reliable_S15','STAR','AGN','MASKED','FLAG_PHOT'],\\\n regex=True)", "| building features\n| features built\n| Rows,Cols 1000 50\n" ], [ "print dataset.features_names", "['FLUX_G_1', 'FLUX_G_2', 'FLUX_G_3', 'FLUX_R_1', 'FLUX_R_2', 'FLUX_R_3', 'FLUX_I_1', 'FLUX_I_2', 'FLUX_I_3', 'FLUX_VIS', 'FLUX_Z_1', 'FLUX_Z_2', 'FLUX_Z_3', 'FLUX_Y_1', 'FLUX_Y_2', 'FLUX_Y_3', 'FLUX_J_1', 'FLUX_J_2', 'FLUX_J_3', 'FLUX_H_1', 'FLUX_H_2', 'FLUX_H_3', 'FLUXERR_G_1', 'FLUXERR_G_2', 'FLUXERR_G_3', 'FLUXERR_R_1', 'FLUXERR_R_2', 'FLUXERR_R_3', 'FLUXERR_I_1', 'FLUXERR_I_2', 'FLUXERR_I_3', 'FLUXERR_VIS', 'FLUXERR_Z_1', 'FLUXERR_Z_2', 'FLUXERR_Z_3', 'FLUXERR_Y_1', 'FLUXERR_Y_2', 'FLUXERR_Y_3', 'FLUXERR_J_1', 'FLUXERR_J_2', 'FLUXERR_J_3', 'FLUXERR_H_1', 'FLUXERR_H_2', 'FLUXERR_H_3', 'FLUX_RADIUS_DETECT', 'reliable_S15', 'STAR', 'AGN', 'MASKED', 'FLAG_PHOT']\n" ] ], [ [ "### using dataset_handler fucntions", "_____no_output_____" ] ], [ [ "Or, columns can be selected using specific selection functions, from the :mod:`~PrimalCore.homogeneous_table.dataset_handler` module", "_____no_output_____" ] ], [ [ "from PrimalCore.homogeneous_table.dataset_handler import drop_features\nfrom PrimalCore.homogeneous_table.dataset_handler import keep_features", "_____no_output_____" ] ], [ [ "for example we decide to drop columns with names matching expression \"FLUX\\*1\\*\" by using the :func:`~PrimalCore.homogeneous_table.dataset_handler.drop_features`", "_____no_output_____" ] ], [ [ "drop_features(dataset,['FLUX*1*'])\ndataset.features_names", "| features initial Rows,Cols= 1000 50\n| removing features ['FLUX_G_1', 'FLUX_R_1', 'FLUX_I_1', 'FLUX_Z_1', 'FLUX_Y_1', 'FLUX_J_1', 'FLUX_H_1', 'FLUXERR_G_1', 'FLUXERR_R_1', 'FLUXERR_I_1', 'FLUXERR_Z_1', 'FLUXERR_Y_1', 'FLUXERR_J_1', 'FLUXERR_H_1']\n| features final Rows,Cols= 1000 36\n\n" ] ], [ [ "Furtger we can decide to keep only columns with names matching the regular expression \"FLUX\\*2\\*\" by using the :func:`~PrimalCore.homogeneous_table.dataset_handler.keep_features` function from the :mod:`PrimalCore.homogeneous_table.dataset_handler` package", "_____no_output_____" ] ], [ [ "keep_features(dataset,['FLUX*2*'],regex=True)\nprint dataset.features_names", "| features initial Rows,Cols= 1000 36\n| removing features ['FLUX_G_3', 'FLUX_R_3', 'FLUX_I_3', 'FLUX_VIS', 'FLUX_Z_3', 'FLUX_Y_3', 'FLUX_J_3', 'FLUX_H_3', 'FLUXERR_G_3', 'FLUXERR_R_3', 'FLUXERR_I_3', 'FLUXERR_VIS', 'FLUXERR_Z_3', 'FLUXERR_Y_3', 'FLUXERR_J_3', 'FLUXERR_H_3', 'FLUX_RADIUS_DETECT', 'reliable_S15', 'STAR', 'AGN', 'MASKED', 'FLAG_PHOT']\n| features final Rows,Cols= 1000 14\n\n['FLUX_G_2', 'FLUX_R_2', 'FLUX_I_2', 'FLUX_Z_2', 'FLUX_Y_2', 'FLUX_J_2', 'FLUX_H_2', 'FLUXERR_G_2', 'FLUXERR_R_2', 'FLUXERR_I_2', 'FLUXERR_Z_2', 'FLUXERR_Y_2', 'FLUXERR_J_2', 'FLUXERR_H_2']\n" ] ], [ [ "## Adding features", "_____no_output_____" ] ], [ [ "And finally we can add a new feature with the :func:`~PrimalCore.homogeneous_table.dataset_handler.add_features` function\nWe can add a single feature:", "_____no_output_____" ] ], [ [ "from PrimalCore.homogeneous_table.dataset_handler import add_features\n\ntest_feature=dataset.get_feature_by_name('FLUXERR_H_2')**2\nadd_features(dataset,'test',test_feature)\ndataset.features_names", "_____no_output_____" ] ], [ [ "Or we can add a 2dim array of features\n", "_____no_output_____" ] ], [ [ "test_feature_2dim=np.zeros((dataset.features_N_rows,5))\ntest_feature_2dim_names=['a','b','c','d','e']\nadd_features(dataset,test_feature_2dim_names,test_feature_2dim)\ndataset.features_names", "_____no_output_____" ] ], [ [ "We can think to a more meaningful example, i.e. we want to add flux ratios. Lets start by defining the list of \ncontigous bands, for the flux evaluation", "_____no_output_____" ] ], [ [ "flux_bands_list_2=['FLUX_G_2','FLUX_R_2','FLUX_I_2','FLUX_Z_2','FLUX_Y_2','FLUX_J_2','FLUX_VIS','FLUX_VIS','FLUX_VIS']\nflux_bands_list_1=['FLUX_R_2','FLUX_I_2','FLUX_Z_2','FLUX_Y_2','FLUX_J_2','FLUX_H_2','FLUX_Y_2','FLUX_J_2','FLUX_H_2']", "_____no_output_____" ] ], [ [ "we import the module where we have defined the FluxRatio class (:mod:`PrimalCore.phz_tools.photometry`)", "_____no_output_____" ] ], [ [ "from PrimalCore.phz_tools.photometry import FluxRatio", "_____no_output_____" ], [ "for f1,f2 in zip(flux_bands_list_1,flux_bands_list_2):\n f1_name=f1.split('_')[1]\n f2_name=f2.split('_')[1]\n if f1 in dataset.features_names and f2 in dataset.features_names:\n f=FluxRatio('F_%s'%(f2_name+'-'+f1_name),f1,f2,features=dataset)\n add_features(dataset,f.name,f.values)", "/Users/orion/Work/Projects/Primal/PrimalCore/python/PrimalCore/phz_tools/photometry.py:110: RuntimeWarning: divide by zero encountered in true_divide\n return features.get_feature_by_name(band_2)/features.get_feature_by_name(band_1)\n/Users/orion/Work/Projects/Primal/PrimalCore/python/PrimalCore/phz_tools/photometry.py:110: RuntimeWarning: invalid value encountered in true_divide\n return features.get_feature_by_name(band_2)/features.get_feature_by_name(band_1)\n" ] ], [ [ ".. note::\n Note that in this example we skipped the selection CLEAN=\" (FLAG_PHOT == 0) & (MASKED == 0) & (STAR == 0) & \n (AGN == 0) & (reliable_S15==1)\", so we have entries with flux values that are zero, and this results in the \n corresponding warning messge due to zero division", "_____no_output_____" ] ], [ [ "dataset.features_names", "_____no_output_____" ] ], [ [ "## Operations on rows\n### filtering NaN/Inf with dataset_preprocessing functions", "_____no_output_____" ] ], [ [ "We can get rid of the NAN/INF rows using the :func:`~PrimalCore.preprocessing.dataset_preprocessing.drop_nan_inf` function", "_____no_output_____" ] ], [ [ "from PrimalCore.preprocessing.dataset_preprocessing import drop_nan_inf", "_____no_output_____" ], [ "drop_nan_inf(dataset)", "| features cleaning for nan/inf\n| features initial Rows,Cols= 1000 26\n| features initial Rows,Cols= 1000 26\n| removing features []\n| features final Rows,Cols= 1000 26\n\n|removed columns []\n|removed rows 468\n| features cleaned Rows,Cols= 532 26\n\n" ] ] ]

[ "raw", "markdown", "raw", "markdown", "raw", "code", "raw", "markdown", "code", "raw", "code", "raw", "code", "markdown", "raw", "code", "markdown", "raw", "code", "markdown", "raw", "code", "raw", "code", "raw", "code", "markdown", "raw", "code", "markdown", "code", "markdown", "code", "raw", "code", "raw", "code", "markdown", "raw", "code" ]

[ [ "raw" ], [ "markdown" ], [ "raw" ], [ "markdown" ], [ "raw" ], [ "code", "code", "code", "code" ], [ "raw" ], [ "markdown" ], [ "code" ], [ "raw", "raw" ], [ "code", "code" ], [ "raw" ], [ "code" ], [ "markdown", "markdown" ], [ "raw" ], [ "code", "code" ], [ "markdown" ], [ "raw" ], [ "code", "code" ], [ "markdown" ], [ "raw" ], [ "code" ], [ "raw" ], [ "code" ], [ "raw" ], [ "code" ], [ "markdown" ], [ "raw" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "raw" ], [ "code", "code" ], [ "raw" ], [ "code" ], [ "markdown" ], [ "raw" ], [ "code", "code" ] ]

4ad813b319ab3025e4bdd20ad9cb3d3059c9acc6

50,728

ipynb

Jupyter Notebook

16_analisis_econometrico.ipynb

PythonistaMX/pyr101

53026bf7044c189c0a2e68a5971b6661ce5e579a

[ "MIT" ]

null

16_analisis_econometrico.ipynb

PythonistaMX/pyr101

53026bf7044c189c0a2e68a5971b6661ce5e579a

[ "MIT" ]

null

16_analisis_econometrico.ipynb

PythonistaMX/pyr101

53026bf7044c189c0a2e68a5971b6661ce5e579a

[ "MIT" ]

null

58.712963

15,696

0.730504

[ [ [ "<a href=\"https://www.pythonista.io\"> <img src=\"img/pythonista.png\"></a>", "_____no_output_____" ], [ "## Análisis econométrico.\n\nUn análisis econométrico consiste en la aplicaciónde técnicas estadísticas para poder crear modelos capaces de predecir con cierto grado de confianza los fenoménos y observados.\n\nhttps://economipedia.com/definiciones/modelo-econometrico.html", "_____no_output_____" ], [ "## Regresión lineal simple.", "_____no_output_____" ] ], [ [ "Data <- read.csv(\"data/16/Regresion.csv\")", "_____no_output_____" ], [ "Data", "_____no_output_____" ], [ "XPrice = Data$AP", "_____no_output_____" ], [ "YPrice = Data$NASDAQ", "_____no_output_____" ], [ "plot(XPrice, YPrice, \nxlab=\"Apple\", \nylab=\"NASDAQ\",\npch =19)", "_____no_output_____" ] ], [ [ "### La función ```lm()```.", "_____no_output_____" ] ], [ [ "help(lm)", "_____no_output_____" ], [ "LinearR.lm = lm(YPrice ~ XPrice, data=Data)", "_____no_output_____" ] ], [ [ "### La función ```coefficients()```.", "_____no_output_____" ] ], [ [ "help(coefficients)", "_____no_output_____" ], [ "coeffs = coefficients(LinearR.lm) \ncoeffs", "_____no_output_____" ] ], [ [ "* La siguiente aplicará la ecuación lineal usando los coeficientes obtenidos en función de ```XPrice```.", "_____no_output_____" ] ], [ [ "YPrice = 4124.10322215869 + 63.954703332101*(XPrice)", "_____no_output_____" ], [ "summary(LinearR.lm)$r.squared", "_____no_output_____" ], [ "summary(LinearR.lm)", "_____no_output_____" ] ], [ [ "[Residuals](https://www.statisticshowto.com/residual/#:~:text=A%20residual%20is%20the%20vertical,are%20below%20the%20regression%20line.&text=In%20other%20words%2C%20the%20residual,explained%20by%20the%20regression%20line)", "_____no_output_____" ] ], [ [ "Predictdata = data.frame(XPrice=75)", "_____no_output_____" ], [ "Predictdata", "_____no_output_____" ] ], [ [ "### La función ```predict()```.", "_____no_output_____" ] ], [ [ "help(predict)", "_____no_output_____" ], [ "predict(LinearR.lm, Predictdata, interval=\"confidence\") ", "_____no_output_____" ] ], [ [ "### la función ```resid()```.", "_____no_output_____" ] ], [ [ "help(resid)", "_____no_output_____" ], [ "LinearR.res = resid(LinearR.lm)", "_____no_output_____" ], [ "summary(LinearR.res)", "_____no_output_____" ], [ "plot(XPrice, LinearR.res, \nylab=\"Residuals\", xlab=\"XPrice\", \nmain=\"Residual Plot\")", "_____no_output_____" ] ], [ [ "## Regresión lineal múltiple.", "_____no_output_____" ] ], [ [ "Data <- read.csv(\"data/16/RegresionMultiple.csv\")", "_____no_output_____" ], [ "Date <- Data$Date", "_____no_output_____" ], [ "StockYPrice <- Data$NASDAQ", "_____no_output_____" ], [ "StockX1Price <- Data$AAPL", "_____no_output_____" ], [ "StockX2Price <- Data$IBM", "_____no_output_____" ], [ "StockX3Price <- Data$MSFT", "_____no_output_____" ], [ "MultipleR.lm = lm(StockYPrice ~ StockX1Price + StockX2Price + StockX3Price, data=Data[2:5])\nsummary(MultipleR.lm)", "_____no_output_____" ], [ "newdata = data.frame(StockX1Price=120, StockX2Price=120, StockX3Price=213)", "_____no_output_____" ], [ "predict(MultipleR.lm, newdata)", "_____no_output_____" ], [ "summary(MultipleR.lm)$r.squared ", "_____no_output_____" ], [ "summary(MultipleR.lm)$adj.r.squared ", "_____no_output_____" ], [ "predict(MultipleR.lm, newdata, interval=\"confidence\")", "_____no_output_____" ] ], [ [ "<a rel=\"license\" href=\"http://creativecommons.org/licenses/by/4.0/\"><img alt=\"Licencia Creative Commons\" style=\"border-width:0\" src=\"https://i.creativecommons.org/l/by/4.0/80x15.png\" /></a> Esta obra está bajo una <a rel=\"license\" href=\"http://creativecommons.org/licenses/by/4.0/\">Licencia Creative Commons Atribución 4.0 Internacional</a>.\n© José Luis Chiquete Valdivieso. 2020.", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown", "markdown", "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ] ]

4ad82003a07f3074c7fbdb53ffb3bf3abe5d8f2d

134,019

ipynb

Jupyter Notebook

week2/FTE/2_Probability.ipynb

kscully-dotcom/msds650

3c03e83cc43502169a4aa1649b7e1312051b4e26

[ "MIT" ]

null

week2/FTE/2_Probability.ipynb

kscully-dotcom/msds650

3c03e83cc43502169a4aa1649b7e1312051b4e26

[ "MIT" ]

null

week2/FTE/2_Probability.ipynb

kscully-dotcom/msds650

3c03e83cc43502169a4aa1649b7e1312051b4e26

[ "MIT" ]

null

93.785164

17,772

0.821563

[ [ [ "# Week 2 -- Probability ", "_____no_output_____" ], [ "<img align=\"right\" style=\"padding-right:10px;\" src=\"figures_wk2/stats_cover.png\" width=200> \n\n**Resources and References**\n>**Practical Statistics for Data Scientists, 2nd Edition** \n>by Peter Bruce, Andrew Bruce, Peter Gedeck \n>Publisher: O'Reilly Media, Inc. \n>Release Date: May 2020 \n>ISBN: 9781492072942 \n\n \n \n\n\n>**Probability for Machine Learning** \n>by Jason Brownle \n>https://machinelearningmastery.com/probability-for-machine-learning/\n<img align=\"right\" style=\"padding-right:10px;\" src=\"figures_wk2/probability_cover.png\" width=200> ", "_____no_output_____" ], [ "## Data Sampling and Distribution", "_____no_output_____" ], [ "### Bias and Random Sampling", "_____no_output_____" ], [ "**Sample**-- subset of data taken from larger data set (usually called a **Population.** NOTE: Different from a population in biology). \n**Population** -- Larger data set (real or theoretical). \n**N(n)** -- size of population or sample. \n**Random Sampling** -- Create a sample by randomly drawing elements from population. \n**Bias** -- Systemic error \n**Sample Bias** -- Sample that misrepresents the population. \n* Recent example: 2016 US. Presidential election polls that placed Hillary Clinton ahead of Donald Trump. **Sample bias** was one of the contributing factors to the incorrect predictions. (Source: \"Harvard Researchers Warn 2016 Polling Mistakes Serve as a 'Cautionary Tale' in 2020\" retrieved from https://www.thecrimson.com/article/2020/11/2/2016-election-polls-kuriwaki-isakov/)\n", "_____no_output_____" ], [ "#### Bias", "_____no_output_____" ], [ "Error due to bias represents something wrong with the data collection or selection system itself. In the \"Practical Statistics for Data Scientists\" book referenced above, the authors use the analogy of two guns shooting at a target X-Y axis:\n\n<table style=\"font-size: 20px\">\n <tr>\n <th>True Aim</th><th>Biased Aim</th>\n </tr>\n <tr>\n <td><img src=\"figures_wk2/true_aim.png\"></td><td><img src=\"figures_wk2/bias_aim.png\"></td>\n </tr>\n</table>\n\nThe \"True Aim\" picture shows us the result of random errors whereas the pattern we see in the \"Biased Aim\" graph ", "_____no_output_____" ], [ "#### Selection and Self-selection bias", "_____no_output_____" ], [ "**Selection bias**: Refers to choosing data favorable to a particular conclusion, whether done deliberately or accidentally. \n\n**Self-selection bias**: Product or place reviews on social media or \"review sites\" like Yelp are not a good source of sample data. These types of reviews are not random -- rather, reviewers typically have a reason for self-selecting. Many times due to either a very good or very bad experience and thus represents a biased sample. \n\nIt is worth noting that most non-compulsory surveys suffer from this same bias. Think of end of course surveys. Only a small number of course attendees usually take the time and effort to fill out a survey, and then usually only due to an extremely good or extremely bad course experience. ", "_____no_output_____" ], [ "#### Random Selection", "_____no_output_____" ], [ "George Gallup proposed random selection as a scientific sampling method after the *Literary Digest* poll of 1936 famously predicted the incorrect outcome of Alf Landon winning the presidential election over Franklin Roosevelt. \n\n**Population** \nA vital point is to correctly define the population from which the sample will be drawn. For example: \n* Surveying 100 random customers to walk in the door of a grocery store may yield an acceptible sample for learning public opinion about general products. \n* Surveying 100 random men to walk in the grocery store about feminine hygiene products will probably yield a less than optimal result.\n\nData quality and appropriate sampling is often more important than data quantity. ", "_____no_output_____" ], [ "### Sampling Distribution", "_____no_output_____" ], [ "**Data Distribution:** Distribution of a sample's *individual data points*.\n\n**Sampling Distribution:** Distribution of a sample statistic, such as mean. Tends to be more regular and bell-shaped than the data itself. \n\nBelow is an example of this, recreated from *Practical Statistics for Data Scientists, 2nd Edition*, using Lending Club data. ", "_____no_output_____" ] ], [ [ "%matplotlib inline\n\nfrom pathlib import Path\nimport pandas as pd\nimport numpy as np\nfrom scipy import stats\nfrom sklearn.utils import resample\n\nimport seaborn as sns\nimport matplotlib.pylab as plt\n\nsns.set()", "_____no_output_____" ], [ "loans_income = pd.read_csv(\"data/loans_income.csv\", squeeze=True)\n\nsample_data = pd.DataFrame({\n 'income': loans_income.sample(1000),\n 'type': 'Data',\n})\n\nsample_mean_05 = pd.DataFrame({\n 'income': [loans_income.sample(5).mean() for _ in range(1000)],\n 'type': 'Mean of 5',\n})\n\nsample_mean_20 = pd.DataFrame({\n 'income': [loans_income.sample(20).mean() for _ in range(1000)],\n 'type': 'Mean of 20',\n})\n\nresults = pd.concat([sample_data, sample_mean_05, sample_mean_20])\nprint(results.head())", " income type\n38798 51000.0 Data\n47991 45000.0 Data\n9912 62000.0 Data\n29199 26000.0 Data\n27846 55000.0 Data\n" ], [ "g = sns.FacetGrid(results, col='type', col_wrap=3, \n height=4, aspect=1)\ng.map(plt.hist, 'income', range=[0, 200000], bins=40)\ng.set_axis_labels('Income', 'Count')\ng.set_titles('{col_name}')\n\nplt.tight_layout()\nplt.show()", "_____no_output_____" ] ], [ [ "* The first graph is the mean of 1000 values.\n* The second graph is 1000 means of 5 values.\n* The third graph is 1000 means of 20 values.", "_____no_output_____" ], [ "#### Central Limit Theorem", "_____no_output_____" ], [ "The **central limit theorem** states that means of multiple samples will be a bell-shaped curve, even if the population isn't normally distributed, if sample size is large enough and not too far off of normal.", "_____no_output_____" ], [ "### Normal (Gaussian) Distribution", "_____no_output_____" ], [ "**Standard Normal Distribution**\n<img style=\"padding-right:10px;\" src=\"figures_wk2/normal_distribution.png\"> \n\n---\n\n$ \\mu = $ The population mean.\n\n\nMany statistical tests, **such as t-distributions and hypothesis testing**, assume sample statistics are normally distributed. Simple mathematics exist to compare data to a standard normal distribution, however, for our purposes, a QQ-plot is faster and easier. \n\nNormality can be checked with a **QQ-plot**. Python's *scipy* package has a QQ-plot function, called `probplot`, seen below:", "_____no_output_____" ] ], [ [ "fig, ax = plt.subplots(figsize=(4, 4))\n\nnorm_sample = stats.norm.rvs(size=100)\nstats.probplot(norm_sample, plot=ax)\n\nplt.tight_layout()\nplt.show()", "_____no_output_____" ] ], [ [ "The blue markers represent *z-scores*, or standardized data points, plotted vs. standard deviations away from the mean. \n", "_____no_output_____" ], [ "### Long-tailed Distributions", "_____no_output_____" ], [ "**Tail:** A long, narrow area of a frequency distribution where extreme cases happen with low frequency. \n**Skew:** Where one tail of a distribution is longer than another.\n\n\nDespite the time and effort spent teaching about normal distributions, most data is **not** normally distributed.\n\nAn example can be seen with a QQ-plot of Netflix stock data. ", "_____no_output_____" ] ], [ [ "sp500_px = pd.read_csv('data/sp500_data.csv.gz')\n\nnflx = sp500_px.NFLX\n# nflx = np.diff(np.log(nflx[nflx>0]))\n\nfig, ax = plt.subplots(figsize=(4, 4))\nstats.probplot(nflx, plot=ax)\n\nplt.tight_layout()\nplt.show()", "_____no_output_____" ], [ "sp500_px.head()", "_____no_output_____" ] ], [ [ "Low values are below the line and high values are above the line. This tells us the data is not normally distributed and we are more likely to see extreme values than if it was normally distributed.\n\nSometimes non-regular data can be normalized using methods like **taking the logarithm of values greater than 0.**", "_____no_output_____" ] ], [ [ "nflx = np.diff(np.log(nflx[nflx>0]))\n\nfig, ax = plt.subplots(figsize=(4, 4))\nstats.probplot(nflx, plot=ax)\n\nplt.tight_layout()\nplt.show()", "_____no_output_____" ] ], [ [ "You can see that helps significantly, but the data is still not very normal. ", "_____no_output_____" ], [ "### Binomial (Bernoulli) Distribution", "_____no_output_____" ], [ "A **binomial outcome** is one for which there are only two possible answers: \n* yes / no \n* true / false \n* buy / don't buy \n* click / don't click \n* etc.\n\nAt its heart, binomial distributions analyze the probability of each outcome under certain conditions. \n\nThe classic example is the coin toss. The outcome will be either heads (H) or tails (T) for any particular toss. \n\nA **trial** is an event of interest with a discrete outcome (e.g. a coin toss).\n\nA **success** is defined as the outcome of interest in the trials. For example, in the coin toss above, we could say we are interested in the number of H outcomes out of 10 trials (tosses). Each H outcome would be a *success*. Also represented as a \"1\" (following binary logic).\n\nA **binary distribution** is the number of successes (*x*) in *n* trials with *p* probability of success for each trial. Also called a *Bernoulli distribution*.\n\n", "_____no_output_____" ], [ "#### Calculating Binomial Probabilies", "_____no_output_____" ], [ "In general, we are concerned with calculating two situations:\n\n1. The probability of *x* successes out of *n* trials. This is called the **probability mass function(pmf).** \n2. The probability of **no more than** _x_ successes out of *n* trials. This is called the **cumulative distribution function (cdf).**\n\nPython uses scipy's `stats.binom.pmf()` and `stats.binom.cdf()` functions, respectively, for that functionality. \n\n**pmf example:** A fair coin has a 50% (.50) chance of coming up heads on a toss. What is the probability of getting a head (H) 7 times out of 10 tosses?\n\nx = 7 \nn = 10 \np = 0.5 ", "_____no_output_____" ] ], [ [ "stats.binom.pmf(7, n=10, p=0.5)", "_____no_output_____" ] ], [ [ "So, there is an 11.7% chance that a coin will land on heads 7 out of 10 tosses.\n\n**cdf example:** Using that same fair coin, what is the probability of getting a head (H) **_no more than_** four times?\n\nx = 4 \nn = 10 \np = 0.5 ", "_____no_output_____" ] ], [ [ "stats.binom.cdf(4, n=10, p=0.5)", "_____no_output_____" ] ], [ [ "There is a 37.6% chance that there will be 4 or fewer heads in 10 trials. Which is the same thing as \n\n`(chance of 0 H) + (chance of 1 H) + (chance of 2 H) + (chance of 3 H) + (chance of 4 H)`", "_____no_output_____" ] ], [ [ "stats.binom.pmf(0, n=10, p=0.5) + stats.binom.pmf(1, n=10, p=0.5) \\\n+ stats.binom.pmf(2, n=10, p=0.5) + stats.binom.pmf(3, n=10, p=0.5) \\\n+ stats.binom.pmf(4, n=10, p=0.5)", "_____no_output_____" ] ], [ [ "There are many other useful data distributions. Students are encouraged to to independently research them. ", "_____no_output_____" ], [ "# Bootstrapping", "_____no_output_____" ], [ "**Bootstrap sample:** A sample taken with replacement from a data set. \n**Resampling:** The process of taking repeated samples from observed data. \n\nHypothesis testing requires some estimate of the sampling distribution. \"Traditional\" hypothesis testing requires formulas to create estimates of sampling distributions. *Bootstrapping* creates a sampling distribution through resampling. \n \nLet's take a look. First, we'll find the median income of the Lending Club data.", "_____no_output_____" ] ], [ [ "loans_income.median()", "_____no_output_____" ] ], [ [ "Next, we'll use scikit-learn's `resample()` function to take 5 samples and print out the median of each.", "_____no_output_____" ] ], [ [ "for _ in range(5):\n sample = resample(loans_income)\n print(sample.median())", "62000.0\n62000.0\n61200.0\n61481.0\n62000.0\n" ] ], [ [ "As you can see, the median is different for each sample. \n\nLet's take 1000 samples and average the medians and see how different it is from the dataset median.", "_____no_output_____" ] ], [ [ "results = []\nfor nrepeat in range(1000):\n sample = resample(loans_income)\n results.append(sample.median())\nresults = pd.Series(results)\nprint('Bootstrap Statistics:')\nprint(f'original: {loans_income.median()}')\nprint(f'mean of medians: {results.mean()}')\nprint(f'bias: {results.mean() - loans_income.median()}')\nprint(f'std. error: {results.std()}')", "Bootstrap Statistics:\noriginal: 62000.0\nmean of medians: 61927.7055\nbias: -72.29450000000361\nstd. error: 213.41770928952386\n" ] ], [ [ "Let's use bootstrapping on that wonky Netflix data. We'll take samples of 100 and store the mean of the sample in a list and do that 20,000 times. \n\nYou'll notice that the data looks much more normally distributed even without the logarithm trick.", "_____no_output_____" ] ], [ [ "len(results)", "_____no_output_____" ], [ "results[90]", "_____no_output_____" ], [ "sp500_px = pd.read_csv('data/sp500_data.csv.gz')\n\nnflx = sp500_px.NFLX.values\n\nmeans = []\n\nfor _ in range(20000): \n mean = resample(nflx, replace=True, n_samples=100).mean()\n means.append(mean)\n \nplt.hist(means)\n\nfig, ax = plt.subplots(figsize=(4, 4))\nstats.probplot(means, plot=ax)\n\nplt.tight_layout()\nplt.show()", "_____no_output_____" ] ], [ [ "# Probability", "_____no_output_____" ], [ "\n**Joint probability:** Probability of two or more events happening at the same time. \n**Marginal probability:** Probability of an event regardless of other variables outcome. \n**Conditional probability:** Probability of an event occurring along with one or more other events. ", "_____no_output_____" ], [ "## Probability for one random variable ", "_____no_output_____" ], [ "Probability shows the likelihood of an event happening. \n\nProbability of one random variable is the likelihood of an event that is independent of other factors. Examples include: \n* Coin toss. \n* Roll of a dice. \n* Drawing one card from a deck of cards. \n\nFor random variable `x`, the function `P(x)` relates probabilities to all values of `x`.\n\n<center>$Probability\\ Density\\ of\\ x = P(x)$</center>\n\nIf `A` is a specific event of `x`, \n\n<center>$Probability\\ of\\ Event\\ A = P(A)$</center>", "_____no_output_____" ], [ "Probability of an event is calculated as *the number of desired outcomes* divided by *total number of possible outcomes*, where all outcomes are equally likely:\n\n<center>$Probability = \\frac{the\\ number\\ of\\ desired\\ outcomes}{total\\ number\\ of\\ possible\\ outcomes}$</center>\n\nIf we apply that principle to our examples above: \n* Coin toss: Probability of heads = 1 (desired outcome) / 2 (possible outcomes) = .50 or 50% \n* Dice roll: Probability of rolling 3 = 1 (specific number) / 6 (possible numbers) = .1666 or 16.66% \n* Cards: Probability of drawing 10 ♦ = 1 (specific card) / 52 (possible cards) = .0192 or 1.92% \n\n<center>$Sum of Probabilities\\ for\\ all\\ outcomes\\ = 1.0$</center>\n\n---\n\nThe probability of an event not occurring is called the **complement** and is calculated:\n\n<center>$Probability\\ of\\ Event\\ not\\ occurring = Probability\\ of\\ all\\ outcomes\\ - Probability\\ of\\ one\\ outcome$</center>\n\nThat is:\n\n<center>$P(not\\ A) = 1 - P(A)$</center>", "_____no_output_____" ], [ "## Probability of multiple random variables", "_____no_output_____" ], [ "Each **column** in a machine learning data set represents a **variable** and each *row* represents an *observation*. Much of the behind-the-scenes math in machine learning deals the probability of one variable in the presence of the observation's other variables. \n\nLet's look again at this section's definitions, in light of what we saw above:\n\n**Joint probability:** Probability of events *A* and *B*. \n**Marginal probability:** Probability of event *A* given variable *Y*. \n**Conditional probability:** Probability of event *A* given event *B*. ", "_____no_output_____" ], [ "### Joint probability", "_____no_output_____" ], [ "Joint probability is the chance that **both** event A and event B happen. This can be written several ways:\n\n<center>\n$$P(A\\ and\\ B)$$\n$$P(A\\ \\cap\\ B)$$\n$$P(A,B)$$\n</center>\n\nJoint probability of A and B can be calculated as *the probability of event A given event B times the probability of event B*. In more mathematical terms:\n\n<center>$P(A\\ \\cap\\ B) = P(A\\ given\\ B)\\ \\times\\ P(B)$</center>\n\n\n", "_____no_output_____" ], [ "## Marginal probability", "_____no_output_____" ], [ "For given fixed event *A* and variable *Y*, marginal probability is the sum of probabilities that one of *Y*'s events will happen along with fixed event *A*. Let's look at that in table form.\n\n* Let's say we ask a group of 60 people which color they like better, **blue** or **pink**.\n\n|Gender| Blue|Pink|Total|\n|------|-----|----|-----|\n|Male|25|10|P(male) = 35 / 60 = 0.5833| \n|Female|5|20|P(female) 25 / 60 = 0.4166|\n|Total|P(blue) = 30 / 60 = .50 | P(pink) = 30 / 60 = .50| total = 60\n\n**Rows** represent the probability that a respondent was a particular gender. \n**Columns** represent the probability of the response being that color. \n\nTo express that more mathematically, \n\n<center>$P(X=A)=\\sum\\limits_{}^{y\\in Y}P(X=A,\\ Y=y)$</center>", "_____no_output_____" ], [ "## Conditional probability", "_____no_output_____" ], [ "Remember, in programming languages, we call `if->then->else` statements *conditionals*.\n\nA **conditional probability** can be thought of as **The probability that event A will happen _if_ event B has happened**.\n\nThe slightly more \"mathy\" way to say that is: **The probability of event A _given_ event B.**\n\nIn formula form, we use **\"|\"** (pipe) as the \"given.\"\n\n<center>\n $P(A\\ given\\ B)$ \n or \n $P(A|B)$ \n </center>\n \n \nThe conditional probability of event A given event B can be calculated by: \n\n\n<center>$P(A|B) = \\frac{P(A \\cap B)}{P(B)}$</center>\n", "_____no_output_____" ], [ "---\n\n**All of the probability above was included simply so we could understand Bayes Theorem (below) and its' application to machine learning.**\n", "_____no_output_____" ], [ "# Bayes Theorem", "_____no_output_____" ], [ "Bayes Theorem gives us a structured way to calculate **conditional probabilities**.\n\nRemember from above, conditional probability is the probability that *event A* will happen *given event B*. In mathematical terms, that is: \n\n<center>$P(A|B) = \\frac{P(A \\cap B)}{P(B)}$</center>\n\nNote that $P(A|B) \\neq P(B|A)$\n\n**Bayes Theorem** gives us another way to calculate conditional probability when the joint probability is not known: \n\n<center>$P(A|B) = \\frac{P(B|A)\\ \\times\\ P(A)}{P(B)}$</center>\n", "_____no_output_____" ], [ "However, we may not know $P(B)$. It can be calculated an alternatve way: \n\n<center>$P(B)=P(B|A)\\ \\times\\ P(A)\\ +\\ P(B|not\\ A)\\ \\times\\ P(not\\ A)$</center> \n\nThen, through the mathematical trickery of substitution, we get: \n\n<center>$P(A|B) = \\frac{P(B|A)\\ \\times\\ P(A)}{P(B|A)\\ \\times\\ P(A)\\ +\\ P(B|not\\ A)\\ \\times\\ P(not\\ A)}$</center>\n\nAlso, remember that \n<center>$P(not\\ A)=1 - P(A)$</center> \n\nFinally, if we have $P(not\\ B|not\\ A)$ we can calculate $P(B|not\\ A)$: \n<center>$P(B|not\\ A) = 1 - P(not\\ B|not\\ A)$</center>", "_____no_output_____" ], [ "### Terminology:", "_____no_output_____" ], [ "The probabilities are given English names to help understand what they are trying to say:\n\n* $P(A|B)$: Posterior probability \n* $P(A)$: Prior probability \n* $P(B|A)$: Likelihood \n* $P(B)$: Evidence \n\nNow, Bayes Theorem can be restated as:\n<center>$Posterior = \\frac{Likelihood\\ \\times\\ Prior}{Evidence}$</center>\n\n---\n\nJason Brownlee gives us the fantastic analogy of the probability that there is fire given that there is smoke. \n\n* $P(Fire)$ is the prior \n* $P(Smoke|Fire)$ is the likelihood \n* $P(Smoke)$ is the evidence \n<center>$P(Fire|Smoke) = \\frac{P(Smoke|Fire)\\ \\times\\ P(Fire)}{P(Smoke)}$</center>", "_____no_output_____" ], [ "# Bayes Theorem as Binary Classifier", "_____no_output_____" ], [ "Bayes Theorem is often used as a **binary classifier** -- the classic example that we will look at in a few moments is detecting spam in email. But first, more terminology.", "_____no_output_____" ], [ "## Terminology", "_____no_output_____" ], [ "* $P(not\\ B|not\\ A)$: True Negative Rate **TNR** (specificity) \n* $P(B|not\\ A)$: False Positive Rate **FPR** \n* $P(not\\ B|A)$: False Negative Rate **FNR** \n* $P(B|A)$: True Positive Rate **TPR** (sensitivity or recall) \n* $P(A|B)$: Positive Predictive Vale **PPV** (precision) ", "_____no_output_____" ], [ "Applying the above to the longer formula above: \n\n<center>$Positive\\ Predictive\\ Value = \\frac{True\\ Positive\\ Rate\\ \\times\\ P(A)}{True\\ Positive\\ Rate\\ \\times\\ P(A)\\ +\\ False\\ Positive\\ Rate\\ \\times\\ P(not\\ A) }$</center>", "_____no_output_____" ], [ "## Examples", "_____no_output_____" ], [ "Let's look at some (contrived) examples, courtesy of Jason Brownlee:", "_____no_output_____" ], [ "### Elderly Fall and Death", "_____no_output_____" ], [ "Let's define elderly as over 80 years of age. What is the probabiity that an elderly person will die from a fall? Let's use 10% as the base rate for elderly death - P(A), and the base rate for elderly falling is 5% - P(B), and 7% of elderly that die had a fall - P(B|A). \n\n<center>$P(A|B) = \\frac{P(B|A)\\ \\times\\ P(A)}{P(B)}$</center> \n\n<center>$P(Die|Fall) = \\frac{P(Fall|Die)\\ \\times\\ P(Die)}{P(Fall)}$</center> \n \n<center>$P(A|B) = \\frac{0.07\\ \\times\\ 0.10}{0.05}$</center> \n\n<center>$P(Die|Fall) = 0.14$</center> \n\nSo, using these completely fake numbers, 14% of elderly falls would end in death.", "_____no_output_____" ], [ "### Spam Detection", "_____no_output_____" ], [ "Let's say our spam filter put an email in the spam folder. What is the probability it was spam?\n\n* 2% of email is spam - P(A).\n* 99% accuracy on the spam filter - P(B|A)\n* 0.1% of email is incorrectly marked as spam - P(B|not A)\n\n<center>$P(A|B) = \\frac{P(B|A)\\ \\times\\ P(A)}{P(B)}$</center> \n\n<center>$P(Spam|Detected) = \\frac{P(Detected|Spam)\\ \\times\\ P(Spam)}{P(Detected)}$</center> \n\nUnfortunately, we don't know P(B) -- P(Detected), but we can figure it out. Recall, \n\n<center>$P(B)=P(B|A)\\ \\times\\ P(A)\\ +\\ P(B|not\\ A)\\ \\times\\ P(not\\ A)$</center> \n\n<center>$P(Detected)=P(Detected|Spam)\\ \\times\\ P(Spam)\\ +\\ P(Detected|not\\ Spam)\\ \\times\\ P(not\\ Spam)$</center> \n\nAnd, we can calculate P(not Spam):\n\n<center>$P(not\\ Spam) = 1 - P(Spam) = 1 - 0.02 = 0.98$</center> \n\n<center>$P(Detected) = 0.99\\ \\times\\ 0.02\\ +\\ 0.001\\ \\times\\\n 0.98$</center> \n \nRemember order of operations here... multiply before addition: \n \n<center>$P(Detected) = 0.0198 + 0.00098 = 0.02078$</center> \n\nWe can finally put it all together: \n\n<center>$P(Spam|Detected) = \\frac{0.99\\ \\times\\ 0.02}{0.02078}$</center> \n\n<center>$P(Spam|Detected) = \\frac{0.0198}{0.02078}$</center> \n\n<center>$P(Spam|Detected) = 0.9528392$</center> \n\nOr, about a 95% chance that the email was classified properly. ", "_____no_output_____" ], [ "# Naive Bayes Classification", "_____no_output_____" ], [ "Supervised machine learning is typically used for prediction or classification, as we will see in Week 7.\n\nBayes Theorem can be used for classification, however even with modern computing advances, figuring out all the probabilities of the dependent variables would be impractical. For this reason, the mathematics of Bayes Theorem is simplified in various ways, including by assuming all variables are independent. \n\nWe will look at Naive Bayes Classification in more depth later in this class and again in MSDS 680 Machine Learning. ", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown" ], [ "code", "code", "code" ], [ "markdown", "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown" ] ]

4ad8206f8a312895882b2f9f8c4f23f647b420bf

24,971

ipynb

Jupyter Notebook

Jupyter_HW2.ipynb

sofrodriguez/JupyterHW2

b9a2b6c8592965a187f6fef4b19fd88a72c8f24d

[ "MIT" ]

null

Jupyter_HW2.ipynb

sofrodriguez/JupyterHW2

b9a2b6c8592965a187f6fef4b19fd88a72c8f24d

[ "MIT" ]

null

Jupyter_HW2.ipynb

sofrodriguez/JupyterHW2

b9a2b6c8592965a187f6fef4b19fd88a72c8f24d

[ "MIT" ]

2

2022-02-22T03:01:14.000Z

2022-02-22T22:08:56.000Z

54.521834

1,295

0.671379

[ [ [ "# Using surface roughness to date landslides\n\n### Overview\nIn March of 2014, unusually high rainfall totals over a period of several weeks triggered a deep-seated landslide that mobilized into a rapidly moving debris flow. The debris flow inundated the town of Oso, Washington, resulting in 43 fatalities and the destruction of 49 houses. Other landslide deposits are visible in the vicinity of the 2014 Oso landslide (see figure below). The goal of this assignment is to estimate the ages of the nearby landslide deposits so that we can say something about the recurrence interval of large, deep-seated landslides in this area. Do they happen roughly every 100 years, 5000 years, or do they only happen once every 100,000 years?\n\n<img src=\"OsoOverviewMap.jpg\" alt=\"Drawing\"/>\n\nOur strategy will be to take advantage of the fact that recent landslides have “rougher” surfaces. Creep and bioturbation smooth the landslide deposits over time in a way that we can predict (using the diffusion equation!). We will use the standard linear diffusion model, shown below, to simulate how the surface of a landslide deposit will change with time:\n\n$$ \\frac{\\partial z}{\\partial t}=D\\frac{\\partial^2z}{\\partial x^2} $$\n\nHere, $z$ denotes elevation, $x$ is distance in the horizontal direction, and $D$ is the colluvial transport coefficient. Recall, that in a previous exercise we estimated the value of $D$ within the San Francisco Volcanic Field (SFVF) in northern Arizona. We found that $D\\approx5$ $\\mathrm{m^2}$ $\\mathrm{kyr}^{-1}$ in the SFVF. In this exercise, we will use a larger value of $D=10$ $\\mathrm{m^2}$ $\\mathrm{kyr}^{-1}$ since our study site near Oso, Washington, is in a wetter climate with more vegetation (and therefore greater rates of bioturbation). Once we have a model that lets us determine how the surface of a landslide deposit will change with time, we may be able to use it to describe how surface roughness varies with age.", "_____no_output_____" ], [ "### Landslide Deposit Morphology\n\nFirst, examine the map below showing the slope in the area of the Oso Landslide. Also pictured is the Rowan Landslide, which is older than the Oso Landslide.\n\n<img src=\"OsoSlopeMap.jpg\" alt=\"Drawing\"/>\n\nNotice how the Oso landslide, which is very recent, is characterized by a number of folds and a very “rough” surface. This type of hummocky topography is common in recent landslide deposits. The plot on the right shows a topographic transect that runs over the Oso Landslide deposit from north to south.", "_____no_output_____" ], [ "### Quantifying Surface Roughness\n\nIf we are ultimately going to use surface roughness to date landslide deposits (i.e. older deposits are less rough, younger deposits are more rought), we first need a way to quantify what we mean by \"roughness\". One way to quantify surface roughness is to extract a transect from the slope data and compute the standard deviation of the slope along that transect. That is what we do here; we compute the standard deviation of the slope (SDS) over each 30-meter interval along a transect and then take the mean of all of these standard deviations to arrive at an estimate of roughness for each landslide deposit that we are interested in dating. The plots below show slope (deg) along transects that run over the 2014 Oso Landslide and the nearby Rowan Landslide (unknown age). Note that the Rowan landslide looks slightly less “rough” and therefore has a lower SDS value associated with it.\n\n<img src=\"RowanOsoTransects.png\" alt=\"Drawing\"/>\n\nDon't worry about understanding exactly how SDS is computed. The most important thing to note here is that SDS gives us a way to objectively define how \"rough\" a surface is. Higher values of SDS correspond to rough surfaces whereas lower values correspond to smoother surfaces.", "_____no_output_____" ], [ "### Estimating the Age of the Rowan Landslide\n\nWe will now estimate the age of the Rowan Landslide using the diffusion model. This is the same model we used to simulate how cinder cones evolve. Since the same processes (creep, bioturbation) are driving sediment transport on landslide deposits, we can apply the same model here. However, when we modeled cinder cones we knew what the initial condition looked like. All cinder cones start with cone shape that is characterized by hillslope angles that are roughly equal to the angle of repose for granular material ($\\approx 30^{\\circ}$). We do not know what each of these landslide deposits looked like when they were first created. So, we will assume that all landslide deposits (including the Rowan Landslide) looked like the Oso Landslide immediately after they were emplaced. Of course, no two landslide deposits ever look exactly the same but it is reasonable to assume that the statistical properties of the initial landslide deposits (i.e. roughness) are similar to each other. We will make this assumption and simulate how the roughness, as quantified using the SDS, of the Oso Landslide deposit will change over time. If we know the relationship between SDS and deposit age, then we can estimate the age of any landside deposit in this region simply by computing its SDS. \n\nLet’s start by using the model to estimate how much the Oso Landslide deposit will change after it is subjected to erosion via diffusive processes (e.g. bioturbation, rain splash, freeze-thaw) for 100 years. The code below is set up to run the diffusion model using a topographic transect through the Oso Landslide as the initial topography. All you need to do is assign realistic values for the colluvial transport coefficient (landscape diffusivity) and choose an age. Use a value of $D=10$ $\\mathrm{m}^2$ $\\mathrm{kyr}^{-1}$ for the colluvial transport coefficient and an age of 0.1 kyr (since we want to know what the deposit will look like when the Oso Landslide is 100 years old). Then run the code block below. ", "_____no_output_____" ] ], [ [ "D=10; # Colluvial transport coefficient [m^2/kyr] (i.e. landscape diffusivity)\nage=0.1; # Age of the simulated landslide deposit [kyr]\n\n# !! YOU DO NOT NEED TO MODIFY THE CODE BELOW THIS LINE !! \n\nfrom diffusion1d import oso\n[distance,elevation,SDS]=oso(D,age)\n\nimport matplotlib.pyplot as plt\nplt.plot(distance,elevation,'b-')\nplt.xlabel('Distance (m)', fontsize=14)\nplt.ylabel('Elevation (m)', fontsize=14)\nplt.title('SDS = '+str(round(SDS,1)), fontsize=14)\nplt.show()", "_____no_output_____" ] ], [ [ "You should see that the SDS value for the topography (shown on the plot) after 0.1 kyr of erosion is slightly smaller than the SDS value of the initial landslide surface (e.g. the SDS value of the Oso Landslide deposit). This is a result of the fact that diffusive processes smooth the surface over time, but 0.1 kyr is not a sufficient amount of time to substantially *smooth* the surface of the landslide deposit. Although the SDS value has decreased over a time period of 0.1 kyr, it is still larger than the SDS value that we have computed for the Rowan Landslide deposit. Therefore, the Rowan Landslide deposit must be older than 0.1 kyr. Continue to run the model using the code cell above with increasing values for the age until you find an age that gives you a SDS value that is close to the one computed for the Rowan Landslide ($SDS\\approx5.2$). Based on this analysis, how old is the Rowan Landslide (you can round your answer to the nearest 0.1 kyr)?", "_____no_output_____" ], [ "INSERT YOUR ANSWER HERE", "_____no_output_____" ], [ "### How does SDS vary with age?\n\nYou have now successfully dated the Rowan Landslide! This process does not take too long, but it can be inefficient if we want to date a large number of landslide deposits. Later, I will give you SDS values for 12 different landslide deposits in this area. We want to date all of them so that we can have more data to accomplish our original goal of saying something about the recurrence interval of large, deep-seated landslides in this area. To do this, we will determine an equation that quantifies the relationship between SDS and age using our diffusion model. Then, we will use this equation to tell us the age of each landslide deposit based on its SDS. \n\nTo get started on this process, lets use the model in the code cell below to determine how the SDS value changes as we change the age of the landslide deposit. Use the model (in the code cell below) to simulate the surface of the Oso Landslide after 1 kyr, 2 kyr, 5 kyr, 10 kyr, and 20 kyr. Continue to use a value of $D=10$ $\\mathrm{m}^2$ $\\mathrm{kyr}^{-1}$. Write down each of the SDS values that you get for these 5 different ages. You will need each of them to complete the next step. Note that it may take 5-10 seconds to compute the SDS when the ages are 10 kyr or 20 kyr since more computations need to be performed to complete these longer simulations.", "_____no_output_____" ] ], [ [ "D=10; # Colluvial transport coefficient [m^2/kyr] (i.e. landscape diffusivity)\nage=0.1; # Age of the simulated landslide deposit [kyr]\n\n# !! YOU DO NOT NEED TO MODIFY THE CODE BELOW THIS LINE !! \nfrom diffusion1d import oso\n[distance,elevation,SDS]=oso(D,age)\n\nimport numpy as np\nitopo=np.loadtxt('osotransect.txt')\n\nimport matplotlib.pyplot as plt\nplt.plot(distance,elevation,'b-',label=\"Modeled Topography\")\nplt.plot(distance,itopo,'--',color='tab:gray',label=\"Modeled Topography\")\nplt.xlabel('Distance (m)', fontsize=14)\nplt.ylabel('Elevation (m)', fontsize=14)\nplt.title('SDS = '+str(round(SDS,1)), fontsize=14)\nplt.show()", "_____no_output_____" ] ], [ [ "### A general method for estimating age based on SDS\n\nIn the code below, we are going to create several variables (\"SDS_0kyr\", \"SDS_1kyr\", etc) so that we can store the information that you obtained in the previous section. Each variable will hold the SDS value of our idealized landslide deposit for different ages. Notice that the variable called *SDS_0kyr* is equal to the SDS value of the Oso transect, which is the same as the SDS value at a time of 0 kyr (since the landslide occured in 2014). The variables *SDS_1kyr*, *SDS_2kyr*,...,*SDS_20kyr* are all set equal to a value of 1. Change these values in the code block below to reflect the SDS values that you computed in the above exercise. For example, if you determined that the landslide deposit has an SDS value of $6.4$ after 5 kyr then set *SDS_5kyr* equal to $6.4$. When you are finished, run the code cell. The code should produce a plot of your data. Verify that the plot appears to be accurate.", "_____no_output_____" ] ], [ [ "SDS_0kyr=9.5 # This is the initial (i.e. t=0) SDS value of our landslide deposit.\nSDS_1kyr=1 # Change this values from \"1\" to the SDS value after 1 kyr.\nSDS_2kyr=1 # Change this values from \"1\" to the SDS value after 2 kyr.\nSDS_5kyr=1 # Change this values from \"1\" to the SDS value after 5 kyr.\nSDS_10kyr=1 # Change this values from \"1\" to the SDS value after 10 kyr.\nSDS_20kyr=1 # Change this values from \"1\" to the SDS value after 20 kyr.\n\n\n# You do not need to modify any code below this point\nimport numpy as np\nage=np.array([0,1,2,5,10,20])\nSDS=np.array([SDS_0kyr,SDS_1kyr,SDS_2kyr,SDS_5kyr,SDS_10kyr,SDS_20kyr])\n\nimport matplotlib.pyplot as plt\nplt.scatter(SDS,age,s=60, c='b', marker='o') # Create the scatter plot, set marker size, set color, set marker type\nplt.xlabel('SDS [-]', fontsize=14)\nplt.ylabel('Landslide Age [kyr]', fontsize=14)\nplt.show()", "_____no_output_____" ] ], [ [ "Now, we need to find a way to use the information above to come up with a more general relationship between SDS and age. Right now we only have 6 points on a graph. We have no way to determine the age of a landslide if its SDS value falls in between any of the points on our plot. One way to proceed is to fit a curve to our 6 data points. Python has routines that can be used to fit a function to X and Y data points. You may have experience using similar techniques in programs like Excel or MATLAB. \n\nBefore proceeding to work with our data, lets examine how this process of curve fitting works for a simple case. Suppose we are given three points, having X coordinates of 1,2, and 3 and corresponding Y coordinates of 2,4, and 6. Below is an example of how to fit a line to data using Python. **Do not worry about understanding how all of the code works. The aim of this part of the exercise is simply to introduce you to the types of tools that are available to you in programming languages like Python. That way, if you run into problems later in your professional or academic career, you will know whether or not using Python or a similar approach will be helpful.** Run the code block below. Then we will examine the output of the code.", "_____no_output_____" ] ], [ [ "# You do not need to modify any code in this cell\n\n# First, define some X data\nX=[1,2,3]\n\n# Then define the corresponding Y data\nY=[3,5,7]\n\n# Use polyfit to find the coefficients of the best fit line (i.e the slope and y-intercept of the line) \nimport numpy as np\npfit=np.polyfit(X,Y,1)\n\n# Print the values contained in the variable \"pfit\"\nprint(pfit)", "_____no_output_____" ] ], [ [ "You should see two values printed at the bottom of the code block. Python has determined the line that best fits the X and Y data that we provided. As you know, a line is described by two numbers: a slope and a y-intercept. Not surprisingly, Python has given us two numbers. The first number, which is a 2, corresponds to the slope of the best fit line. The second number, which is a 1, corresponds to the y-intercept. Thus, we now know that the best fit line for this X and Y data is given by\n\n$$ Y=2X+1$$\n", "_____no_output_____" ], [ "### Fitting a line to your data\n\nNow that we know how to interpret the output from the *polyfit* function, we can see what information it gives us about the relationship between SDS and age. Look at the plot that you created earlier that shows age as a function of SDS. Age is the Y variable (i.e. the dependent variable) and SDS is the X variable (or independent variable). This is what we want because we ultimately want to be able to estimate the age of a landslide based on its SDS value. \n\nIn the code below, we use polyfit to find the line that best describes the relationship between age and SDS. Notice that the code looks exactly like the code for the simple curve fitting example shown above except that *SDS* has been substituted for X and *age* has been substituted for Y. Run the code below (you don't need to make any changes to it) and then we will discuss the output.", "_____no_output_____" ] ], [ [ "# You do not need to modify any code in this cell\n\n# You do not need to modify any code in this cell\npfit=np.polyfit(SDS,age,1)\n\n# Print the values contained in the variable \"pfit\"\nprint(pfit)", "_____no_output_____" ] ], [ [ "Python has determined the line that best desribes the relationship between SDS and age. The first number in the output represents the slope of the line and the second number is the y intercept. The first number (the slope) should be roughly $-2.65$. The second number (the y-intercept) should be roughly $21.5$. This means that \n$$\n\\mathrm{AGE}=21.5-2.65 \\cdot{} \\mathrm{SDS}\n$$\n\nwhere AGE denotes the age of the landslide deposit in kyr. The code below will plot your best fit line on top of the actual data. Run the code below to see what your best fit line looks like.", "_____no_output_____" ] ], [ [ "# You do not need to modify any code in this cell\npfit=np.polyfit(SDS,age,1);\n\nimport matplotlib.pyplot as plt\nplt.scatter(SDS,age,s=60, c='b', marker='o',label=\"Original Data\") # Create the scatter plot, set marker size, set color, set marker type\nplt.plot(SDS,pfit[1]+pfit[0]*SDS,'k-',label=\"Best Fit Line\")\nplt.xlabel('SDS [-]', fontsize=14)\nplt.ylabel('Landslide Age [kyr]', fontsize=14)\nplt.legend()\nplt.show()", "_____no_output_____" ] ], [ [ "You should see that a line does not fit the data very well. If you have correctly completed the assignment to this point, you will notice that the data points (blue circles) in the above plot show that age decreases rapidly with SDS at first and then decreases more slowly at higher SDS values. This pattern suggests that age varies nonlinearly with SDS. Motivated by this observation, let’s see if a 2nd order polynomial (i.e. a quadratic function) will provide a better fit to our data. ", "_____no_output_____" ], [ "### Fitting a quadratic function to your SDS data\n\nWe use *polyfit* in the code below in much the same way as before. The only difference is that we want Python to find the quadratic funciton that best describes our data. We still need to provide the X data (i.e. \"SDS\") and the Y data (i.e. age). The only difference is that we change the third input for the *polyfit* function from a 1 (indicating that you want your data fit to a 1st order polynomial, which is a line) to a 2 (which indicates that you want your data fit to a 2nd order polynomial, which is a quadratic). Run the code block below and Python will determine the quadratic function that best fits your data.", "_____no_output_____" ] ], [ [ "# You do not need to modify any code in this cell\npfit=np.polyfit(SDS,age,2)\n\n# Print the values contained in the variable \"pfit\"\nprint(pfit)", "_____no_output_____" ] ], [ [ "Notice that Python returns three numbers. This is because three numbers are required to define a quadratic function, which looks like:\n\n$$ AGE=A\\cdot (SDS)^2+B\\cdot SDS+C $$\n\nThe first number above is the coefficient $A$. The second number is equal to $B$ and the third is equal to $C$. In your notes, write down the equation of the best fit quadratic function. You will need to use this equation to finish the exercise. ", "_____no_output_____" ], [ "Let's see how well this quadratic function fits our data. Run the code below to plot the best fit quadratic function on the same plot as your data. Verify that the best fit quadratic looks reasonable in comparison to the actual data points. In other words, it should look like the curve fits the data reasonably well.", "_____no_output_____" ] ], [ [ "# You do not need to modify any code in this cell\npfit=np.polyfit(SDS,age,2);\n\nimport matplotlib.pyplot as plt\nplt.scatter(SDS,age,s=60, c='b', marker='o',label=\"Original Data\") # Create the scatter plot, set marker size, set color, set marker type\nplt.plot(SDS,pfit[2]+pfit[1]*SDS+pfit[0]*SDS**2,'k-',label=\"Best Fit Quadratic\")\nplt.xlabel('SDS [-]', fontsize=14)\nplt.ylabel('Landslide Age [kyr]', fontsize=14)\nplt.legend()\nplt.show()", "_____no_output_____" ] ], [ [ "You now have a model (i.e. the quadratic function that you just found) that can be used to predict the age of a landslide based on its SDS. It is definitely not a perfect model but it will be ok for our purposes today. If we had more time, it would be beneficial to try to fit our data to a function that looks like:\n$$\n\\displaystyle{AGE=Ae^{-B*SDS}}\n$$\nDespite its limitations, the best quadratic function that we have found will be ok for our purposes. It will allow us to come up with rough estimates for the ages of other landslides in this area.", "_____no_output_____" ], [ "### Landslide recurrence", "_____no_output_____" ], [ "Below is a list of SDS values for $12$ other landslide deposits in the area of the Oso landslide. Use the best-fit quadratic function from your above analysis to compute the age of each deposit based on its SDS. You can make your computations any way that you like (excel, calculator, etc) - you do not need to write code to do this.\n\n5.72\n\n5.11\n\n4.40\n\n4.57\n\n5.53\n\n5.55\n\n4.38\n\n6.13\n\n6.57\n\n6.08\n\n6.47\n\n5.81\n", "_____no_output_____" ], [ "How many landslides have ages less than 1 kyr?", "_____no_output_____" ], [ "INSERT YOUR ANSWER HERE", "_____no_output_____" ], [ "How many landslides have ages between 1 kyr and 3 kyr?", "_____no_output_____" ], [ "INSERT YOUR ANSWER HERE", "_____no_output_____" ], [ "How many landslides have ages greater than 3 kyr?", "_____no_output_____" ], [ "INSERT YOUR ANSWER HERE", "_____no_output_____" ], [ "Based on your analysis, is it likely that this area will experience another large, deep-seated landslide within the next one thousand years? Since we are not doing a rigorous analsysis of recurrence intervals, you do not need to do any calculations (other than those needed to answer the three questions above). In the space below, include a parapgraph of text (3-5 sentences) in which you answer this question and provide some justification for your reasoning. Your justificaiton could include (but does not have to) some discussion about the limitations of the above analysis.", "_____no_output_____" ], [ "INSERT YOUR ANSWER HERE", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown", "markdown" ] ]

4ad83714c20c0d65b520bffcd701528a8ea7a3f7

312,585

ipynb

Jupyter Notebook

data and database/.ipynb_checkpoints/database class 8 June16-checkpoint.ipynb

sz2472/foundations-homework

3b33175d6b0a7d0fbdef8c5380ba87aa371b459e

[ "MIT" ]

null

data and database/.ipynb_checkpoints/database class 8 June16-checkpoint.ipynb

sz2472/foundations-homework

3b33175d6b0a7d0fbdef8c5380ba87aa371b459e

[ "MIT" ]

null

data and database/.ipynb_checkpoints/database class 8 June16-checkpoint.ipynb

sz2472/foundations-homework

3b33175d6b0a7d0fbdef8c5380ba87aa371b459e

[ "MIT" ]

null

43.547646

1,075

0.507791

[ [ [ "## regular expressions", "_____no_output_____" ] ], [ [ "input_str = \"Yes, my zip code is 12345. I heard that Gary's zip code is 23456. But 212 is not a zip code.\"", "_____no_output_____" ], [ "import re\nzips= re.findall(r\"\\d{5}\", input_str)\nzips", "_____no_output_____" ], [ "from urllib.request import urlretrieve\nurlretrieve(\"https://raw.githubusercontent.com/ledeprogram/courses/master/databases/data/enronsubjects.txt\", \"enronsubjects.txt\")", "_____no_output_____" ], [ "subjects = [x.strip() for x in open(\"enronsubjects.txt\").readlines()] #x.trip()［\\n]", "_____no_output_____" ], [ "subjects[:10]", "_____no_output_____" ], [ "subjects[-10]", "_____no_output_____" ], [ "[line for line in subjects if line.startswith(\"Hi!\")]", "_____no_output_____" ], [ "import re", "_____no_output_____" ], [ "[line for line in subjects if re.search(\"shipping\", line)] # if line string match the \"\" parameter", "_____no_output_____" ] ], [ [ "### metacharacters\nspecial characters that you can use in regular expressions that have a \nspecial meaning: they stand in for multiple different characters of the same \"class\"", "_____no_output_____" ], [ ".: any char\n\\w any alphanumeric char (a-z, A-Z, 0-9)\n\\s any whitespace char (\"_\", \\t,\\n)\n\\S any non-whitespace char\n\\d any digit(0-9)\n\\. actual period", "_____no_output_____" ] ], [ [ "[line for line in subjects if re.search(\"sh.pping\", line)] #. means any single character sh.pping is called class", "_____no_output_____" ], [ "# subjects that contain a time, e.g., 5: 52pm or 12:06am \n[line for line in subjects if re.search(\"\\d:\\d\\d\\wm\", line)] # \\d:\\d\\d\\wm a template read character by character", "_____no_output_____" ], [ "[line for line in subjects if re.search(\"\\.\\.\\.\\.\\.\", line)]", "_____no_output_____" ], [ "# subject lines that have dates, e.g. 12/01/99\n[line for line in subjects if re.search(\"\\d\\d/\\d\\d/\\d\\d\", line)]", "_____no_output_____" ], [ "[line for line in subjects if re.search(\"6/\\d\\d/\\d\\d\", line)]", "_____no_output_____" ] ], [ [ "### define your own character classes\ninside your regular expression, write [aeiou]`", "_____no_output_____" ] ], [ [ "[line for line in subjects if re.search(\"[aeiou][aeiou][aeiou][aeiou]\",line)] ", "_____no_output_____" ], [ "[line for line in subjects if re.search(\"F[wW]:\", line)] #F followed by either a lowercase w followed by a uppercase W", "_____no_output_____" ], [ "# subjects that contain a time, e.g., 5: 52pm or 12:06am \n[line for line in subjects if re.search(\"\\d:[012345]\\d[apAP][mM]\", line)]", "_____no_output_____" ] ], [ [ "### metacharacters 2: anchors\n^ beginning of a str\n$ end of str\n\\b word boundary", "_____no_output_____" ] ], [ [ "# begin with the New York character #anchor the search of the particular string\n[line for line in subjects if re.search(\"^[Nn]ew [Yy]ork\", line)]", "_____no_output_____" ], [ "[line for line in subjects if re.search(\"\\.\\.\\.$\", line)]", "_____no_output_____" ], [ "[line for line in subjects if re.search(\"!!!!!$\", line)]", "_____no_output_____" ], [ "# find sequence of characters that match \"oil\"\n[line for line in subjects if re.search(\"\\boil\\b\", line)]", "_____no_output_____" ] ], [ [ "### aside: matacharacters and escape characters", "_____no_output_____" ], [ "#### escape sequences \\n: new line; \\t: tab \\\\backslash \\b: word boundary", "_____no_output_____" ] ], [ [ "x = \"this is \\na test\"\nprint(x)", "this is \na test\n" ], [ "x= \"this is\\t\\t\\tanother test\"\nprint(x)", "this is\t\t\tanother test\n" ], [ "# ascii backspace\nprint(\"hello there\\b\\b\\b\\b\\bhi\")", "hello there\b\b\b\b\bhi\n" ], [ "print(\"hello\\nthere\")", "hello\nthere\n" ], [ "print(\"hello\\\\nthere\")", "hello\\nthere\n" ], [ "normal = \"hello\\nthere\"\nraw = r\"hello\\nthere\" #don't interpret any escape character in the raw string\nprint(\"normal:\", normal)\nprint(\"raw:\", raw)", "normal: hello\nthere\nraw: hello\\nthere\n" ], [ "[line for line in subjects if re.search(r\"\\boil\\b\", line)] #r for regular expression, include r for regular expression all the time", "_____no_output_____" ], [ "[line for line in subjects if re.search(r\"\\b\\.\\.\\.\\b\", line)]", "_____no_output_____" ], [ "[line for line in subjects if re.search(r\"\\banti\", line)] #\\b only search anti at the beginning of the word", "_____no_output_____" ] ], [ [ "### metacharacters 3: quantifiers\n{n} matches exactly n times\n{n.m} matches at least n times, but no more than m times\n{n,} matches at least n times, but maybe infinite times\n+ match at least onece ({1})\n* match zero or more times\n? match one time or zero times", "_____no_output_____" ] ], [ [ "[line for line in subjects if re.search(r\"[A-Z]{15,}\", line)]", "_____no_output_____" ], [ "[line for line in subjects if re.search(r\"[aeiou]{4}\", line)] #find words that have 4 characters from aeiou for each line", "_____no_output_____" ], [ "[line for line in subjects if re.search(r\"^F[wW]d?:\", line)] # find method that begins with F followed by either w or W and either a d or not d is not there; ? means find that character d or not", "_____no_output_____" ], [ "[line for line in subjects if re.search(r\"[nN]ews.*!$\", line)] # * means any characters between ews and !", "_____no_output_____" ], [ "[line in line in subjects if re.search(r\"^R[eE]:.*\\b[iI]nvestor\", line)]", "_____no_output_____" ], [ "### more metacharacters: alternation\n(?:x|y) match either x or y\n(?:x|y|z) match x,y, or z", "_____no_output_____" ], [ "[line for line in subjects if re.search(r\"\\b(?:[Cc]at|[kK]itty|[kK]itten)\\b\", line)]", "_____no_output_____" ], [ "[line for line in subjects if re.search(r\"(energy|oil|electricity)\\b\", line)]", "_____no_output_____" ] ], [ [ "### capturing", "_____no_output_____" ], [ "#### read the whole corpus in as one big string", "_____no_output_____" ] ], [ [ "all_subjects = open(\"enronsubjects.txt\").read()", "_____no_output_____" ], [ "all_subjects[:1000]", "_____no_output_____" ], [ "# domain names: foo.org, cheese.net, stuff.come\nre.findall(r\"\\b\\w+\\.(?:come|net|org)\\b\", all_subjects)", "_____no_output_____" ], [ "## differences between re.search() yes/no\n##re.findall []", "_____no_output_____" ], [ "input_str = \"Yes, my zip code is 12345. I heard that Gary's zip code is 23456. But 212 is not a zip code.\"", "_____no_output_____" ], [ "re.search(r\"\\b\\d{5}\\b\", input_str)", "_____no_output_____" ] ], [ [ "re.findall(r\"\\b\\d{5}\\b\", input_str)", "_____no_output_____" ] ], [ [ "re.findall(r\"New York \\b\\w+\\b\", all_subjects)", "_____no_output_____" ], [ "re.findall(r\"New York (\\b\\w+\\b)\", all_subjects) #the things in (): to group for the findall method", "_____no_output_____" ] ], [ [ "### using re.search to capture", "_____no_output_____" ] ], [ [ "src = \"this example has been used 423 times\"\nif re.search(r\"\\d\\d\\d\\d\", src):\n print(\"yup\")\nelse:\n print(\"nope\")", "nope\n" ], [ "src = \"this example has been used 423 times\"\nmatch = re.search(r\"\\d\\d\\d\", src)\ntype(match)", "_____no_output_____" ], [ "print(match.start())\nprint(match.end())", "27\n30\n" ], [ "print(match.group())", "423\n" ], [ "for line in subjects:\n match = re.search(r\"[A-Z]{15,}\", line)\n if match: #if find that match\n print(match.group())", "CONGRATULATIONS\nCONGRATULATIONS\nPLEEEEEEEEEEEEEEEASE\nACCOMPLISHMENTS\nACCOMPLISHMENTS\nCONFIDENTIALITY\nCONFIDENTIALITY\nCONGRATULATIONS\nCONGRATULATIONS\nACKNOWLEDGEMENT\nACKNOWLEDGEMENT\nCONGRATULATIONS\nCONGRATULATIONS\nCONGRATULATIONS\nCONGRATULATIONS\nCONGRATULATIONS\nCONGRATULATIONS\nCONGRATULATIONS\nCONGRATULATIONS\nCONGRATULATIONS\nCONGRATULATIONS\nINTERCONNECTION\nINTERCONNECTION\nINTERCONNECTION\nINTERCONNECTION\nINTERCONNECTION\nCONGRATULATIONS\nWASSSAAAAAAAAAAAAAABI\nWASSSAAAAAAAAAAAAAABI\nWASSSAAAAAAAAAAAAAABI\nWASSSAAAAAAAAAAAAAABI\nWASSSAAAAAAAAAAAAAABI\nWASSSAAAAAAAAAAAAAABI\nWASSSAAAAAAAAAAAAAABI\nNOOOOOOOOOOOOOOOO\nNOOOOOOOOOOOOOOOO\nNOOOOOOOOOOOOOOOO\nCONGRATULATIONS\nCONGRATULATIONS\nCONGRATULATIONS\nCONGRATULATIONS\nCONFIDENTIALITY\nCONFIDENTIALITY\nACCOMPLISHMENTS\nACCOMPLISHMENTS\nCONGRATULATIONS\nSTANDARDIZATION\nSTANDARDIZATION\nSTANDARDIZATION\nSTANDARDIZATION\nBRRRRRRRRRRRRRRRRRRRRR\nCONGRATULATIONS\nCONGRATULATIONS\nNETCOTRANSMISSION\nNETCOTRANSMISSION\nNETCOTRANSMISSION\nINTERCONTINENTAL\nINTERCONTINENTAL\n" ], [ "courses=[\n \n]\nprint \"Course catalog report:\"\nfor item in courses:\n match = re.search(r\"^(\\w+) (\\d+): (.*)$\", item)\n print(match.group(1)) #group 1: find the item in first group\n print(\"Course dept\", match.group(1))\n print(\"Course #\", match.group(2))\n print(\"Course title\", match.group(3))", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "raw", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code", "code" ], [ "raw" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code" ] ]

4ad8474a122eb47b27839e07c7f44a0638049039

105,318

ipynb

Jupyter Notebook

_build/jupyter_execute/12_binomial_soln.ipynb

myamullaciencia/Bayesian-statistics

1439716e377884465316cf9512f8e3a9199e79d6

[ "Apache-2.0" ]

null

_build/jupyter_execute/12_binomial_soln.ipynb

myamullaciencia/Bayesian-statistics

1439716e377884465316cf9512f8e3a9199e79d6

[ "Apache-2.0" ]

null

_build/jupyter_execute/12_binomial_soln.ipynb

myamullaciencia/Bayesian-statistics

1439716e377884465316cf9512f8e3a9199e79d6

[ "Apache-2.0" ]

null

74.272214

19,264

0.817629

[ [ [ "# Bite Size Bayes\n\nCopyright 2020 Allen B. Downey\n\nLicense: [Attribution-NonCommercial-ShareAlike 4.0 International (CC BY-NC-SA 4.0)](https://creativecommons.org/licenses/by-nc-sa/4.0/)", "_____no_output_____" ] ], [ [ "import numpy as np\nimport pandas as pd\nimport matplotlib.pyplot as plt", "_____no_output_____" ] ], [ [ "## The Euro problem\n\nIn [a previous notebook](https://colab.research.google.com/github/AllenDowney/BiteSizeBayes/blob/master/07_euro.ipynb) I presented a problem from David MacKay's book, [*Information Theory, Inference, and Learning Algorithms*](http://www.inference.org.uk/mackay/itila/p0.html):\n\n> A statistical statement appeared in The Guardian on\nFriday January 4, 2002:\n>\n> >\"When spun on edge 250 times, a Belgian one-euro coin came\nup heads 140 times and tails 110. ‘It looks very suspicious\nto me’, said Barry Blight, a statistics lecturer at the London\nSchool of Economics. ‘If the coin were unbiased the chance of\ngetting a result as extreme as that would be less than 7%’.\"\n>\n> But [asks MacKay] do these data give evidence that the coin is biased rather than fair?", "_____no_output_____" ], [ "To answer this question, we made these modeling decisions:\n\n* If you spin a coin on edge, there is some probability, $x$, that it will land heads up.\n\n* The value of $x$ varies from one coin to the next, depending on how the coin is balanced and other factors.\n\nWe started with a uniform prior distribution for $x$, then updated it 250 times, once for each spin of the coin. Then we used the posterior distribution to compute the MAP, posterior mean, and a credible interval.\n\nBut we never really answered MacKay's question.\n\nIn this notebook, I introduce the binomial distribution and we will use it to solve the Euro problem more efficiently. Then we'll get back to MacKay's question and see if we can find a more satisfying answer.", "_____no_output_____" ], [ "## Binomial distribution\n\nSuppose I tell you that a coin is \"fair\", that is, the probability of heads is 50%. If you spin it twice, there are four outcomes: `HH`, `HT`, `TH`, and `TT`.\n\nAll four outcomes have the same probability, 25%. If we add up the total number of heads, it is either 0, 1, or 2. The probability of 0 and 2 is 25%, and the probability of 1 is 50%.\n\nMore generally, suppose the probability of heads is `p` and we spin the coin `n` times. What is the probability that we get a total of `k` heads?\n\nThe answer is given by the binomial distribution:\n\n$P(k; n, p) = \\binom{n}{k} p^k (1-p)^{n-k}$\n\nwhere $\\binom{n}{k}$ is the [binomial coefficient](https://en.wikipedia.org/wiki/Binomial_coefficient), usually pronounced \"n choose k\".\n\nWe can compute this expression ourselves, but we can also use the SciPy function `binom.pmf`:", "_____no_output_____" ] ], [ [ "from scipy.stats import binom\n\nn = 2\np = 0.5\nks = np.arange(n+1)\n\na = binom.pmf(ks, n, p)\na", "_____no_output_____" ] ], [ [ "If we put this result in a Series, the result is the distribution of `k` for the given values of `n` and `p`.", "_____no_output_____" ] ], [ [ "pmf_k = pd.Series(a, index=ks)\npmf_k", "_____no_output_____" ] ], [ [ "The following function computes the binomial distribution for given values of `n` and `p`:", "_____no_output_____" ] ], [ [ "def make_binomial(n, p):\n \"\"\"Make a binomial PMF.\n \n n: number of spins\n p: probability of heads\n \n returns: Series representing a PMF\n \"\"\"\n ks = np.arange(n+1)\n\n a = binom.pmf(ks, n, p)\n pmf_k = pd.Series(a, index=ks)\n return pmf_k", "_____no_output_____" ] ], [ [ "And here's what it looks like with `n=250` and `p=0.5`:", "_____no_output_____" ] ], [ [ "pmf_k = make_binomial(n=250, p=0.5)\npmf_k.plot()\n\nplt.xlabel('Number of heads (k)')\nplt.ylabel('Probability')\nplt.title('Binomial distribution');", "_____no_output_____" ] ], [ [ "The most likely value in this distribution is 125:", "_____no_output_____" ] ], [ [ "pmf_k.idxmax()", "_____no_output_____" ] ], [ [ "But even though it is the most likely value, the probability that we get exactly 125 heads is only about 5%.", "_____no_output_____" ] ], [ [ "pmf_k[125]", "_____no_output_____" ] ], [ [ "In MacKay's example, we got 140 heads, which is less likely than 125:", "_____no_output_____" ] ], [ [ "pmf_k[140]", "_____no_output_____" ] ], [ [ "In the article MacKay quotes, the statistician says, ‘If the coin were unbiased the chance of getting a result as extreme as that would be less than 7%’.\n\nWe can use the binomial distribution to check his math. The following function takes a PMF and computes the total probability of values greater than or equal to `threshold`. ", "_____no_output_____" ] ], [ [ "def prob_ge(pmf, threshold):\n \"\"\"Probability of values greater than a threshold.\n \n pmf: Series representing a PMF\n threshold: value to compare to\n \n returns: probability\n \"\"\"\n ge = (pmf.index >= threshold)\n total = pmf[ge].sum()\n return total", "_____no_output_____" ] ], [ [ "Here's the probability of getting 140 heads or more:", "_____no_output_____" ] ], [ [ "prob_ge(pmf_k, 140)", "_____no_output_____" ] ], [ [ "It's about 3.3%, which is less than 7%. The reason is that the statistician includes all values \"as extreme as\" 140, which includes values less than or equal to 110, because 140 exceeds the expected value by 15 and 110 falls short by 15.", "_____no_output_____" ], [ "The probability of values less than or equal to 110 is also 3.3%,\nso the total probability of values \"as extreme\" as 140 is about 7%.\n\nThe point of this calculation is that these extreme values are unlikely if the coin is fair.\n\nThat's interesting, but it doesn't answer MacKay's question. Let's see if we can.", "_____no_output_____" ], [ "## Estimating x\n\nAs promised, we can use the binomial distribution to solve the Euro problem more efficiently. Let's start again with a uniform prior:", "_____no_output_____" ] ], [ [ "xs = np.arange(101) / 100\nuniform = pd.Series(1, index=xs)\nuniform /= uniform.sum()", "_____no_output_____" ] ], [ [ "We can use `binom.pmf` to compute the likelihood of the data for each possible value of $x$.", "_____no_output_____" ] ], [ [ "k = 140\nn = 250\nxs = uniform.index\n\nlikelihood = binom.pmf(k, n, p=xs)", "_____no_output_____" ] ], [ [ "Now we can do the Bayesian update in the usual way, multiplying the priors and likelihoods,", "_____no_output_____" ] ], [ [ "posterior = uniform * likelihood", "_____no_output_____" ] ], [ [ "Computing the total probability of the data,", "_____no_output_____" ] ], [ [ "total = posterior.sum()\ntotal", "_____no_output_____" ] ], [ [ "And normalizing the posterior,", "_____no_output_____" ] ], [ [ "posterior /= total", "_____no_output_____" ] ], [ [ "Here's what it looks like.", "_____no_output_____" ] ], [ [ "posterior.plot(label='Uniform')\n\nplt.xlabel('Probability of heads (x)')\nplt.ylabel('Probability')\nplt.title('Posterior distribution, uniform prior')\nplt.legend()", "_____no_output_____" ] ], [ [ "**Exercise:** Based on what we know about coins in the real world, it doesn't seem like every value of $x$ is equally likely. I would expect values near 50% to be more likely and values near the extremes to be less likely. \n\nIn Notebook 7, we used a triangle prior to represent this belief about the distribution of $x$. The following code makes a PMF that represents a triangle prior.", "_____no_output_____" ] ], [ [ "ramp_up = np.arange(50)\nramp_down = np.arange(50, -1, -1)\n\na = np.append(ramp_up, ramp_down)\n\ntriangle = pd.Series(a, index=xs)\ntriangle /= triangle.sum()", "_____no_output_____" ] ], [ [ "Update this prior with the likelihoods we just computed and plot the results. ", "_____no_output_____" ] ], [ [ "# Solution\n\nposterior2 = triangle * likelihood\ntotal2 = posterior2.sum()\ntotal2", "_____no_output_____" ], [ "# Solution\n\nposterior2 /= total2", "_____no_output_____" ], [ "# Solution\n\nposterior.plot(label='Uniform')\nposterior2.plot(label='Triangle')\n\nplt.xlabel('Probability of heads (x)')\nplt.ylabel('Probability')\nplt.title('Posterior distribution, uniform prior')\nplt.legend();", "_____no_output_____" ] ], [ [ "## Evidence\n\nFinally, let's get back to MacKay's question: do these data give evidence that the coin is biased rather than fair?\n\nI'll use a Bayes table to answer this question, so here's the function that makes one:", "_____no_output_____" ] ], [ [ "def make_bayes_table(hypos, prior, likelihood):\n \"\"\"Make a Bayes table.\n \n hypos: sequence of hypotheses\n prior: prior probabilities\n likelihood: sequence of likelihoods\n \n returns: DataFrame\n \"\"\"\n table = pd.DataFrame(index=hypos)\n table['prior'] = prior\n table['likelihood'] = likelihood\n table['unnorm'] = table['prior'] * table['likelihood']\n prob_data = table['unnorm'].sum()\n table['posterior'] = table['unnorm'] / prob_data\n return table", "_____no_output_____" ] ], [ [ "Recall that data, $D$, is considered evidence in favor of a hypothesis, `H`, if the posterior probability is greater than the prior, that is, if\n\n$P(H|D) > P(H)$\n\nFor this example, I'll call the hypotheses `fair` and `biased`:", "_____no_output_____" ] ], [ [ "hypos = ['fair', 'biased']", "_____no_output_____" ] ], [ [ "And just to get started, I'll assume that the prior probabilities are 50/50.", "_____no_output_____" ] ], [ [ "prior = [0.5, 0.5]", "_____no_output_____" ] ], [ [ "Now we have to compute the probability of the data under each hypothesis.\n\nIf the coin is fair, the probability of heads is 50%, and we can compute the probability of the data (140 heads out of 250 spins) using the binomial distribution:", "_____no_output_____" ] ], [ [ "k = 140\nn = 250\n\nlike_fair = binom.pmf(k, n, p=0.5)\nlike_fair", "_____no_output_____" ] ], [ [ "So that's the probability of the data, given that the coin is fair.\n\nBut if the coin is biased, what's the probability of the data? Well, that depends on what \"biased\" means.\n\nIf we know ahead of time that \"biased\" means the probability of heads is 56%, we can use the binomial distribution again:", "_____no_output_____" ] ], [ [ "like_biased = binom.pmf(k, n, p=0.56)\nlike_biased", "_____no_output_____" ] ], [ [ "Now we can put the likelihoods in the Bayes table:", "_____no_output_____" ] ], [ [ "likes = [like_fair, like_biased]\n\nmake_bayes_table(hypos, prior, likes)", "_____no_output_____" ] ], [ [ "The posterior probability of `biased` is about 86%, so the data is evidence that the coin is biased, at least for this definition of \"biased\".\n\nBut we used the data to define the hypothesis, which seems like cheating. To be fair, we should define \"biased\" before we see the data.", "_____no_output_____" ], [ "## Uniformly distributed bias\n\nSuppose \"biased\" means that the probability of heads is anything except 50%, and all other values are equally likely.\n\nWe can represent that definition by making a uniform distribution and removing 50%.", "_____no_output_____" ] ], [ [ "biased_uniform = uniform.copy()\nbiased_uniform[50] = 0\nbiased_uniform /= biased_uniform.sum()", "_____no_output_____" ] ], [ [ "Now, to compute the probability of the data under this hypothesis, we compute the probability of the data for each value of $x$.", "_____no_output_____" ] ], [ [ "xs = biased_uniform.index\nlikelihood = binom.pmf(k, n, xs)", "_____no_output_____" ] ], [ [ "And then compute the total probability in the usual way:", "_____no_output_____" ] ], [ [ "like_uniform = np.sum(biased_uniform * likelihood)\nlike_uniform", "_____no_output_____" ] ], [ [ "So that's the probability of the data under the \"biased uniform\" hypothesis.\n\nNow we make a Bayes table that compares the hypotheses `fair` and `biased uniform`:", "_____no_output_____" ] ], [ [ "hypos = ['fair', 'biased uniform']\nlikes = [like_fair, like_uniform]\n\nmake_bayes_table(hypos, prior, likes)", "_____no_output_____" ] ], [ [ "Using this definition of `biased`, the posterior is less than the prior, so the data are evidence that the coin is *fair*.\n\nIn this example, the data might support the fair hypothesis or the biased hypothesis, depending on the definition of \"biased\".", "_____no_output_____" ], [ "**Exercise:** Suppose \"biased\" doesn't mean every value of $x$ is equally likely. Maybe values near 50% are more likely and values near the extremes are less likely. In the previous exercise we created a PMF that represents a triangle-shaped distribution.\n\nWe can use it to represent an alternative definition of \"biased\":", "_____no_output_____" ] ], [ [ "biased_triangle = triangle.copy()\nbiased_triangle[50] = 0\nbiased_triangle /= biased_triangle.sum()", "_____no_output_____" ] ], [ [ "Compute the total probability of the data under this definition of \"biased\" and use a Bayes table to compare it with the fair hypothesis.\n\nIs the data evidence that the coin is biased?", "_____no_output_____" ] ], [ [ "# Solution\n\nlike_triangle = np.sum(biased_triangle * likelihood)\nlike_triangle", "_____no_output_____" ], [ "# Solution\n\nhypos = ['fair', 'biased triangle']\nlikes = [like_fair, like_triangle]\n\nmake_bayes_table(hypos, prior, likes)", "_____no_output_____" ], [ "# Solution\n\n# For this definition of \"biased\", \n# the data are slightly in favor of the fair hypothesis.", "_____no_output_____" ] ], [ [ "## Bayes factor\n\nIn the previous section, we used a Bayes table to see whether the data are in favor of the fair or biased hypothesis.\n\nI assumed that the prior probabilities were 50/50, but that was an arbitrary choice. \n\nAnd it was unnecessary, because we don't really need a Bayes table to say whether the data favor one hypothesis or another: we can just look at the likelihoods.\n\nUnder the first definition of biased, `x=0.56`, the likelihood of the biased hypothesis is higher: ", "_____no_output_____" ] ], [ [ "like_fair, like_biased", "_____no_output_____" ] ], [ [ "Under the biased uniform definition, the likelihood of the fair hypothesis is higher.", "_____no_output_____" ] ], [ [ "like_fair, like_uniform", "_____no_output_____" ] ], [ [ "The ratio of these likelihoods tells us which hypothesis the data support.\n\nIf the ratio is less than 1, the data support the second hypothesis:", "_____no_output_____" ] ], [ [ "like_fair / like_biased", "_____no_output_____" ] ], [ [ "If the ratio is greater than 1, the data support the first hypothesis:", "_____no_output_____" ] ], [ [ "like_fair / like_uniform", "_____no_output_____" ] ], [ [ "This likelihood ratio is called a [Bayes factor](https://en.wikipedia.org/wiki/Bayes_factor); it provides a concise way to present the strength of a dataset as evidence for or against a hypothesis.", "_____no_output_____" ], [ "## Summary\n\nIn this notebook I introduced the binomial disrtribution and used it to solve the Euro problem more efficiently.\n\nThen we used the results to (finally) answer the original version of the Euro problem, considering whether the data support the hypothesis that the coin is fair or biased. We found that the answer depends on how we define \"biased\". And we summarized the results using a Bayes factor, which quantifies the strength of the evidence.\n\n[In the next notebook](https://colab.research.google.com/github/AllenDowney/BiteSizeBayes/blob/master/13_price.ipynb) we'll start on a new problem based on the television game show *The Price Is Right*.", "_____no_output_____" ], [ "## Exercises\n\n**Exercise:** In preparation for an alien invasion, the Earth Defense League has been working on new missiles to shoot down space invaders. Of course, some missile designs are better than others; let's assume that each design has some probability of hitting an alien ship, `x`.\n\nBased on previous tests, the distribution of `x` in the population of designs is roughly uniform between 10% and 40%.\n\nNow suppose the new ultra-secret Alien Blaster 9000 is being tested. In a press conference, a Defense League general reports that the new design has been tested twice, taking two shots during each test. The results of the test are confidential, so the general won't say how many targets were hit, but they report: \"The same number of targets were hit in the two tests, so we have reason to think this new design is consistent.\"\n\nIs this data good or bad; that is, does it increase or decrease your estimate of `x` for the Alien Blaster 9000?\n\nPlot the prior and posterior distributions, and use the following function to compute the prior and posterior means.", "_____no_output_____" ] ], [ [ "def pmf_mean(pmf):\n \"\"\"Compute the mean of a PMF.\n \n pmf: Series representing a PMF\n \n return: float\n \"\"\"\n return np.sum(pmf.index * pmf)", "_____no_output_____" ], [ "# Solution\n\nxs = np.linspace(0.1, 0.4)\nprior = pd.Series(1, index=xs)\nprior /= prior.sum()", "_____no_output_____" ], [ "# Solution\n\nlikelihood = xs**2 + (1-xs)**2", "_____no_output_____" ], [ "# Solution\n\nposterior = prior * likelihood\nposterior /= posterior.sum()", "_____no_output_____" ], [ "# Solution\n\nprior.plot(color='gray', label='prior')\nposterior.plot(label='posterior')\n\nplt.xlabel('Probability of success (x)')\nplt.ylabel('Probability')\nplt.ylim(0, 0.027)\nplt.title('Distribution of before and after testing')\nplt.legend();", "_____no_output_____" ], [ "# Solution\n\npmf_mean(prior), pmf_mean(posterior)", "_____no_output_____" ], [ "# With this prior, being \"consistent\" is more likely\n# to mean \"consistently bad\".", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown" ], [ "code", "code", "code", "code", "code", "code", "code" ] ]

4ad856d91e881057cfec861256dc1c7e236400f0

106,684

ipynb

Jupyter Notebook

Small Projects/Cost of Capital/Monte Carlo Simulation of DDM.ipynb

daniel-rovilla/Financial_Modeling

55272c83ccca10d4b37d356e95c27512fb765405

[ "MIT" ]

null

Small Projects/Cost of Capital/Monte Carlo Simulation of DDM.ipynb

daniel-rovilla/Financial_Modeling

55272c83ccca10d4b37d356e95c27512fb765405

[ "MIT" ]

null

Small Projects/Cost of Capital/Monte Carlo Simulation of DDM.ipynb

daniel-rovilla/Financial_Modeling

55272c83ccca10d4b37d356e95c27512fb765405

[ "MIT" ]

null

88.755408

9,688

0.694322

[ [ [ "# Monte Carlo Simulation of Dividend Discount Model\n\n## Description\nYou are trying to determine the value of a mature company. The company has had stable dividend growth for a long time so you select the dividend discount model (DDM).\n$$P = \\frac{d_1}{r_s - g}$$\n\n### Level 1\n- The next dividend will be \\\\$1 and your baseline estimates of the cost of capital and growth are 9% and 4%, respectively\n- Write a function which is able to get the price based on values of the inputs\n- Then you are concerned about mis-estimation of the inputs and how it could affect the price. So then assume that the growth rate has a mean of 4% but a standard deviation of 1%\n- Visualize and summarize the resulting probability distribution of the price\n\n### Level 2\nContinue from the first exercise:\n- Now you are also concerned you have mis-estimated the cost of capital. So now use a mean of 9% and standard deviation of 2%, in addition to varying the growth\n- Visualize and summarize the resulting probability distribution of the price\n- Be careful as in some cases, the drawn cost of capital will be lower than the drawn growth rate, which breaks the DDM.\n - You will need to modify your logic to throw out these cases.", "_____no_output_____" ], [ "## Setup", "_____no_output_____" ] ], [ [ "from IPython.display import HTML, display\nimport matplotlib.pyplot as plt\n%matplotlib inline\nimport pandas as pd\nimport random", "_____no_output_____" ] ], [ [ "### Level 1", "_____no_output_____" ] ], [ [ "def price_ddm(dividend, cost_of_capital, growth):\n '''\n Function to determine the price of a company based on the dividend\n discount model, according to the example, inputs are placed by default\n '''\n return dividend / (cost_of_capital - growth)\nprice_ddm(1, 0.09, 0.04)", "_____no_output_____" ], [ "dividend = 1\ncost_of_capital = 0.09\ngrowth_mean = 0.04\ngrowth_std = 0.01\nn_iter = 1000000\ndef price_ddm_simulations(dividend, cost_of_capital, growth_mean, growth_std, n_iter):\n outputs = []\n for i in range(n_iter):\n growth_prob = random.normalvariate(growth_mean, growth_std)\n result = price_ddm(dividend, cost_of_capital, growth = growth_prob)\n outputs.append((growth_prob,result))\n return outputs", "_____no_output_____" ], [ "l1_results = price_ddm_simulations(dividend, cost_of_capital, growth_mean, growth_std, n_iter)\nprint(f'There are {len(l1_results)} results. First five:')\nl1_results[:5]", "There are 1000000 results. First five:\n" ] ], [ [ "### Visualize the Outputs\n\nUsually a first good way to analyze the outputs is to visualize them. Let's plot the results. A histogram or KDE plot is usually most appropriate for visualizing a single output. The KDE is basically a smoothed out histogram.", "_____no_output_____" ] ], [ [ "df = pd.DataFrame(l1_results, columns = ['Growth', 'Price'])\ndf['Price'].plot.hist(bins=100)", "_____no_output_____" ], [ "df['Price'].plot.kde()", "_____no_output_____" ] ], [ [ "As we can see, the results look basically like a normal distribution. Which makes sense because we only have one input changing and it is following a normal distribution.", "_____no_output_____" ], [ "### Probability Table\nWe would like to see two other kinds of outputs. One is a table of probabilities, along with the result which is achived at that probability in the distribution. E.g. at 5%, only 5% of cases are lower than the given value. At 75%, 75% of cases are lower than the given value. \nFirst we'll get the percentiles we want to explore in the table.", "_____no_output_____" ] ], [ [ "percentiles = [i/20 for i in range(1, 20)]\npercentiles", "_____no_output_____" ] ], [ [ "Now we can use the `.quantile` method from `pandas` to get this table easily. \nFor a better look of the dataframe we can create a function which will style the DF", "_____no_output_____" ] ], [ [ "def styled_df(df):\n return df.style.format({\n 'Growth' : '{:.2%}',\n 'Price' : '${:,.2f}'\n })\nstyled_df(df.quantile(percentiles))", "_____no_output_____" ] ], [ [ "### Level 2", "_____no_output_____" ] ], [ [ "# We will now modify the previous function to include variation in the cost of capital\ndividend = 1\ncost_capital_mean = 0.09\ncost_capital_std = 0.02\ngrowth_mean = 0.04\ngrowth_std = 0.01\nn_iter = 1000000\ndef price_ddm_simulations(dividend, cost_capital_mean, cost_capital_std, growth_mean, growth_std, n_iter):\n outputs = []\n for i in range(n_iter):\n growth_prob = random.normalvariate(growth_mean, growth_std)\n rate_prob = random.normalvariate(cost_capital_mean, cost_capital_std)\n result = price_ddm(dividend, cost_of_capital=rate_prob, growth = growth_prob)\n outputs.append((growth_prob, rate_prob, result))\n return outputs", "_____no_output_____" ], [ "l2_results = price_ddm_simulations(dividend, cost_capital_mean, cost_capital_std, growth_mean, growth_std, n_iter)\ndf_2 = pd.DataFrame(l2_results, columns = ['Growth', 'Cost of Capital', 'Price'])\ndef styled_df(df):\n return df.style.format({\n 'Growth' : '{:.2%}',\n 'Price' : '${:,.2f}',\n 'Cost of Capital' : '{:.2%}'\n })\nstyled_df(df_2.quantile(percentiles))", "_____no_output_____" ], [ "df_2['Price'].plot.hist(bins=100)", "_____no_output_____" ] ], [ [ "We notice that some price values are negative when adding variation to the cost of capital, this is due to growth rate being greater than the cost of capital which will \"break\" the DDM model and show negative prices. \nWe can look where did that happen in the next line of code:", "_____no_output_____" ] ], [ [ "df_2[df_2['Growth'] > df_2['Cost of Capital']]", "_____no_output_____" ], [ "# We will now modify the previous function to get positive values\ndef price_ddm_simulations(dividend, cost_capital_mean, cost_capital_std, growth_mean, growth_std, n_iter):\n outputs = []\n for i in range(n_iter):\n growth_prob = random.normalvariate(growth_mean, growth_std)\n rate_prob = random.normalvariate(cost_capital_mean, cost_capital_std)\n if growth_prob > rate_prob:\n continue\n result = price_ddm(dividend, cost_of_capital=rate_prob, growth = growth_prob)\n outputs.append((growth_prob, rate_prob, result))\n return outputs", "_____no_output_____" ], [ "new_l2_results = price_ddm_simulations(dividend, cost_capital_mean, cost_capital_std, growth_mean, growth_std, n_iter)\ndf_3 = pd.DataFrame(new_l2_results, columns = ['Growth', 'Cost of Capital', 'Price'])\nstyled_df(df_3.quantile(percentiles))", "_____no_output_____" ] ], [ [ "We notice now there is no instance where growth is greater than the cost of capital.", "_____no_output_____" ] ], [ [ "df_3[df_3['Growth'] > df_3['Cost of Capital']]", "_____no_output_____" ], [ "df_3['Price'].plot.hist(bins=100)", "_____no_output_____" ] ], [ [ "There are no longer negative values, however the price range is still looking odd. While growth rate is no longer greater than cost of capital, we can get really close values, which translates to the previous histogram. \nTo make the model better, we will ignore if the growth rate and the cost of capital are quite similar (0.5%)", "_____no_output_____" ] ], [ [ "# We will now modify the previous function to get positive values\ndef price_ddm_simulations(dividend, cost_capital_mean, cost_capital_std, growth_mean, growth_std, n_iter):\n outputs = []\n for i in range(n_iter):\n growth_prob = random.normalvariate(growth_mean, growth_std)\n rate_prob = random.normalvariate(cost_capital_mean, cost_capital_std)\n if growth_prob > (rate_prob - 0.005):\n continue\n result = price_ddm(dividend, cost_of_capital=rate_prob, growth = growth_prob)\n outputs.append((growth_prob, rate_prob, result))\n return outputs", "_____no_output_____" ], [ "new_l2_results = price_ddm_simulations(dividend, cost_capital_mean, cost_capital_std, growth_mean, growth_std, n_iter)\ndf_3 = pd.DataFrame(new_l2_results, columns = ['Growth', 'Cost of Capital', 'Price'])\ndf_3['Price'].plot.hist(bins=100)\nstyled_df(df_3.quantile(percentiles))", "_____no_output_____" ] ], [ [ "This range of prices make much more sense.", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ] ]

4ad85826320ccb39e1d788d269e1bf4d53797200

494,772

ipynb

Jupyter Notebook

graficoDinamicoTalita.ipynb

brunabellini/Projeto_FaleSobre

333409608c2a6fa63e0ca944e751ec5486870bf5

[ "MIT" ]

null

graficoDinamicoTalita.ipynb

brunabellini/Projeto_FaleSobre

333409608c2a6fa63e0ca944e751ec5486870bf5

[ "MIT" ]

null

graficoDinamicoTalita.ipynb

brunabellini/Projeto_FaleSobre

333409608c2a6fa63e0ca944e751ec5486870bf5

[ "MIT" ]

1

2021-05-18T23:59:04.000Z

84.302607

586

0.822019

[ [ [ "import pandas as pd\r\nimport bar_chart_race as bcr\r\nimport ffmpeg", "_____no_output_____" ], [ "#eua e japão não tem dados de suicidio de 2018 , então considerei os dados no ano de 2017\nnome_dataset = 'Média de suícidios oficial(Br.Jap.Eua.Bel).txt'\ndf = pd.read_csv(nome_dataset,sep=\";\")\ndf=df.set_index('Ano')\ndf", "_____no_output_____" ], [ "bcr.bar_chart_race(df=df,title=\"Evolução dos suicídios por número de habitantes\")\n", "C:\\Users\\thiag\\anaconda3\\lib\\site-packages\\bar_chart_race\\_make_chart.py:286: UserWarning: FixedFormatter should only be used together with FixedLocator\n ax.set_yticklabels(self.df_values.columns)\nC:\\Users\\thiag\\anaconda3\\lib\\site-packages\\bar_chart_race\\_make_chart.py:287: UserWarning: FixedFormatter should only be used together with FixedLocator\n ax.set_xticklabels([max_val] * len(ax.get_xticks()))\n" ] ] ]

[ "code" ]

[ [ "code", "code", "code" ] ]

4ad85dd64fd290c88842493687e81c4294260b2c

15,491

ipynb

Jupyter Notebook

Deep_Learning/TensorFlow-aymericdamien/notebooks/3_NeuralNetworks/dynamic_rnn.ipynb

Chau-Xochitl/INFO_7390

3ab4a3bf7af5c0b57d9604def26d64dc31405333

[ "MIT" ]

null

Deep_Learning/TensorFlow-aymericdamien/notebooks/3_NeuralNetworks/dynamic_rnn.ipynb

Chau-Xochitl/INFO_7390

3ab4a3bf7af5c0b57d9604def26d64dc31405333

[ "MIT" ]

null

Deep_Learning/TensorFlow-aymericdamien/notebooks/3_NeuralNetworks/dynamic_rnn.ipynb

Chau-Xochitl/INFO_7390

3ab4a3bf7af5c0b57d9604def26d64dc31405333

[ "MIT" ]

1

2019-12-02T06:48:30.000Z

43.759887

254

0.576335

[ [ [ "# Dynamic Recurrent Neural Network.\n\nTensorFlow implementation of a Recurrent Neural Network (LSTM) that performs dynamic computation over sequences with variable length. This example is using a toy dataset to classify linear sequences. The generated sequences have variable length.\n\n- Author: Aymeric Damien\n- Project: https://github.com/aymericdamien/TensorFlow-Examples/\n\nThese lessons are adapted from [aymericdamien TensorFlow tutorials](https://github.com/aymericdamien/TensorFlow-Examples) \n / [GitHub](https://github.com/aymericdamien/TensorFlow-Examples) \nwhich are published under the [MIT License](https://github.com/Hvass-Labs/TensorFlow-Tutorials/blob/master/LICENSE) which allows very broad use for both academic and commercial purposes.\n", "_____no_output_____" ], [ "## RNN Overview\n\n<img src=\"http://colah.github.io/posts/2015-08-Understanding-LSTMs/img/RNN-unrolled.png\" alt=\"nn\" style=\"width: 600px;\"/>\n\nReferences:\n- [Long Short Term Memory](http://deeplearning.cs.cmu.edu/pdfs/Hochreiter97_lstm.pdf), Sepp Hochreiter & Jurgen Schmidhuber, Neural Computation 9(8): 1735-1780, 1997.", "_____no_output_____" ] ], [ [ "from __future__ import print_function\n\nimport tensorflow as tf\nimport random", "_____no_output_____" ], [ "# ====================\n# TOY DATA GENERATOR\n# ====================\n\nclass ToySequenceData(object):\n \"\"\" Generate sequence of data with dynamic length.\n This class generate samples for training:\n - Class 0: linear sequences (i.e. [0, 1, 2, 3,...])\n - Class 1: random sequences (i.e. [1, 3, 10, 7,...])\n\n NOTICE:\n We have to pad each sequence to reach 'max_seq_len' for TensorFlow\n consistency (we cannot feed a numpy array with inconsistent\n dimensions). The dynamic calculation will then be perform thanks to\n 'seqlen' attribute that records every actual sequence length.\n \"\"\"\n def __init__(self, n_samples=1000, max_seq_len=20, min_seq_len=3,\n max_value=1000):\n self.data = []\n self.labels = []\n self.seqlen = []\n for i in range(n_samples):\n # Random sequence length\n len = random.randint(min_seq_len, max_seq_len)\n # Monitor sequence length for TensorFlow dynamic calculation\n self.seqlen.append(len)\n # Add a random or linear int sequence (50% prob)\n if random.random() < .5:\n # Generate a linear sequence\n rand_start = random.randint(0, max_value - len)\n s = [[float(i)/max_value] for i in\n range(rand_start, rand_start + len)]\n # Pad sequence for dimension consistency\n s += [[0.] for i in range(max_seq_len - len)]\n self.data.append(s)\n self.labels.append([1., 0.])\n else:\n # Generate a random sequence\n s = [[float(random.randint(0, max_value))/max_value]\n for i in range(len)]\n # Pad sequence for dimension consistency\n s += [[0.] for i in range(max_seq_len - len)]\n self.data.append(s)\n self.labels.append([0., 1.])\n self.batch_id = 0\n\n def next(self, batch_size):\n \"\"\" Return a batch of data. When dataset end is reached, start over.\n \"\"\"\n if self.batch_id == len(self.data):\n self.batch_id = 0\n batch_data = (self.data[self.batch_id:min(self.batch_id +\n batch_size, len(self.data))])\n batch_labels = (self.labels[self.batch_id:min(self.batch_id +\n batch_size, len(self.data))])\n batch_seqlen = (self.seqlen[self.batch_id:min(self.batch_id +\n batch_size, len(self.data))])\n self.batch_id = min(self.batch_id + batch_size, len(self.data))\n return batch_data, batch_labels, batch_seqlen", "_____no_output_____" ], [ "# ==========\n# MODEL\n# ==========\n\n# Parameters\nlearning_rate = 0.01\ntraining_steps = 10000\nbatch_size = 128\ndisplay_step = 200\n\n# Network Parameters\nseq_max_len = 20 # Sequence max length\nn_hidden = 64 # hidden layer num of features\nn_classes = 2 # linear sequence or not\n\ntrainset = ToySequenceData(n_samples=1000, max_seq_len=seq_max_len)\ntestset = ToySequenceData(n_samples=500, max_seq_len=seq_max_len)\n\n# tf Graph input\nx = tf.placeholder(\"float\", [None, seq_max_len, 1])\ny = tf.placeholder(\"float\", [None, n_classes])\n# A placeholder for indicating each sequence length\nseqlen = tf.placeholder(tf.int32, [None])\n\n# Define weights\nweights = {\n 'out': tf.Variable(tf.random_normal([n_hidden, n_classes]))\n}\nbiases = {\n 'out': tf.Variable(tf.random_normal([n_classes]))\n}", "_____no_output_____" ], [ "def dynamicRNN(x, seqlen, weights, biases):\n\n # Prepare data shape to match `rnn` function requirements\n # Current data input shape: (batch_size, n_steps, n_input)\n # Required shape: 'n_steps' tensors list of shape (batch_size, n_input)\n \n # Unstack to get a list of 'n_steps' tensors of shape (batch_size, n_input)\n x = tf.unstack(x, seq_max_len, 1)\n\n # Define a lstm cell with tensorflow\n lstm_cell = tf.contrib.rnn.BasicLSTMCell(n_hidden)\n\n # Get lstm cell output, providing 'sequence_length' will perform dynamic\n # calculation.\n outputs, states = tf.contrib.rnn.static_rnn(lstm_cell, x, dtype=tf.float32,\n sequence_length=seqlen)\n\n # When performing dynamic calculation, we must retrieve the last\n # dynamically computed output, i.e., if a sequence length is 10, we need\n # to retrieve the 10th output.\n # However TensorFlow doesn't support advanced indexing yet, so we build\n # a custom op that for each sample in batch size, get its length and\n # get the corresponding relevant output.\n\n # 'outputs' is a list of output at every timestep, we pack them in a Tensor\n # and change back dimension to [batch_size, n_step, n_input]\n outputs = tf.stack(outputs)\n outputs = tf.transpose(outputs, [1, 0, 2])\n\n # Hack to build the indexing and retrieve the right output.\n batch_size = tf.shape(outputs)[0]\n # Start indices for each sample\n index = tf.range(0, batch_size) * seq_max_len + (seqlen - 1)\n # Indexing\n outputs = tf.gather(tf.reshape(outputs, [-1, n_hidden]), index)\n\n # Linear activation, using outputs computed above\n return tf.matmul(outputs, weights['out']) + biases['out']", "_____no_output_____" ], [ "pred = dynamicRNN(x, seqlen, weights, biases)\n\n# Define loss and optimizer\ncost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))\noptimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate).minimize(cost)\n\n# Evaluate model\ncorrect_pred = tf.equal(tf.argmax(pred,1), tf.argmax(y,1))\naccuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))\n\n# Initialize the variables (i.e. assign their default value)\ninit = tf.global_variables_initializer()", "/Users/aymeric.damien/anaconda2/lib/python2.7/site-packages/tensorflow/python/ops/gradients_impl.py:93: UserWarning: Converting sparse IndexedSlices to a dense Tensor of unknown shape. This may consume a large amount of memory.\n \"Converting sparse IndexedSlices to a dense Tensor of unknown shape. \"\n" ], [ "# Start training\nwith tf.Session() as sess:\n\n # Run the initializer\n sess.run(init)\n\n for step in range(1, training_steps+1):\n batch_x, batch_y, batch_seqlen = trainset.next(batch_size)\n # Run optimization op (backprop)\n sess.run(optimizer, feed_dict={x: batch_x, y: batch_y,\n seqlen: batch_seqlen})\n if step % display_step == 0 or step == 1:\n # Calculate batch accuracy & loss\n acc, loss = sess.run([accuracy, cost], feed_dict={x: batch_x, y: batch_y,\n seqlen: batch_seqlen})\n print(\"Step \" + str(step) + \", Minibatch Loss= \" + \\\n \"{:.6f}\".format(loss) + \", Training Accuracy= \" + \\\n \"{:.5f}\".format(acc))\n\n print(\"Optimization Finished!\")\n\n # Calculate accuracy\n test_data = testset.data\n test_label = testset.labels\n test_seqlen = testset.seqlen\n print(\"Testing Accuracy:\", \\\n sess.run(accuracy, feed_dict={x: test_data, y: test_label,\n seqlen: test_seqlen}))", "Step 1, Minibatch Loss= 0.864517, Training Accuracy= 0.42188\nStep 200, Minibatch Loss= 0.686012, Training Accuracy= 0.43269\nStep 400, Minibatch Loss= 0.682970, Training Accuracy= 0.48077\nStep 600, Minibatch Loss= 0.679640, Training Accuracy= 0.50962\nStep 800, Minibatch Loss= 0.675208, Training Accuracy= 0.53846\nStep 1000, Minibatch Loss= 0.668636, Training Accuracy= 0.56731\nStep 1200, Minibatch Loss= 0.657525, Training Accuracy= 0.62500\nStep 1400, Minibatch Loss= 0.635423, Training Accuracy= 0.67308\nStep 1600, Minibatch Loss= 0.580433, Training Accuracy= 0.75962\nStep 1800, Minibatch Loss= 0.475599, Training Accuracy= 0.81731\nStep 2000, Minibatch Loss= 0.434865, Training Accuracy= 0.83654\nStep 2200, Minibatch Loss= 0.423690, Training Accuracy= 0.85577\nStep 2400, Minibatch Loss= 0.417472, Training Accuracy= 0.85577\nStep 2600, Minibatch Loss= 0.412906, Training Accuracy= 0.85577\nStep 2800, Minibatch Loss= 0.409193, Training Accuracy= 0.85577\nStep 3000, Minibatch Loss= 0.406035, Training Accuracy= 0.86538\nStep 3200, Minibatch Loss= 0.403287, Training Accuracy= 0.87500\nStep 3400, Minibatch Loss= 0.400862, Training Accuracy= 0.87500\nStep 3600, Minibatch Loss= 0.398704, Training Accuracy= 0.86538\nStep 3800, Minibatch Loss= 0.396768, Training Accuracy= 0.86538\nStep 4000, Minibatch Loss= 0.395017, Training Accuracy= 0.86538\nStep 4200, Minibatch Loss= 0.393422, Training Accuracy= 0.86538\nStep 4400, Minibatch Loss= 0.391957, Training Accuracy= 0.85577\nStep 4600, Minibatch Loss= 0.390600, Training Accuracy= 0.85577\nStep 4800, Minibatch Loss= 0.389334, Training Accuracy= 0.86538\nStep 5000, Minibatch Loss= 0.388143, Training Accuracy= 0.86538\nStep 5200, Minibatch Loss= 0.387015, Training Accuracy= 0.86538\nStep 5400, Minibatch Loss= 0.385940, Training Accuracy= 0.86538\nStep 5600, Minibatch Loss= 0.384907, Training Accuracy= 0.86538\nStep 5800, Minibatch Loss= 0.383904, Training Accuracy= 0.85577\nStep 6000, Minibatch Loss= 0.382921, Training Accuracy= 0.86538\nStep 6200, Minibatch Loss= 0.381941, Training Accuracy= 0.86538\nStep 6400, Minibatch Loss= 0.380947, Training Accuracy= 0.86538\nStep 6600, Minibatch Loss= 0.379912, Training Accuracy= 0.86538\nStep 6800, Minibatch Loss= 0.378796, Training Accuracy= 0.86538\nStep 7000, Minibatch Loss= 0.377540, Training Accuracy= 0.86538\nStep 7200, Minibatch Loss= 0.376041, Training Accuracy= 0.86538\nStep 7400, Minibatch Loss= 0.374130, Training Accuracy= 0.85577\nStep 7600, Minibatch Loss= 0.371514, Training Accuracy= 0.85577\nStep 7800, Minibatch Loss= 0.367723, Training Accuracy= 0.85577\nStep 8000, Minibatch Loss= 0.362049, Training Accuracy= 0.85577\nStep 8200, Minibatch Loss= 0.353558, Training Accuracy= 0.85577\nStep 8400, Minibatch Loss= 0.341072, Training Accuracy= 0.86538\nStep 8600, Minibatch Loss= 0.323062, Training Accuracy= 0.87500\nStep 8800, Minibatch Loss= 0.299278, Training Accuracy= 0.89423\nStep 9000, Minibatch Loss= 0.273857, Training Accuracy= 0.90385\nStep 9200, Minibatch Loss= 0.248392, Training Accuracy= 0.91346\nStep 9400, Minibatch Loss= 0.221348, Training Accuracy= 0.92308\nStep 9600, Minibatch Loss= 0.191947, Training Accuracy= 0.92308\nStep 9800, Minibatch Loss= 0.159308, Training Accuracy= 0.93269\nStep 10000, Minibatch Loss= 0.136938, Training Accuracy= 0.96154\nOptimization Finished!\nTesting Accuracy: 0.952\n" ] ] ]

[ "markdown", "code" ]

[ [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code", "code" ] ]

4ad876b7e35395b6a88e32c434cb892759a8f0f3

463,541

ipynb

Jupyter Notebook

selection/mutation-drift.ipynb

tethig/sismid

6104195619ae2ddbd91921872ffde2b7c83b01ac

[ "CC-BY-4.0", "MIT" ]

null

selection/mutation-drift.ipynb

tethig/sismid

6104195619ae2ddbd91921872ffde2b7c83b01ac

[ "CC-BY-4.0", "MIT" ]

null

selection/mutation-drift.ipynb

tethig/sismid

6104195619ae2ddbd91921872ffde2b7c83b01ac

[ "CC-BY-4.0", "MIT" ]

null

266.862982

332,082

0.919099

[ [ [ "# Wright-Fisher model of mutation and random genetic drift", "_____no_output_____" ], [ "A Wright-Fisher model has a fixed population size *N* and discrete non-overlapping generations. Each generation, each individual has a random number of offspring whose mean is proportional to the individual's fitness. Each generation, mutation may occur.", "_____no_output_____" ], [ "## Setup", "_____no_output_____" ] ], [ [ "import numpy as np\ntry:\n import itertools.izip as zip\nexcept ImportError:\n import itertools", "_____no_output_____" ] ], [ [ "## Make population dynamic model", "_____no_output_____" ], [ "### Basic parameters", "_____no_output_____" ] ], [ [ "pop_size = 100", "_____no_output_____" ], [ "seq_length = 10", "_____no_output_____" ], [ "alphabet = ['A', 'T', 'G', 'C']", "_____no_output_____" ], [ "base_haplotype = \"AAAAAAAAAA\"", "_____no_output_____" ] ], [ [ "### Setup a population of sequences", "_____no_output_____" ], [ "Store this as a lightweight Dictionary that maps a string to a count. All the sequences together will have count *N*.", "_____no_output_____" ] ], [ [ "pop = {}", "_____no_output_____" ], [ "pop[\"AAAAAAAAAA\"] = 40", "_____no_output_____" ], [ "pop[\"AAATAAAAAA\"] = 30", "_____no_output_____" ], [ "pop[\"AATTTAAAAA\"] = 30", "_____no_output_____" ], [ "pop[\"AAATAAAAAA\"]", "_____no_output_____" ] ], [ [ "### Add mutation", "_____no_output_____" ], [ "Mutations occur each generation in each individual in every basepair.", "_____no_output_____" ] ], [ [ "mutation_rate = 0.005 # per gen per individual per site", "_____no_output_____" ] ], [ [ "Walk through population and mutate basepairs. Use Poisson splitting to speed this up (you may be familiar with Poisson splitting from its use in the [Gillespie algorithm](https://en.wikipedia.org/wiki/Gillespie_algorithm)). \n\n * In naive scenario A: take each element and check for each if event occurs. For example, 100 elements, each with 1% chance. This requires 100 random numbers.\n * In Poisson splitting scenario B: Draw a Poisson random number for the number of events that occur and distribute them randomly. In the above example, this will most likely involve 1 random number draw to see how many events and then a few more draws to see which elements are hit.", "_____no_output_____" ], [ "First off, we need to get random number of total mutations", "_____no_output_____" ] ], [ [ "def get_mutation_count():\n mean = mutation_rate * pop_size * seq_length\n return np.random.poisson(mean)", "_____no_output_____" ] ], [ [ "Here we use Numpy's [Poisson random number](http://docs.scipy.org/doc/numpy/reference/generated/numpy.random.poisson.html).", "_____no_output_____" ] ], [ [ "get_mutation_count()", "_____no_output_____" ] ], [ [ "We need to get random haplotype from the population.", "_____no_output_____" ] ], [ [ "pop.keys()", "_____no_output_____" ], [ "[x/float(pop_size) for x in pop.values()]", "_____no_output_____" ], [ "def get_random_haplotype():\n haplotypes = list(pop.keys()) \n frequencies = [x/float(pop_size) for x in pop.values()]\n total = sum(frequencies)\n frequencies = [x / total for x in frequencies]\n return np.random.choice(haplotypes, p=frequencies)", "_____no_output_____" ] ], [ [ "Here we use Numpy's [weighted random choice](http://docs.scipy.org/doc/numpy/reference/generated/numpy.random.choice.html).", "_____no_output_____" ] ], [ [ "get_random_haplotype()", "_____no_output_____" ] ], [ [ "Here, we take a supplied haplotype and mutate a site at random.", "_____no_output_____" ] ], [ [ "def get_mutant(haplotype):\n site = np.random.randint(seq_length)\n possible_mutations = list(alphabet)\n possible_mutations.remove(haplotype[site])\n mutation = np.random.choice(possible_mutations)\n new_haplotype = haplotype[:site] + mutation + haplotype[site+1:]\n return new_haplotype", "_____no_output_____" ], [ "get_mutant(\"AAAAAAAAAA\")", "_____no_output_____" ] ], [ [ "Putting things together, in a single mutation event, we grab a random haplotype from the population, mutate it, decrement its count, and then check if the mutant already exists in the population. If it does, increment this mutant haplotype; if it doesn't create a new haplotype of count 1. ", "_____no_output_____" ] ], [ [ "def mutation_event():\n haplotype = get_random_haplotype()\n if pop[haplotype] > 1:\n pop[haplotype] -= 1\n new_haplotype = get_mutant(haplotype)\n if new_haplotype in pop:\n pop[new_haplotype] += 1\n else:\n pop[new_haplotype] = 1", "_____no_output_____" ], [ "mutation_event()", "_____no_output_____" ], [ "pop", "_____no_output_____" ] ], [ [ "To create all the mutations that occur in a single generation, we draw the total count of mutations and then iteratively add mutation events.", "_____no_output_____" ] ], [ [ "def mutation_step():\n mutation_count = get_mutation_count()\n for i in range(mutation_count):\n mutation_event()", "_____no_output_____" ], [ "mutation_step()", "_____no_output_____" ], [ "pop", "_____no_output_____" ] ], [ [ "### Add genetic drift", "_____no_output_____" ], [ "Given a list of haplotype frequencies currently in the population, we can take a [multinomial draw](https://en.wikipedia.org/wiki/Multinomial_distribution) to get haplotype counts in the following generation.", "_____no_output_____" ] ], [ [ "def get_offspring_counts():\n haplotypes = list(pop.keys())\n frequencies = [x/float(pop_size) for x in pop.values()]\n return list(np.random.multinomial(pop_size, frequencies))", "_____no_output_____" ] ], [ [ "Here we use Numpy's [multinomial random sample](http://docs.scipy.org/doc/numpy/reference/generated/numpy.random.multinomial.html).", "_____no_output_____" ] ], [ [ "get_offspring_counts()", "_____no_output_____" ] ], [ [ "We then need to assign this new list of haplotype counts to the `pop` dictionary. To save memory and computation, if a haplotype goes to 0, we remove it entirely from the `pop` dictionary.", "_____no_output_____" ] ], [ [ "def offspring_step():\n haplotypes = list(pop.keys())\n counts = get_offspring_counts()\n for (haplotype, count) in zip(haplotypes, counts):\n if (count > 0):\n pop[haplotype] = count\n else:\n del pop[haplotype]", "_____no_output_____" ], [ "offspring_step()", "_____no_output_____" ], [ "pop", "_____no_output_____" ] ], [ [ "### Combine and iterate", "_____no_output_____" ], [ "Each generation is simply a mutation step where a random number of mutations are thrown down, and an offspring step where haplotype counts are updated.", "_____no_output_____" ] ], [ [ "def time_step():\n mutation_step()\n offspring_step()", "_____no_output_____" ] ], [ [ "Can iterate this over a number of generations.", "_____no_output_____" ] ], [ [ "generations = 5", "_____no_output_____" ], [ "def simulate():\n for i in range(generations):\n time_step()", "_____no_output_____" ], [ "simulate()", "_____no_output_____" ], [ "pop", "_____no_output_____" ] ], [ [ "### Record", "_____no_output_____" ], [ "We want to keep a record of past population frequencies to understand dynamics through time. At each step in the simulation, we append to a history object.", "_____no_output_____" ] ], [ [ "pop = {\"AAAAAAAAAA\": pop_size}", "_____no_output_____" ], [ "history = []", "_____no_output_____" ], [ "def simulate():\n clone_pop = dict(pop)\n history.append(clone_pop)\n for i in range(generations):\n time_step()\n clone_pop = dict(pop)\n history.append(clone_pop)", "_____no_output_____" ], [ "simulate()", "_____no_output_____" ], [ "pop", "_____no_output_____" ], [ "history[0]", "_____no_output_____" ], [ "history[1]", "_____no_output_____" ], [ "history[2]", "_____no_output_____" ] ], [ [ "## Analyze trajectories", "_____no_output_____" ], [ "### Calculate diversity", "_____no_output_____" ], [ "Here, diversity in population genetics is usually shorthand for the statistic *π*, which measures pairwise differences between random individuals in the population. *π* is usually measured as substitutions per site.", "_____no_output_____" ] ], [ [ "pop", "_____no_output_____" ] ], [ [ "First, we need to calculate the number of differences per site between two arbitrary sequences.", "_____no_output_____" ] ], [ [ "def get_distance(seq_a, seq_b):\n diffs = 0\n length = len(seq_a)\n assert len(seq_a) == len(seq_b)\n for chr_a, chr_b in zip(seq_a, seq_b):\n if chr_a != chr_b:\n diffs += 1\n return diffs / float(length)", "_____no_output_____" ], [ "get_distance(\"AAAAAAAAAA\", \"AAAAAAAAAB\")", "_____no_output_____" ] ], [ [ "We calculate diversity as a weighted average between all pairs of haplotypes, weighted by pairwise haplotype frequency.", "_____no_output_____" ] ], [ [ "def get_diversity(population):\n haplotypes = list(population.keys())\n haplotype_count = len(haplotypes)\n diversity = 0\n for i in range(haplotype_count):\n for j in range(haplotype_count):\n haplotype_a = haplotypes[i]\n haplotype_b = haplotypes[j]\n frequency_a = population[haplotype_a] / float(pop_size)\n frequency_b = population[haplotype_b] / float(pop_size)\n frequency_pair = frequency_a * frequency_b\n diversity += frequency_pair * get_distance(haplotype_a, haplotype_b)\n return diversity", "_____no_output_____" ], [ "get_diversity(pop)", "_____no_output_____" ], [ "def get_diversity_trajectory():\n trajectory = [get_diversity(generation) for generation in history]\n return trajectory", "_____no_output_____" ], [ "get_diversity_trajectory()", "_____no_output_____" ] ], [ [ "### Plot diversity", "_____no_output_____" ], [ "Here, we use [matplotlib](http://matplotlib.org/) for all Python plotting.", "_____no_output_____" ] ], [ [ "%matplotlib inline\nimport matplotlib.pyplot as plt\nimport matplotlib as mpl", "_____no_output_____" ] ], [ [ "Here, we make a simple line plot using matplotlib's `plot` function.", "_____no_output_____" ] ], [ [ "plt.plot(get_diversity_trajectory())", "_____no_output_____" ] ], [ [ "Here, we style the plot a bit with x and y axes labels.", "_____no_output_____" ] ], [ [ "def diversity_plot():\n mpl.rcParams['font.size']=14\n trajectory = get_diversity_trajectory()\n plt.plot(trajectory, \"#447CCD\") \n plt.ylabel(\"diversity\")\n plt.xlabel(\"generation\")", "_____no_output_____" ], [ "diversity_plot()", "_____no_output_____" ] ], [ [ "### Analyze and plot divergence", "_____no_output_____" ], [ "In population genetics, divergence is generally the number of substitutions away from a reference sequence. In this case, we can measure the average distance of the population to the starting haplotype. Again, this will be measured in terms of substitutions per site.", "_____no_output_____" ] ], [ [ "def get_divergence(population):\n haplotypes = population.keys()\n divergence = 0\n for haplotype in haplotypes:\n frequency = population[haplotype] / float(pop_size)\n divergence += frequency * get_distance(base_haplotype, haplotype)\n return divergence", "_____no_output_____" ], [ "def get_divergence_trajectory():\n trajectory = [get_divergence(generation) for generation in history]\n return trajectory", "_____no_output_____" ], [ "get_divergence_trajectory()", "_____no_output_____" ], [ "def divergence_plot():\n mpl.rcParams['font.size']=14\n trajectory = get_divergence_trajectory()\n plt.plot(trajectory, \"#447CCD\")\n plt.ylabel(\"divergence\")\n plt.xlabel(\"generation\") ", "_____no_output_____" ], [ "divergence_plot()", "_____no_output_____" ] ], [ [ "### Plot haplotype trajectories", "_____no_output_____" ], [ "We also want to directly look at haplotype frequencies through time.", "_____no_output_____" ] ], [ [ "def get_frequency(haplotype, generation):\n pop_at_generation = history[generation]\n if haplotype in pop_at_generation:\n return pop_at_generation[haplotype]/float(pop_size)\n else:\n return 0", "_____no_output_____" ], [ "get_frequency(\"AAAAAAAAAA\", 4)", "_____no_output_____" ], [ "def get_trajectory(haplotype):\n trajectory = [get_frequency(haplotype, gen) for gen in range(generations)]\n return trajectory", "_____no_output_____" ], [ "get_trajectory(\"AAAAAAAAAA\")", "_____no_output_____" ] ], [ [ "We want to plot all haplotypes seen during the simulation.", "_____no_output_____" ] ], [ [ "def get_all_haplotypes():\n haplotypes = set() \n for generation in history:\n for haplotype in generation:\n haplotypes.add(haplotype)\n return haplotypes", "_____no_output_____" ], [ "get_all_haplotypes()", "_____no_output_____" ] ], [ [ "Here is a simple plot of their overall frequencies.", "_____no_output_____" ] ], [ [ "haplotypes = get_all_haplotypes()\nfor haplotype in haplotypes:\n plt.plot(get_trajectory(haplotype))\nplt.show()", "_____no_output_____" ], [ "colors = [\"#781C86\", \"#571EA2\", \"#462EB9\", \"#3F47C9\", \"#3F63CF\", \"#447CCD\", \"#4C90C0\", \"#56A0AE\", \"#63AC9A\", \"#72B485\", \"#83BA70\", \"#96BD60\", \"#AABD52\", \"#BDBB48\", \"#CEB541\", \"#DCAB3C\", \"#E49938\", \"#E68133\", \"#E4632E\", \"#DF4327\", \"#DB2122\"]", "_____no_output_____" ], [ "colors_lighter = [\"#A567AF\", \"#8F69C1\", \"#8474D1\", \"#7F85DB\", \"#7F97DF\", \"#82A8DD\", \"#88B5D5\", \"#8FC0C9\", \"#97C8BC\", \"#A1CDAD\", \"#ACD1A0\", \"#B9D395\", \"#C6D38C\", \"#D3D285\", \"#DECE81\", \"#E8C77D\", \"#EDBB7A\", \"#EEAB77\", \"#ED9773\", \"#EA816F\", \"#E76B6B\"]", "_____no_output_____" ] ], [ [ "We can use `stackplot` to stack these trajectoies on top of each other to get a better picture of what's going on.", "_____no_output_____" ] ], [ [ "def stacked_trajectory_plot(xlabel=\"generation\"):\n mpl.rcParams['font.size']=18\n haplotypes = get_all_haplotypes()\n trajectories = [get_trajectory(haplotype) for haplotype in haplotypes]\n plt.stackplot(range(generations), trajectories, colors=colors_lighter)\n plt.ylim(0, 1)\n plt.ylabel(\"frequency\")\n plt.xlabel(xlabel)", "_____no_output_____" ], [ "stacked_trajectory_plot()", "_____no_output_____" ] ], [ [ "### Plot SNP trajectories", "_____no_output_____" ] ], [ [ "def get_snp_frequency(site, generation):\n minor_allele_frequency = 0.0\n pop_at_generation = history[generation]\n for haplotype in pop_at_generation.keys():\n allele = haplotype[site]\n frequency = pop_at_generation[haplotype] / float(pop_size)\n if allele != \"A\":\n minor_allele_frequency += frequency\n return minor_allele_frequency", "_____no_output_____" ], [ "get_snp_frequency(3, 5)", "_____no_output_____" ], [ "def get_snp_trajectory(site):\n trajectory = [get_snp_frequency(site, gen) for gen in range(generations)]\n return trajectory", "_____no_output_____" ], [ "get_snp_trajectory(3)", "_____no_output_____" ] ], [ [ "Find all variable sites.", "_____no_output_____" ] ], [ [ "def get_all_snps():\n snps = set() \n for generation in history:\n for haplotype in generation:\n for site in range(seq_length):\n if haplotype[site] != \"A\":\n snps.add(site)\n return snps", "_____no_output_____" ], [ "def snp_trajectory_plot(xlabel=\"generation\"):\n mpl.rcParams['font.size']=18\n snps = get_all_snps()\n trajectories = [get_snp_trajectory(snp) for snp in snps]\n data = []\n for trajectory, color in zip(trajectories, itertools.cycle(colors)):\n data.append(range(generations))\n data.append(trajectory) \n data.append(color)\n plt.plot(*data) \n plt.ylim(0, 1)\n plt.ylabel(\"frequency\")\n plt.xlabel(xlabel)", "_____no_output_____" ], [ "snp_trajectory_plot()", "_____no_output_____" ] ], [ [ "## Scale up", "_____no_output_____" ], [ "Here, we scale up to more interesting parameter values.", "_____no_output_____" ] ], [ [ "pop_size = 50\nseq_length = 100\ngenerations = 500\nmutation_rate = 0.0001 # per gen per individual per site", "_____no_output_____" ] ], [ [ "In this case there are $\\mu$ = 0.01 mutations entering the population every generation.", "_____no_output_____" ] ], [ [ "seq_length * mutation_rate", "_____no_output_____" ] ], [ [ "And the population genetic parameter $\\theta$, which equals $2N\\mu$, is 1.", "_____no_output_____" ] ], [ [ "2 * pop_size * seq_length * mutation_rate", "_____no_output_____" ], [ "base_haplotype = ''.join([\"A\" for i in range(seq_length)])\npop.clear()\ndel history[:]\npop[base_haplotype] = pop_size", "_____no_output_____" ], [ "simulate()", "_____no_output_____" ], [ "plt.figure(num=None, figsize=(14, 14), dpi=80, facecolor='w', edgecolor='k')\nplt.subplot2grid((3,2), (0,0), colspan=2)\nstacked_trajectory_plot(xlabel=\"\")\nplt.subplot2grid((3,2), (1,0), colspan=2)\nsnp_trajectory_plot(xlabel=\"\")\nplt.subplot2grid((3,2), (2,0))\ndiversity_plot()\nplt.subplot2grid((3,2), (2,1))\ndivergence_plot()", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code" ] ]

4ad87713efc29e73e4d576d04c78fdadcbc9192a

64,193

ipynb

Jupyter Notebook

labs/laboratorio_04.ipynb

MaximilianoRamirezN/mat281_portfolio

b7e62a45216f5fa96569eee57ac6405316aafc79

[ "MIT" ]

1

2021-12-06T01:03:51.000Z

labs/laboratorio_04.ipynb

MaximilianoRamirezN/mat281_portfolio

b7e62a45216f5fa96569eee57ac6405316aafc79

[ "MIT" ]

null

labs/laboratorio_04.ipynb

MaximilianoRamirezN/mat281_portfolio

b7e62a45216f5fa96569eee57ac6405316aafc79

[ "MIT" ]

null

31.778713

165

0.304519

[ [ [ "<img src=\"images/usm.jpg\" width=\"480\" height=\"240\" align=\"left\"/>", "_____no_output_____" ], [ "# MAT281 - Laboratorio N°04\n\n## Objetivos de la clase\n\n* Reforzar los conceptos básicos de los módulos de pandas.", "_____no_output_____" ], [ "## Contenidos\n\n* [Problema 01](#p1)\n* [Problema 02](#p2)\n", "_____no_output_____" ], [ "## Problema 01\n\n\n<img src=\"https://image.freepik.com/vector-gratis/varios-automoviles-dibujos-animados_23-2147613095.jpg\" width=\"360\" height=\"360\" align=\"center\"/>\n\n\nEL conjunto de datos se denomina `Automobile_data.csv`, el cual contiene información tal como: compañia, precio, kilometraje, etc.\n\nLo primero es cargar el conjunto de datos y ver las primeras filas que lo componen:", "_____no_output_____" ] ], [ [ "import pandas as pd\nimport numpy as np\nimport os", "_____no_output_____" ], [ "# cargar datos\ndf = pd.read_csv(os.path.join(\"data\",\"Automobile_data.csv\")).set_index('index')\ndf.head()", "_____no_output_____" ] ], [ [ "El objetivo es tratar de obtener la mayor información posible de este conjunto de datos. Para cumplir este objetivo debe resolver las siguientes problemáticas:", "_____no_output_____" ], [ "1. Elimine los valores nulos (Nan)", "_____no_output_____" ] ], [ [ "df=df.dropna()", "_____no_output_____" ] ], [ [ "2. Encuentra el nombre de la compañía de automóviles más cara", "_____no_output_____" ] ], [ [ "df.groupby(['company']).mean()['price'].idxmax()", "_____no_output_____" ] ], [ [ "3. Imprimir todos los detalles de Toyota Cars", "_____no_output_____" ] ], [ [ "df[df['company']=='toyota']", "_____no_output_____" ] ], [ [ "4. Cuente el total de automóviles por compañía", "_____no_output_____" ] ], [ [ "comp[['company']].count()", "_____no_output_____" ] ], [ [ "5. Encuentra el coche con el precio más alto por compañía", "_____no_output_____" ] ], [ [ "df[['company','price']].groupby('company').max()", "_____no_output_____" ] ], [ [ "6. Encuentre el kilometraje promedio (**average-mileage**) de cada compañía automotriz", "_____no_output_____" ] ], [ [ "comp=df.groupby('company')\ndf_leng = comp.agg({'average-mileage':[np.mean]}).reset_index()\ndf_leng ", "_____no_output_____" ] ], [ [ "7. Ordenar todos los autos por columna de precio (**price**)", "_____no_output_____" ] ], [ [ "df.sort_values(by='price')", "_____no_output_____" ] ], [ [ "## Problema 02\n\n\nSiguiendo la temática de los automóviles, resuelva los siguientes problemas:\n\n#### a) Subproblema 01\nA partir de los siguientes diccionarios:", "_____no_output_____" ] ], [ [ "GermanCars = {'Company': ['Ford', 'Mercedes', 'BMV', 'Audi'],\n 'Price': [23845, 171995, 135925 , 71400]}\njapaneseCars = {'Company': ['Toyota', 'Honda', 'Nissan', 'Mitsubishi '], \n 'Price': [29995, 23600, 61500 , 58900]}", "_____no_output_____" ] ], [ [ "* Cree dos dataframes (**carsDf1** y **carsDf2**) según corresponda.\n* Concatene ambos dataframes (**carsDf**) y añada una llave [\"Germany\", \"Japan\"], según corresponda.", "_____no_output_____" ] ], [ [ "carsDf1=pd.DataFrame(GermanCars)\ncarsDf2=pd.DataFrame(japaneseCars)\n\ncarsDf= pd.concat([carsDf1,carsDf2],keys=['Germany','Japan']) \ncarsDf", "_____no_output_____" ] ], [ [ "#### b) Subproblema 02\n\nA partir de los siguientes diccionarios:", "_____no_output_____" ] ], [ [ "Car_Price = {'Company': ['Toyota', 'Honda', 'BMV', 'Audi'], 'Price': [23845, 17995, 135925 , 71400]}\ncar_Horsepower = {'Company': ['Toyota', 'Honda', 'BMV', 'Audi'], 'horsepower': [141, 80, 182 , 160]}", "_____no_output_____" ] ], [ [ "* Cree dos dataframes (**carsDf1** y **carsDf2**) según corresponda.\n* Junte ambos dataframes (**carsDf**) por la llave **Company**.", "_____no_output_____" ] ], [ [ "carsDf1=pd.DataFrame(Car_Price)\ncarsDf2=pd.DataFrame(car_Horsepower)\n\ncarsDf= pd.merge(carsDf1, carsDf2, on='Company')\ncarsDf", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown", "markdown", "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad88f0448ee1c7eec21ad6856f89527cbfb9506

274,772

ipynb

Jupyter Notebook

Datasets/jigsaw-dataset-split-pb-roberta-large-192-ratio-2.ipynb

dimitreOliveira/Jigsaw-Multilingual-Toxic-Comment-Classification

44422e6aeeff227e22dbb5c05101322e9d4aabbe

[ "MIT" ]

4

2020-06-23T02:31:07.000Z

2020-07-04T11:50:08.000Z

Datasets/jigsaw-dataset-split-pb-roberta-large-192-ratio-2.ipynb

dimitreOliveira/Jigsaw-Multilingual-Toxic-Comment-Classification

44422e6aeeff227e22dbb5c05101322e9d4aabbe

[ "MIT" ]

null

Datasets/jigsaw-dataset-split-pb-roberta-large-192-ratio-2.ipynb

dimitreOliveira/Jigsaw-Multilingual-Toxic-Comment-Classification

44422e6aeeff227e22dbb5c05101322e9d4aabbe

[ "MIT" ]

null

189.759669

29,264

0.890859

[ [ [ "# Dependencies", "_____no_output_____" ] ], [ [ "import os, warnings, shutil\nimport numpy as np\nimport pandas as pd\nimport seaborn as sns\nimport matplotlib.pyplot as plt\nfrom transformers import AutoTokenizer\nfrom sklearn.utils import shuffle\nfrom sklearn.model_selection import StratifiedKFold\n\n\nSEED = 0\nwarnings.filterwarnings(\"ignore\")", "_____no_output_____" ] ], [ [ "# Parameters", "_____no_output_____" ] ], [ [ "MAX_LEN = 192\ntokenizer_path = 'jplu/tf-xlm-roberta-large'\n\npositive1 = 21384 * 2\npositive2 = 112226 * 2", "_____no_output_____" ] ], [ [ "# Load data", "_____no_output_____" ] ], [ [ "train1 = pd.read_csv(\"/kaggle/input/jigsaw-multilingual-toxic-comment-classification/jigsaw-toxic-comment-train.csv\")\ntrain2 = pd.read_csv(\"/kaggle/input/jigsaw-multilingual-toxic-comment-classification/jigsaw-unintended-bias-train.csv\")\n\ntrain_df = pd.concat([train1[['comment_text', 'toxic']].query('toxic == 1'),\n train1[['comment_text', 'toxic']].query('toxic == 0').sample(n=positive1, random_state=SEED),\n train2[['comment_text', 'toxic']].query('toxic > .5'),\n train2[['comment_text', 'toxic']].query('toxic <= .5').sample(n=positive2, random_state=SEED)\n ])\ntrain_df = shuffle(train_df, random_state=SEED).reset_index(drop=True)\ntrain_df['toxic_int'] = train_df['toxic'].round().astype(int)\n\nprint('Train samples %d' % len(train_df))\ndisplay(train_df.head())", "Train samples 400830\n" ] ], [ [ "# Tokenizer", "_____no_output_____" ] ], [ [ "tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)", "_____no_output_____" ] ], [ [ "# Data generation sanity check", "_____no_output_____" ] ], [ [ "for idx in range(5):\n print('\\nRow %d' % idx)\n max_seq_len = 22\n comment_text = train_df['comment_text'].loc[idx]\n \n enc = tokenizer.encode_plus(comment_text, \n return_token_type_ids=False, \n pad_to_max_length=True, \n max_length=max_seq_len)\n \n print('comment_text : \"%s\"' % comment_text)\n print('input_ids : \"%s\"' % enc['input_ids'])\n print('attention_mask: \"%s\"' % enc['attention_mask'])\n \n assert len(enc['input_ids']) == len(enc['attention_mask']) == max_seq_len", "\nRow 0\ncomment_text : \"And you have something alien on your lips Chance. Geez dude give it up. Better yet why don't you ask the Krotch brothers to buy you another drink.\"\ninput_ids : \"[0, 3493, 398, 765, 9844, 66961, 98, 935, 66265, 76710, 5, 2206, 1255, 115, 112, 8337, 442, 1257, 5, 177154, 14373, 2]\"\nattention_mask: \"[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]\"\n\nRow 1\ncomment_text : \"no income tax for anyone that comes and works and leaves. no income tax for retirees that come and work and leave. no cutting in waste. sick of them, too.\"\ninput_ids : \"[0, 110, 91763, 24643, 100, 35672, 450, 32497, 136, 43240, 136, 31358, 7, 5, 110, 91763, 24643, 100, 90223, 90, 450, 2]\"\nattention_mask: \"[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]\"\n\nRow 2\ncomment_text : \"Rubbish, Brandon. Nothing stopped Comey of the FBI of dumping the phony story about a new investigation in Hillary's email at a time when he knew it would influence the election. \nNothing stopped Putin's trolls from putting out fake news articles to undermine the Democrats. \nNothing either stopped Trump from lying over seventy percent of the time during the campaign. \nSo, even if this turns out to be untrue, it's completely fair to publish it.\"\ninput_ids : \"[0, 65787, 6454, 127, 4, 23243, 191, 5, 182747, 118066, 10592, 53, 111, 70, 83085, 111, 115, 44287, 70, 53073, 299, 2]\"\nattention_mask: \"[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]\"\n\nRow 3\ncomment_text : \"poster \n\nI have sent the e-mail. -Garrett\"\ninput_ids : \"[0, 40908, 87, 765, 9325, 70, 28, 9, 2065, 5, 20, 82166, 20574, 2, 1, 1, 1, 1, 1, 1, 1, 1]\"\nattention_mask: \"[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0]\"\n\nRow 4\ncomment_text : \"== ECB exposure == \n\n I responded to a comment of yours here. Cheers.\"\ninput_ids : \"[0, 6, 69112, 241, 47307, 197826, 6, 69112, 87, 45739, 71, 47, 10, 6868, 111, 935, 7, 3688, 5, 5024, 1314, 2]\"\nattention_mask: \"[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]\"\n" ] ], [ [ "# 5-Fold split", "_____no_output_____" ] ], [ [ "folds = StratifiedKFold(n_splits=5, shuffle=True, random_state=SEED)\n\nfor fold_n, (train_idx, val_idx) in enumerate(folds.split(train_df, train_df['toxic_int'])):\n print('Fold: %s, Train size: %s, Validation size %s' % (fold_n+1, len(train_idx), len(val_idx)))\n train_df[('fold_%s' % str(fold_n+1))] = 0\n train_df[('fold_%s' % str(fold_n+1))].loc[train_idx] = 'train'\n train_df[('fold_%s' % str(fold_n+1))].loc[val_idx] = 'validation'", "Fold: 1, Train size: 320664, Validation size 80166\nFold: 2, Train size: 320664, Validation size 80166\nFold: 3, Train size: 320664, Validation size 80166\nFold: 4, Train size: 320664, Validation size 80166\nFold: 5, Train size: 320664, Validation size 80166\n" ] ], [ [ "# Label distribution", "_____no_output_____" ] ], [ [ "for fold_n in range(folds.n_splits):\n fold_n += 1\n fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(18, 6))\n fig.suptitle('Fold %s' % fold_n, fontsize=22) \n sns.countplot(x=\"toxic_int\", data=train_df[train_df[('fold_%s' % fold_n)] == 'train'], palette=\"GnBu_d\", ax=ax1).set_title('Train')\n sns.countplot(x=\"toxic_int\", data=train_df[train_df[('fold_%s' % fold_n)] == 'validation'], palette=\"GnBu_d\", ax=ax2).set_title('Validation')\n sns.despine()\n plt.show()", "_____no_output_____" ], [ "for fold_n in range(folds.n_splits):\n fold_n += 1\n fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(18, 6))\n fig.suptitle('Fold %s' % fold_n, fontsize=22) \n sns.distplot(train_df[train_df[('fold_%s' % fold_n)] == 'train']['toxic'], ax=ax1).set_title('Train')\n sns.distplot(train_df[train_df[('fold_%s' % fold_n)] == 'validation']['toxic'], ax=ax2).set_title('Validation')\n sns.despine()\n plt.show()", "_____no_output_____" ] ], [ [ "# Output 5-fold set", "_____no_output_____" ] ], [ [ "train_df.to_csv('5-fold.csv', index=False)\ndisplay(train_df.head())\n\nfor fold_n in range(folds.n_splits):\n if fold_n < 1:\n fold_n += 1\n base_path = 'fold_%d/' % fold_n\n\n # Create dir\n os.makedirs(base_path)\n\n x_train = tokenizer.batch_encode_plus(train_df[train_df[('fold_%s' % fold_n)] == 'train']['comment_text'].values, \n return_token_type_ids=False, \n pad_to_max_length=True, \n max_length=MAX_LEN)\n\n x_train = np.array([np.array(x_train['input_ids']), \n np.array(x_train['attention_mask'])])\n\n x_valid = tokenizer.batch_encode_plus(train_df[train_df[('fold_%s' % fold_n)] == 'validation']['comment_text'].values, \n return_token_type_ids=False, \n pad_to_max_length=True, \n max_length=MAX_LEN)\n\n x_valid = np.array([np.array(x_valid['input_ids']), \n np.array(x_valid['attention_mask'])])\n\n y_train = train_df[train_df[('fold_%s' % fold_n)] == 'train']['toxic'].values\n y_valid = train_df[train_df[('fold_%s' % fold_n)] == 'validation']['toxic'].values\n\n np.save(base_path + 'x_train', np.asarray(x_train))\n np.save(base_path + 'y_train', y_train)\n np.save(base_path + 'x_valid', np.asarray(x_valid))\n np.save(base_path + 'y_valid', y_valid)\n\n print('\\nFOLD: %d' % (fold_n))\n print('x_train shape:', x_train.shape)\n print('y_train shape:', y_train.shape)\n print('x_valid shape:', x_valid.shape)\n print('y_valid shape:', y_valid.shape)\n \n# Compress logs dir\n!tar -cvzf fold_1.tar.gz fold_1\n# !tar -cvzf fold_2.tar.gz fold_2\n# !tar -cvzf fold_3.tar.gz fold_3\n# !tar -cvzf fold_4.tar.gz fold_4\n# !tar -cvzf fold_5.tar.gz fold_5\n\n# Delete logs dir\nshutil.rmtree('fold_1')\n# shutil.rmtree('fold_2')\n# shutil.rmtree('fold_3')\n# shutil.rmtree('fold_4')\n# shutil.rmtree('fold_5')", "_____no_output_____" ] ], [ [ "# Validation set", "_____no_output_____" ] ], [ [ "valid_df = pd.read_csv(\"/kaggle/input/jigsaw-multilingual-toxic-comment-classification/validation.csv\", usecols=['comment_text', 'toxic', 'lang'])\ndisplay(valid_df.head())\n\nx_valid = tokenizer.batch_encode_plus(valid_df['comment_text'].values, \n return_token_type_ids=False, \n pad_to_max_length=True, \n max_length=MAX_LEN)\n\nx_valid = np.array([np.array(x_valid['input_ids']), \n np.array(x_valid['attention_mask'])])\n\ny_valid = valid_df['toxic'].values\n\nnp.save('x_valid', np.asarray(x_valid))\nnp.save('y_valid', y_valid)\nprint('x_valid shape:', x_valid.shape)\nprint('y_valid shape:', y_valid.shape)", "_____no_output_____" ] ], [ [ "# Test set", "_____no_output_____" ] ], [ [ "test_df = pd.read_csv(\"/kaggle/input/jigsaw-multilingual-toxic-comment-classification/test.csv\", usecols=['content'])\ndisplay(test_df.head())\n\nx_test = tokenizer.batch_encode_plus(test_df['content'].values, \n return_token_type_ids=False, \n pad_to_max_length=True, \n max_length=MAX_LEN)\n\nx_test = np.array([np.array(x_test['input_ids']), \n np.array(x_test['attention_mask'])])\n\n\nnp.save('x_test', np.asarray(x_test))\nprint('x_test shape:', x_test.shape)", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad8907b8e42941cebdcd9680fed587022713577

174,572

ipynb

Jupyter Notebook

code/lqr_policy_comparisons.ipynb

benjamin-recht/benjamin-recht.github.io

fdc0c325f14ef7ddf19c18f39833588c2169bc9e

[ "MIT" ]

5

2017-12-11T16:07:35.000Z

2018-11-16T14:32:34.000Z

code/lqr_policy_comparisons.ipynb

benjamin-recht/benjamin-recht.github.io

fdc0c325f14ef7ddf19c18f39833588c2169bc9e

[ "MIT" ]

null

code/lqr_policy_comparisons.ipynb

benjamin-recht/benjamin-recht.github.io

fdc0c325f14ef7ddf19c18f39833588c2169bc9e

[ "MIT" ]

11

2016-04-11T16:31:51.000Z

2021-12-08T06:24:25.000Z

431.041975

53,662

0.923069

[ [ [ "import numpy as np\nimport lqrpols\nimport matplotlib.pyplot as plt", "_____no_output_____" ] ], [ [ "Here is a link to [lqrpols.py](http://www.argmin.net/code/lqrpols.py)\n", "_____no_output_____" ] ], [ [ "np.random.seed(1337)\n\n# state transition matrices for linear system: \n# x(t+1) = A x (t) + B u(t)\nA = np.array([[1,1],[0,1]])\nB = np.array([[0],[1]])\nd,p = B.shape\n\n# LQR quadratic cost per state\nQ = np.array([[1,0],[0,0]])\n\n# initial condition for system\nz0 = -1 # initial position\nv0 = 0 # initial velocity\nx0 = np.vstack((z0,v0))\n\nR = np.array([[1.0]])\n\n# number of time steps to simulate\nT = 10\n\n# amount of Gaussian noise in dynamics\neq_err = 1e-2", "_____no_output_____" ], [ "# N_vals = np.floor(np.linspace(1,75,num=7)).astype(int)\nN_vals = [1,2,5,7,12,25,50,75]\nN_trials = 10\n\n### Bunch of matrices for storing costs\nJ_finite_nom = np.zeros((N_trials,len(N_vals)))\nJ_finite_nomK = np.zeros((N_trials,len(N_vals)))\nJ_finite_rs = np.zeros((N_trials,len(N_vals)))\nJ_finite_ur = np.zeros((N_trials,len(N_vals)))\nJ_finite_pg = np.zeros((N_trials,len(N_vals)))\nJ_inf_nom = np.zeros((N_trials,len(N_vals)))\nJ_inf_rs = np.zeros((N_trials,len(N_vals)))\nJ_inf_ur = np.zeros((N_trials,len(N_vals)))\nJ_inf_pg = np.zeros((N_trials,len(N_vals)))\n\n\n# cost for finite time horizon, true model\nJ_finite_opt = lqrpols.cost_finite_model(A,B,Q,R,x0,T,A,B)\n\n### Solve for optimal infinite time horizon LQR controller\nK_opt = -lqrpols.lqr_gain(A,B,Q,R)\n# cost for infinite time horizon, true model\nJ_inf_opt = lqrpols.cost_inf_K(A,B,Q,R,K_opt) \n\n# cost for zero control\nbaseline = lqrpols.cost_finite_K(A,B,Q,R,x0,T,np.zeros((p,d)))\n\n# model for nominal control with 1 rollout\nA_nom1,B_nom1 = lqrpols.lsqr_estimator(A,B,Q,R,x0,eq_err,1,T)\nprint(A_nom1)\nprint(B_nom1)\n\n# cost for finite time horizon, one rollout, nominal control\none_rollout_cost = lqrpols.cost_finite_model(A,B,Q,R,x0,T,A_nom1,B_nom1)\nK_nom1 = -lqrpols.lqr_gain(A_nom1,B_nom1,Q,R)\none_rollout_cost_inf = lqrpols.cost_inf_K(A,B,Q,R,K_nom1)\n\nfor N in range(len(N_vals)):\n for trial in range(N_trials):\n \n # nominal model, N x 40 to match sample budget of policy gradient\n A_nom,B_nom = lqrpols.lsqr_estimator(A,B,Q,R,x0,eq_err,N_vals[N]*40,T);\n # finite time horizon cost with nominal model\n J_finite_nom[trial,N] = lqrpols.cost_finite_model(A,B,Q,R,x0,T,A_nom,B_nom)\n # Solve for infinite time horizon nominal LQR controller\n K_nom = -lqrpols.lqr_gain(A_nom,B_nom,Q,R)\n # cost of using the infinite time horizon solution for finite time horizon\n J_finite_nomK[trial,N] = lqrpols.cost_finite_K(A,B,Q,R,x0,T,K_nom)\n # infinite time horizon cost of nominal model\n J_inf_nom[trial,N] = lqrpols.cost_inf_K(A,B,Q,R,K_nom)\n\n # policy gradient, batchsize 40 per iteration\n K_pg = lqrpols.policy_gradient_adam_linear_policy(A,B,Q,R,x0,eq_err,N_vals[N]*5,T)\n J_finite_pg[trial,N] = lqrpols.cost_finite_K(A,B,Q,R,x0,T,K_pg)\n J_inf_pg[trial,N] = lqrpols.cost_inf_K(A,B,Q,R,K_pg)\n \n # random search, batchsize 4, so uses 8 rollouts per iteration\n K_rs = lqrpols.random_search_linear_policy(A,B,Q,R,x0,eq_err,N_vals[N]*5,T)\n J_finite_rs[trial,N] = lqrpols.cost_finite_K(A,B,Q,R,x0,T,K_rs)\n J_inf_rs[trial,N] = lqrpols.cost_inf_K(A,B,Q,R,K_rs)\n\n # uniformly random sampling, N x 40 to match sample budget of policy gradient\n K_ur = lqrpols.uniform_random_linear_policy(A,B,Q,R,x0,eq_err,N_vals[N]*40,T)\n J_finite_ur[trial,N] = lqrpols.cost_finite_K(A,B,Q,R,x0,T,K_ur)\n J_inf_ur[trial,N] = lqrpols.cost_inf_K(A,B,Q,R,K_ur)", "[[ 1.00032885e+00 9.99484822e-01]\n [ 2.11465750e-04 9.99153277e-01]]\n[[-0.00174996]\n [ 0.99998723]]\n" ], [ "colors = [ '#2D328F', '#F15C19',\"#81b13c\",\"#ca49ac\"]\n \nlabel_fontsize = 18\ntick_fontsize = 14\nlinewidth = 3\nmarkersize = 10\n\ntot_samples = 40*np.array(N_vals)\n\nplt.plot(tot_samples,np.amin(J_finite_pg,axis=0),'o-',color=colors[0],linewidth=linewidth,\n markersize=markersize,label='policy gradient')\n\nplt.plot(tot_samples,np.amin(J_finite_ur,axis=0),'>-',color=colors[1],linewidth=linewidth,\n markersize=markersize,label='uniform sampling')\n\nplt.plot(tot_samples,np.amin(J_finite_rs,axis=0),'s-',color=colors[2],linewidth=linewidth,\n markersize=markersize,label='random search')\n\nplt.plot([tot_samples[0],tot_samples[-1]],[baseline, baseline],color='#000000',linewidth=linewidth,\n linestyle='--',label='zero control')\nplt.plot([tot_samples[0],tot_samples[-1]],[J_finite_opt, J_finite_opt],color='#000000',linewidth=linewidth,\n linestyle=':',label='optimal')\n\nplt.axis([0,2000,0,12])\n\nplt.xlabel('rollouts',fontsize=label_fontsize)\nplt.ylabel('cost',fontsize=label_fontsize)\nplt.legend(fontsize=18, bbox_to_anchor=(1.0, 0.54))\nplt.xticks(fontsize=tick_fontsize)\nplt.yticks(fontsize=tick_fontsize)\nplt.grid(True)\n\nfig = plt.gcf()\nfig.set_size_inches(9, 6)\n\nplt.show()", "_____no_output_____" ], [ "plt.plot(tot_samples,np.median(J_finite_pg,axis=0),'o-',color=colors[0],linewidth=linewidth,\n markersize=markersize,label='policy gradient')\nplt.fill_between(tot_samples, np.amin(J_finite_pg,axis=0), np.amax(J_finite_pg,axis=0), alpha=0.25)\n\nplt.plot(tot_samples,np.median(J_finite_ur,axis=0),'>-',color=colors[1],linewidth=linewidth,\n markersize=markersize,label='uniform sampling')\nplt.fill_between(tot_samples, np.amin(J_finite_ur,axis=0), np.amax(J_finite_ur,axis=0), alpha=0.25)\n\nplt.plot(tot_samples,np.median(J_finite_rs,axis=0),'s-',color=colors[2],linewidth=linewidth,\n markersize=markersize,label='random search')\nplt.fill_between(tot_samples, np.amin(J_finite_rs,axis=0), np.amax(J_finite_rs,axis=0), alpha=0.25)\n\n\nplt.plot([tot_samples[0],tot_samples[-1]],[baseline, baseline],color='#000000',linewidth=linewidth,\n linestyle='--',label='zero control')\nplt.plot([tot_samples[0],tot_samples[-1]],[J_finite_opt, J_finite_opt],color='#000000',linewidth=linewidth,\n linestyle=':',label='optimal')\n\nplt.axis([0,2000,0,12])\n\nplt.xlabel('rollouts',fontsize=label_fontsize)\nplt.ylabel('cost',fontsize=label_fontsize)\nplt.legend(fontsize=18, bbox_to_anchor=(1.0, 0.54))\nplt.xticks(fontsize=tick_fontsize)\nplt.yticks(fontsize=tick_fontsize)\nplt.grid(True)\n\nfig = plt.gcf()\nfig.set_size_inches(9, 6)\n\nplt.show()", "_____no_output_____" ], [ "plt.plot(tot_samples,np.median(J_inf_pg,axis=0),'o-',color=colors[0],linewidth=linewidth,\n markersize=markersize,label='policy gradient')\nplt.fill_between(tot_samples, np.amin(J_inf_pg,axis=0), np.minimum(np.amax(J_inf_pg,axis=0),15), alpha=0.25)\n\nplt.plot(tot_samples,np.median(J_inf_ur,axis=0),'>-',color=colors[1],linewidth=linewidth,\n markersize=markersize,label='uniform sampling')\nplt.fill_between(tot_samples, np.amin(J_inf_ur,axis=0), np.minimum(np.amax(J_inf_ur,axis=0),15), alpha=0.25)\n\nplt.plot(tot_samples,np.median(J_inf_rs,axis=0),'s-',color=colors[2],linewidth=linewidth,\n markersize=markersize,label='random search')\nplt.fill_between(tot_samples, np.amin(J_inf_rs,axis=0), np.minimum(np.amax(J_inf_rs,axis=0),15), alpha=0.25)\n\nplt.plot([tot_samples[0],tot_samples[-1]],[J_inf_opt, J_inf_opt],color='#000000',linewidth=linewidth,\n linestyle=':',label='optimal')\n\nplt.axis([0,3000,5,10])\n\nplt.xlabel('rollouts',fontsize=label_fontsize)\nplt.ylabel('cost',fontsize=label_fontsize)\nplt.legend(fontsize=18, bbox_to_anchor=(1.0, 0.54))\nplt.xticks(fontsize=tick_fontsize)\nplt.yticks(fontsize=tick_fontsize)\nplt.grid(True)\n\nfig = plt.gcf()\nfig.set_size_inches(9, 6)\n\nplt.show()", "_____no_output_____" ], [ "plt.plot(tot_samples,1-np.sum(np.isinf(J_inf_pg),axis=0)/10,'o-',color=colors[0],linewidth=linewidth,\n markersize=markersize,label='policy gradient')\n\nplt.plot(tot_samples,1-np.sum(np.isinf(J_inf_ur),axis=0)/10,'>-',color=colors[1],linewidth=linewidth,\n markersize=markersize,label='uniform sampling')\n\nplt.plot(tot_samples,1-np.sum(np.isinf(J_inf_rs),axis=0)/10,'s-',color=colors[2],linewidth=linewidth,\n markersize=markersize,label='random search')\n\nplt.axis([0,3000,0,1])\n\nplt.xlabel('rollouts',fontsize=label_fontsize)\nplt.ylabel('fraction stable',fontsize=label_fontsize)\nplt.legend(fontsize=18, bbox_to_anchor=(1.0, 0.54))\nplt.xticks(fontsize=tick_fontsize)\nplt.yticks(fontsize=tick_fontsize)\nplt.grid(True)\n\nfig = plt.gcf()\nfig.set_size_inches(9, 6)\n\nplt.show()", "_____no_output_____" ], [ "one_rollout_cost-J_finite_opt", "_____no_output_____" ], [ "one_rollout_cost_inf-J_inf_opt", "_____no_output_____" ] ] ]

[ "code", "markdown", "code" ]

[ [ "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad890839595d4546deb48ed439541856cf2c0b8

111,585

ipynb

Jupyter Notebook

Learning from disaster.ipynb

daftintrovert/Learning-from-disaster

4180df58c329ced4193a1caac4d1358545833606

[ "Apache-2.0" ]

null

Learning from disaster.ipynb

daftintrovert/Learning-from-disaster

4180df58c329ced4193a1caac4d1358545833606

[ "Apache-2.0" ]

null

Learning from disaster.ipynb

daftintrovert/Learning-from-disaster

4180df58c329ced4193a1caac4d1358545833606

[ "Apache-2.0" ]

null

58.636364

31,148

0.701456

[ [ [ "import pandas as pd\nimport numpy as np\nimport matplotlib.pyplot as plt\n%matplotlib inline", "_____no_output_____" ], [ "import seaborn as sns\nsns.set_style('darkgrid')", "_____no_output_____" ], [ "import warnings\nwarnings.filterwarnings('ignore')", "_____no_output_____" ], [ "tit = sns.load_dataset('titanic')\ntit.head()", "_____no_output_____" ], [ "tit.info()", "<class 'pandas.core.frame.DataFrame'>\nRangeIndex: 891 entries, 0 to 890\nData columns (total 15 columns):\nsurvived 891 non-null int64\npclass 891 non-null int64\nsex 891 non-null object\nage 714 non-null float64\nsibsp 891 non-null int64\nparch 891 non-null int64\nfare 891 non-null float64\nembarked 889 non-null object\nclass 891 non-null category\nwho 891 non-null object\nadult_male 891 non-null bool\ndeck 203 non-null category\nembark_town 889 non-null object\nalive 891 non-null object\nalone 891 non-null bool\ndtypes: bool(2), category(2), float64(2), int64(4), object(5)\nmemory usage: 80.6+ KB\n" ], [ "tit.shape", "_____no_output_____" ], [ "tit.isnull().head()", "_____no_output_____" ], [ "plt.figure(figsize = (7,7))\nsns.heatmap(tit.isnull(),yticklabels = False,cbar = False,cmap = 'viridis')", "_____no_output_____" ], [ "tit.drop(['sibsp','parch','embarked','who','deck','alive','alone'],axis = 1,inplace = True)", "_____no_output_____" ], [ "tit.head()", "_____no_output_____" ], [ "tit.dropna(inplace = True)", "_____no_output_____" ], [ "tit.fillna(tit.mean()).head()", "_____no_output_____" ], [ "tit.info()", "<class 'pandas.core.frame.DataFrame'>\nInt64Index: 712 entries, 0 to 890\nData columns (total 8 columns):\nsurvived 712 non-null int64\npclass 712 non-null int64\nsex 712 non-null object\nage 712 non-null float64\nfare 712 non-null float64\nclass 712 non-null category\nadult_male 712 non-null bool\nembark_town 712 non-null object\ndtypes: bool(1), category(1), float64(2), int64(2), object(2)\nmemory usage: 40.4+ KB\n" ], [ "plt.figure(figsize = (7,7))\nsns.heatmap(tit.isnull(), yticklabels = False,cbar = False,cmap = 'viridis') #our data is clean now", "_____no_output_____" ], [ "tit.dropna(inplace = True)", "_____no_output_____" ], [ "sex = pd.get_dummies(tit['sex'],drop_first = True)", "_____no_output_____" ], [ "embark = pd.get_dummies(tit['embark_town'],drop_first = True)", "_____no_output_____" ], [ "embark.head()", "_____no_output_____" ], [ "sex.head()", "_____no_output_____" ], [ "tit = pd.concat([tit,sex,embark],axis = 1)", "_____no_output_____" ], [ "tit.head()", "_____no_output_____" ], [ "clas = pd.get_dummies(tit['class'],drop_first=True)", "_____no_output_____" ], [ "clas.head(2)", "_____no_output_____" ], [ "tit.drop(['sex','embark_town','class','adult_male'],axis = 1,inplace = True)", "_____no_output_____" ], [ "tit.head()", "_____no_output_____" ], [ "tit = pd.concat([tit,clas],axis = 1)", "_____no_output_____" ], [ "tit.head()", "_____no_output_____" ], [ "X = tit.drop(['survived'],axis = 1)\ny = tit['survived']", "_____no_output_____" ], [ "X.shape", "_____no_output_____" ], [ "y.shape", "_____no_output_____" ], [ "from sklearn.model_selection import train_test_split", "_____no_output_____" ], [ "X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.3, random_state=101)", "_____no_output_____" ], [ "from sklearn.linear_model import LogisticRegression", "_____no_output_____" ], [ "logmodel = LogisticRegression()", "_____no_output_____" ], [ "logmodel.fit(X_train,y_train)", "_____no_output_____" ], [ "predictions = logmodel.predict(X_test)", "_____no_output_____" ], [ "from sklearn.metrics import classification_report", "_____no_output_____" ], [ "print(classification_report(y_test,predictions))", " precision recall f1-score support\n\n 0 0.81 0.82 0.81 128\n 1 0.73 0.71 0.72 86\n\n micro avg 0.78 0.78 0.78 214\n macro avg 0.77 0.76 0.77 214\nweighted avg 0.77 0.78 0.78 214\n\n" ], [ "from sklearn.metrics import confusion_matrix", "_____no_output_____" ], [ "confusion_matrix(y_test,predictions)", "_____no_output_____" ], [ "from sklearn.linear_model import LinearRegression\n", "_____no_output_____" ], [ "lm = LinearRegression()\nlm", "_____no_output_____" ], [ "lm.fit(X_train,y_train)", "_____no_output_____" ], [ "lm.intercept_", "_____no_output_____" ], [ "lm.coef_", "_____no_output_____" ], [ "predictions = lm.predict(X_test)", "_____no_output_____" ], [ "predictions.shape", "_____no_output_____" ], [ "y_test.head()", "_____no_output_____" ], [ "plt.scatter(y_test,predictions)", "_____no_output_____" ], [ "plt.figure(figsize=(10,10))\nplt.title('Prediction Plot')\nsns.distplot(y_test-predictions,rug = True,color = 'r',bins = 10,label = 'Prediction plot')", "_____no_output_____" ], [ "from sklearn import metrics", "_____no_output_____" ], [ "metrics.mean_absolute_error(y_test,predictions)", "_____no_output_____" ], [ "metrics.mean_squared_error(y_test,predictions)", "_____no_output_____" ], [ "np.sqrt(metrics.mean_squared_error(y_test,predictions))", "_____no_output_____" ], [ "import numpy as np\nimport pandas as pd\nprediction = pd.DataFrame(predictions, columns=['predictions']).to_csv('prediction.csv')", "_____no_output_____" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad890f69f01688179274b9913a09e0724feb9c1

29,619

ipynb

Jupyter Notebook

demoFAIR_EBAII_n2.ipynb

IFB-ElixirFr/FAIR_EBAII_n2

2a550b04abc47a7a0d75a67322b7c46fac358497

[ "BSD-3-Clause" ]

null

demoFAIR_EBAII_n2.ipynb

IFB-ElixirFr/FAIR_EBAII_n2

2a550b04abc47a7a0d75a67322b7c46fac358497

[ "BSD-3-Clause" ]

null

demoFAIR_EBAII_n2.ipynb

IFB-ElixirFr/FAIR_EBAII_n2

2a550b04abc47a7a0d75a67322b7c46fac358497

[ "BSD-3-Clause" ]

null

40.966805

318

0.567203

[ [ [ "empty" ] ] ]

[ "empty" ]

[ [ "empty" ] ]

4ad8b7c46d023961639523755863d2172c329739

31,544

ipynb

Jupyter Notebook

section_uncertainty/exponential.ipynb

kentaroy47/LNPR_BOOK_CODES

f0d1bef336423ebdf04539ce833f0ce4cffc51f5

[ "MIT" ]

148

2019-03-27T00:20:16.000Z

2022-03-30T22:34:11.000Z

section_uncertainty/exponential.ipynb

kentaroy47/LNPR_BOOK_CODES

f0d1bef336423ebdf04539ce833f0ce4cffc51f5

[ "MIT" ]

3

2018-11-07T04:33:13.000Z

2018-12-31T01:35:16.000Z

section_uncertainty/exponential.ipynb

kentaroy47/LNPR_BOOK_CODES

f0d1bef336423ebdf04539ce833f0ce4cffc51f5

[ "MIT" ]

116

2019-04-18T08:35:53.000Z

2022-03-24T05:17:46.000Z

294.803738

16,052

0.941479

[ [ [ "import matplotlib.pyplot as plt\nimport numpy as np\nfrom scipy.stats import expon", "_____no_output_____" ], [ "expected_travel_distance = 5.0 #小石を踏むまで進む量の期待値\np = expon(scale=expected_travel_distance)", "_____no_output_____" ], [ "zs = np.arange(0, 10, 0.01)\nys = [p.pdf(z) for z in zs]\n\nplt.plot(zs,ys)\nplt.title(\"probaility density function\")\nplt.show()", "_____no_output_____" ], [ "ys = [p.cdf(z) for z in zs]\n\nplt.plot(zs,ys)\nplt.title(\"cumulative distribution function\")\nplt.show()", "_____no_output_____" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code" ] ]

4ad8b7d27ca1b2e9eafb8b74bd5aff96d96ca6ad

69,012

ipynb

Jupyter Notebook

dev/01_core.ipynb

DineshChauhan/fastai_docs

cf4d88073fb6f3ef7331b5360618b8dd95eb9345

[ "Apache-2.0" ]

null

dev/01_core.ipynb

DineshChauhan/fastai_docs

cf4d88073fb6f3ef7331b5360618b8dd95eb9345

[ "Apache-2.0" ]

null

dev/01_core.ipynb

DineshChauhan/fastai_docs

cf4d88073fb6f3ef7331b5360618b8dd95eb9345

[ "Apache-2.0" ]

null

28.98446

3,296

0.554541

[ [ [ "#default_exp core", "_____no_output_____" ], [ "#export\nfrom local.test import *\nfrom local.imports import *\nfrom local.notebook.showdoc import show_doc", "_____no_output_____" ] ], [ [ "# Core\n\n> Basic functions used in the fastai library", "_____no_output_____" ] ], [ [ "# export\ndefaults = SimpleNamespace()", "_____no_output_____" ] ], [ [ "## Metaclasses", "_____no_output_____" ] ], [ [ "#export\nclass PrePostInitMeta(type):\n \"A metaclass that calls optional `__pre_init__` and `__post_init__` methods\"\n def __new__(cls, name, bases, dct):\n x = super().__new__(cls, name, bases, dct)\n def _pass(self, *args,**kwargs): pass\n for o in ('__init__', '__pre_init__', '__post_init__'):\n if not hasattr(x,o): setattr(x,o,_pass)\n old_init = x.__init__\n \n @functools.wraps(old_init)\n def _init(self,*args,**kwargs):\n self.__pre_init__()\n old_init(self, *args,**kwargs)\n self.__post_init__()\n setattr(x, '__init__', _init)\n return x", "_____no_output_____" ], [ "show_doc(PrePostInitMeta, title_level=3)", "_____no_output_____" ], [ "class _T(metaclass=PrePostInitMeta):\n def __pre_init__(self): self.a = 0; assert self.a==0\n def __init__(self): self.a += 1; assert self.a==1\n def __post_init__(self): self.a += 1; assert self.a==2\n\nt = _T()\nt.a", "_____no_output_____" ], [ "#export\nclass PrePostInit(metaclass=PrePostInitMeta):\n \"Base class that provides `PrePostInitMeta` metaclass to subclasses\"\n pass", "_____no_output_____" ], [ "class _T(PrePostInit):\n def __pre_init__(self): self.a = 0; assert self.a==0\n def __init__(self): self.a += 1; assert self.a==1\n def __post_init__(self): self.a += 1; assert self.a==2\n\nt = _T()\nt.a", "_____no_output_____" ], [ "#export\nclass NewChkMeta(PrePostInitMeta):\n \"Metaclass to avoid recreating object passed to constructor (plus all `PrePostInitMeta` functionality)\"\n def __new__(cls, name, bases, dct):\n x = super().__new__(cls, name, bases, dct)\n old_init,old_new = x.__init__,x.__new__\n\n @functools.wraps(old_init)\n def _new(cls, x=None, *args, **kwargs):\n if x is not None and isinstance(x,cls):\n x._newchk = 1\n return x\n res = old_new(cls)\n res._newchk = 0\n return res\n\n @functools.wraps(old_init)\n def _init(self,*args,**kwargs):\n if self._newchk: return\n old_init(self, *args, **kwargs)\n\n x.__init__,x.__new__ = _init,_new\n return x", "_____no_output_____" ], [ "class _T(metaclass=NewChkMeta):\n \"Testing\"\n def __init__(self, o=None): self.foo = getattr(o,'foo',0) + 1\n\nclass _T2():\n def __init__(self, o): self.foo = getattr(o,'foo',0) + 1\n\nt = _T(1)\ntest_eq(t.foo,1)\nt2 = _T(t)\ntest_eq(t2.foo,1)\ntest_is(t,t2)\n\nt = _T2(1)\ntest_eq(t.foo,1)\nt2 = _T2(t)\ntest_eq(t2.foo,2)\n\ntest_eq(_T.__doc__, \"Testing\")\ntest_eq(str(inspect.signature(_T)), '(o=None)')", "_____no_output_____" ] ], [ [ "## Foundational functions", "_____no_output_____" ], [ "### Decorators", "_____no_output_____" ] ], [ [ "import copy", "_____no_output_____" ], [ "#export\ndef patch_to(cls):\n \"Decorator: add `f` to `cls`\"\n def _inner(f):\n nf = copy.copy(f)\n # `functools.update_wrapper` when passing patched function to `Pipeline`, so we do it manually\n for o in functools.WRAPPER_ASSIGNMENTS: setattr(nf, o, getattr(f,o))\n nf.__qualname__ = f\"{cls.__name__}.{f.__name__}\"\n setattr(cls, f.__name__, nf)\n return f\n return _inner", "_____no_output_____" ], [ "@patch_to(_T2)\ndef func1(x, a:bool): return a+2\n\nt = _T2(1)\ntest_eq(t.func1(1), 3)", "_____no_output_____" ], [ "#export\ndef patch(f):\n \"Decorator: add `f` to the first parameter's class (based on f's type annotations)\"\n cls = next(iter(f.__annotations__.values()))\n return patch_to(cls)(f)", "_____no_output_____" ], [ "@patch\ndef func(x:_T2, a:bool):\n \"test\"\n return a+2\n\nt = _T2(1)\ntest_eq(t.func(1), 3)\ntest_eq(t.func.__qualname__, '_T2.func')", "_____no_output_____" ] ], [ [ "### Type checking", "_____no_output_____" ], [ "Runtime type checking is handy, so let's make it easy!", "_____no_output_____" ] ], [ [ "#export core\n#NB: Please don't move this to a different line or module, since it's used in testing `get_source_link`\ndef chk(f): return typechecked(always=True)(f)", "_____no_output_____" ] ], [ [ "Decorator for a function to check that type-annotated arguments receive arguments of the right type.", "_____no_output_____" ] ], [ [ "@chk\ndef test_chk(a:int=1): return a\n\ntest_eq(test_chk(2), 2)\ntest_eq(test_chk(), 1)\ntest_fail(lambda: test_chk('a'), contains='\"a\" must be int')", "_____no_output_____" ] ], [ [ "Decorated functions will pickle correctly.", "_____no_output_____" ] ], [ [ "t = pickle.loads(pickle.dumps(test_chk))\ntest_eq(t(2), 2)\ntest_eq(t(), 1)", "_____no_output_____" ] ], [ [ "### Context managers", "_____no_output_____" ] ], [ [ "@contextmanager\ndef working_directory(path):\n \"Change working directory to `path` and return to previous on exit.\"\n prev_cwd = Path.cwd()\n os.chdir(path)\n try: yield\n finally: os.chdir(prev_cwd)", "_____no_output_____" ] ], [ [ "### Monkey-patching", "_____no_output_____" ] ], [ [ "#export\n#NB: Please don't move this to a different line or module, since it's used in testing `get_source_link`\n@patch\ndef ls(self:Path):\n \"Contents of path as a list\"\n return list(self.iterdir())", "_____no_output_____" ] ], [ [ "We add an `ls()` method to `pathlib.Path` which is simply defined as `list(Path.iterdir())`, mainly for convenience in REPL environments such as notebooks.", "_____no_output_____" ] ], [ [ "path = Path()\nt = path.ls()\nassert len(t)>0\nt[0]", "_____no_output_____" ], [ "#hide\npkl = pickle.dumps(path)\np2 =pickle.loads(pkl)\ntest_eq(path.ls()[0], p2.ls()[0])", "_____no_output_____" ], [ "#export\ndef tensor(x, *rest):\n \"Like `torch.as_tensor`, but handle lists too, and can pass multiple vector elements directly.\"\n if len(rest): x = (x,)+rest\n # Pytorch bug in dataloader using num_workers>0\n if isinstance(x, (tuple,list)) and len(x)==0: return tensor(0)\n res = torch.tensor(x) if isinstance(x, (tuple,list)) else as_tensor(x)\n if res.dtype is torch.int32:\n warn('Tensor is int32: upgrading to int64; for better performance use int64 input')\n return res.long()\n return res", "_____no_output_____" ], [ "test_eq(tensor(array([1,2,3])), torch.tensor([1,2,3]))\ntest_eq(tensor(1,2,3), torch.tensor([1,2,3]))", "_____no_output_____" ] ], [ [ "#### `Tensor.ndim`", "_____no_output_____" ] ], [ [ "#export\nTensor.ndim = property(lambda x: x.dim())", "_____no_output_____" ] ], [ [ "We add an `ndim` property to `Tensor` with same semantics as [numpy ndim](https://docs.scipy.org/doc/numpy/reference/generated/numpy.ndarray.ndim.html), which allows tensors to be used in matplotlib and other places that assume this property exists.", "_____no_output_____" ] ], [ [ "test_eq(torch.tensor([1,2]).ndim,1)\ntest_eq(torch.tensor(1).ndim,0)\ntest_eq(torch.tensor([[1]]).ndim,2)", "_____no_output_____" ] ], [ [ "### Documentation functions", "_____no_output_____" ] ], [ [ "#export core\ndef add_docs(cls, cls_doc=None, **docs):\n \"Copy values from `docs` to `cls` docstrings, and confirm all public methods are documented\"\n if cls_doc is not None: cls.__doc__ = cls_doc\n for k,v in docs.items():\n f = getattr(cls,k)\n if hasattr(f,'__func__'): f = f.__func__ # required for class methods\n f.__doc__ = v\n # List of public callables without docstring\n nodoc = [c for n,c in vars(cls).items() if isinstance(c,Callable)\n and not n.startswith('_') and c.__doc__ is None]\n assert not nodoc, f\"Missing docs: {nodoc}\"\n assert cls.__doc__ is not None, f\"Missing class docs: {cls}\"", "_____no_output_____" ], [ "#export core\ndef docs(cls):\n \"Decorator version of `add_docs\"\n add_docs(cls, **cls._docs)\n return cls", "_____no_output_____" ], [ "class _T:\n def f(self): pass\n @classmethod\n def g(cls): pass\nadd_docs(_T, \"a\", f=\"f\", g=\"g\")\n\ntest_eq(_T.__doc__, \"a\")\ntest_eq(_T.f.__doc__, \"f\")\ntest_eq(_T.g.__doc__, \"g\")", "_____no_output_____" ], [ "#export\ndef custom_dir(c, add:List):\n \"Implement custom `__dir__`, adding `add` to `cls`\"\n return dir(type(c)) + list(c.__dict__.keys()) + add", "_____no_output_____" ], [ "# export\ndef is_iter(o):\n \"Test whether `o` can be used in a `for` loop\"\n #Rank 0 tensors in PyTorch are not really iterable\n return isinstance(o, (Iterable,Generator)) and getattr(o,'ndim',1)", "_____no_output_____" ], [ "assert is_iter([1])\nassert not is_iter(torch.tensor(1))\nassert is_iter(torch.tensor([1,2]))\nassert (o for o in range(3))", "_____no_output_____" ], [ "# export\ndef coll_repr(c, max=1000):\n \"String repr of up to `max` items of (possibly lazy) collection `c`\"\n return f'(#{len(c)}) [' + ','.join(itertools.islice(map(str,c), 10)) + ('...'\n if len(c)>10 else '') + ']'", "_____no_output_____" ], [ "test_eq(coll_repr(range(1000)), '(#1000) [0,1,2,3,4,5,6,7,8,9...]')", "_____no_output_____" ] ], [ [ "## GetAttr -", "_____no_output_____" ] ], [ [ "#export\nclass GetAttr:\n \"Inherit from this to have all attr accesses in `self._xtra` passed down to `self.default`\"\n _xtra=[]\n def __getattr__(self,k):\n assert self._xtra, \"Inherited from `GetAttr` but no `_xtra` attrs listed\"\n if k in self._xtra: return getattr(self.default, k)\n raise AttributeError(k)\n def __dir__(self): return custom_dir(self, self._xtra)", "_____no_output_____" ], [ "class _C(GetAttr): default,_xtra = 'Hi',['lower']\n\nt = _C()\ntest_eq(t.lower(), 'hi')\ntest_fail(lambda: t.upper())\nassert 'lower' in dir(t)", "_____no_output_____" ] ], [ [ "## L -", "_____no_output_____" ] ], [ [ "# export\ndef _mask2idxs(mask):\n mask = list(mask)\n if len(mask)==0: return []\n if isinstance(mask[0],bool): return [i for i,m in enumerate(mask) if m]\n return [int(i) for i in mask]\n\ndef _listify(o):\n if o is None: return []\n if isinstance(o, list): return o\n if isinstance(o, (str,np.ndarray,Tensor)): return [o]\n if is_iter(o): return list(o)\n return [o]", "_____no_output_____" ], [ "#export\nclass L(GetAttr, metaclass=NewChkMeta):\n \"Behaves like a list of `items` but can also index with list of indices or masks\"\n _xtra = [o for o in dir(list) if not o.startswith('_')]\n\n def __init__(self, items=None, *rest, use_list=False, match=None):\n items = [] if items is None else items\n self.items = self.default = list(items) if use_list else _listify(items)\n self.items += list(rest)\n if match is not None:\n if len(self.items)==1: self.items = self.items*len(match)\n else: assert len(self.items)==len(match), 'Match length mismatch'\n \n def __len__(self): return len(self.items)\n def __delitem__(self, i): del(self.items[i])\n def __repr__(self): return f'{coll_repr(self)}'\n def __eq__(self,b): return all_equal(b,self)\n def __iter__(self): return (self[i] for i in range(len(self)))\n def __invert__(self): return L(not i for i in self)\n def __mul__ (a,b): return L(a.items*b)\n def __add__ (a,b): return L(a.items+_listify(b))\n def __radd__(a,b): return L(b)+a\n def __addi__(a,b):\n a.items += list(b)\n return a\n \n def __getitem__(self, idx):\n \"Retrieve `idx` (can be list of indices, or mask, or int) items\"\n res = [self.items[i] for i in _mask2idxs(idx)] if is_iter(idx) else self.items[idx]\n if isinstance(res,(tuple,list)) and not isinstance(res,L): res = L(res)\n return res\n \n def __setitem__(self, idx, o):\n \"Set `idx` (can be list of indices, or mask, or int) items to `o` (which is broadcast if not iterable)\"\n idx = idx if isinstance(idx,L) else _listify(idx) \n if not is_iter(o): o = [o]*len(idx)\n for i,o_ in zip(idx,o): self.items[i] = o_\n\n def sorted(self, key=None, reverse=False):\n \"New `L` sorted by `key`. If key is str then use `attrgetter`. If key is int then use `itemgetter`.\"\n if isinstance(key,str): k=lambda o:getattr(o,key,0)\n elif isinstance(key,int): k=itemgetter(key)\n else: k=key\n return L(sorted(self.items, key=k, reverse=reverse))\n \n def mapped(self, f, *args, **kwargs): return L(map(partial(f,*args,**kwargs), self))\n def zipped(self): return L(zip(*self))\n def itemgot(self, idx): return self.mapped(itemgetter(idx))\n def attrgot(self, k): return self.mapped(lambda o:getattr(o,k,0))\n def tensored(self): return self.mapped(tensor)\n def stack(self, dim=0): return torch.stack(list(self.tensored()), dim=dim)\n def cat (self, dim=0): return torch.cat (list(self.tensored()), dim=dim)", "_____no_output_____" ], [ "add_docs(L,\n mapped=\"Create new `L` with `f` applied to all `items`, passing `args` and `kwargs` to `f`\",\n zipped=\"Create new `L` with `zip(*items)`\",\n itemgot=\"Create new `L` with item `idx` of all `items`\",\n attrgot=\"Create new `L` with attr `k` of all `items`\",\n tensored=\"`mapped(tensor)`\",\n stack=\"Same as `torch.stack`\",\n cat=\"Same as `torch.cat`\")", "_____no_output_____" ] ], [ [ "You can create an `L` from an existing iterable (e.g. a list, range, etc) and access or modify it with an int list/tuple index, mask, int, or slice. All `list` methods can also be used with `L`.", "_____no_output_____" ] ], [ [ "t = L(range(12))\ntest_eq(t, list(range(12)))\ntest_ne(t, list(range(11)))\nt.reverse()\ntest_eq(t[0], 11)\nt[3] = \"h\"\ntest_eq(t[3], \"h\")\nt[3,5] = (\"j\",\"k\")\ntest_eq(t[3,5], [\"j\",\"k\"])\ntest_eq(t, L(t))\nt", "_____no_output_____" ] ], [ [ "You can also modify an `L` with `append`, `+`, and `*`.", "_____no_output_____" ] ], [ [ "t = L()\ntest_eq(t, [])\nt.append(1)\ntest_eq(t, [1])\nt += [3,2]\ntest_eq(t, [1,3,2])\nt = t + [4]\ntest_eq(t, [1,3,2,4])\nt = 5 + t\ntest_eq(t, [5,1,3,2,4])\ntest_eq(L(1,2,3), [1,2,3])\ntest_eq(L(1,2,3), L(1,2,3))\nt = L(1)*5\nt = t.mapped(operator.neg)\ntest_eq(t,[-1]*5)\ntest_eq(~L([True,False,False]), L([False,True,True]))", "_____no_output_____" ], [ "def _f(x,a=0): return x+a\nt = L(1)*5\ntest_eq(t.mapped(_f), t)\ntest_eq(t.mapped(_f,1), [2]*5)\ntest_eq(t.mapped(_f,a=2), [3]*5)", "_____no_output_____" ] ], [ [ "An `L` can be constructed from anything iterable, although tensors and arrays will not be iterated over on construction, unless you pass `use_list` to the constructor.", "_____no_output_____" ] ], [ [ "test_eq(L([1,2,3]),[1,2,3])\ntest_eq(L(L([1,2,3])),[1,2,3])\ntest_ne(L([1,2,3]),[1,2,])\ntest_eq(L('abc'),['abc'])\ntest_eq(L(range(0,3)),[0,1,2])\ntest_eq(L(o for o in range(0,3)),[0,1,2])\ntest_eq(L(tensor(0)),[tensor(0)])\ntest_eq(L([tensor(0),tensor(1)]),[tensor(0),tensor(1)])\ntest_eq(L(tensor([0.,1.1]))[0],tensor([0.,1.1]))\ntest_eq(L(tensor([0.,1.1]), use_list=True), [0.,1.1]) # `use_list=True` to unwrap arrays/tensors", "_____no_output_____" ] ], [ [ "If `match` is not `None` then the created list is same len as `match`, either by:\n\n- If `len(items)==1` then `items` is replicated,\n- Otherwise an error is raised if `match` and `items` are not already the same size.", "_____no_output_____" ] ], [ [ "test_eq(L(1,match=[1,2,3]),[1,1,1])\ntest_eq(L([1,2],match=[2,3]),[1,2])\ntest_fail(lambda: L([1,2],match=[1,2,3]))", "_____no_output_____" ] ], [ [ "If you create an `L` from an existing `L` then you'll get back the original object (since `L` uses the `NewChkMeta` metaclass).", "_____no_output_____" ] ], [ [ "test_is(L(t), t)", "_____no_output_____" ] ], [ [ "### Methods", "_____no_output_____" ] ], [ [ "show_doc(L.__getitem__)", "_____no_output_____" ], [ "t = L(range(12))\ntest_eq(t[1,2], [1,2]) # implicit tuple\ntest_eq(t[[1,2]], [1,2]) # list\ntest_eq(t[:3], [0,1,2]) # slice\ntest_eq(t[[False]*11 + [True]], [11]) # mask\ntest_eq(t[tensor(3)], 3)", "_____no_output_____" ], [ "show_doc(L.__setitem__)", "_____no_output_____" ], [ "t[4,6] = 0\ntest_eq(t[4,6], [0,0])\nt[4,6] = [1,2]\ntest_eq(t[4,6], [1,2])", "_____no_output_____" ], [ "show_doc(L.mapped)", "_____no_output_____" ], [ "test_eq(L(range(4)).mapped(operator.neg), [0,-1,-2,-3])", "_____no_output_____" ], [ "show_doc(L.zipped)", "_____no_output_____" ], [ "t = L([[1,2,3],'abc'])\ntest_eq(t.zipped(), [(1, 'a'),(2, 'b'),(3, 'c')])", "_____no_output_____" ], [ "show_doc(L.itemgot)", "_____no_output_____" ], [ "test_eq(t.itemgot(1), [2,'b'])", "_____no_output_____" ], [ "show_doc(L.attrgot)", "_____no_output_____" ], [ "a = [SimpleNamespace(a=3,b=4),SimpleNamespace(a=1,b=2)]\ntest_eq(L(a).attrgot('b'), [4,2])", "_____no_output_____" ], [ "show_doc(L.sorted)", "_____no_output_____" ], [ "test_eq(L(a).sorted('a').attrgot('b'), [2,4])", "_____no_output_____" ] ], [ [ "### Tensor methods", "_____no_output_____" ], [ "There are shortcuts for `torch.stack` and `torch.cat` if your `L` contains tensors or something convertible. You can manually convert with `tensored`.", "_____no_output_____" ] ], [ [ "t = L(([1,2],[3,4]))", "_____no_output_____" ], [ "test_eq(t.tensored(), [tensor(1,2),tensor(3,4)])\ntest_eq(t.stack(), tensor([[1,2],[3,4]]))\ntest_eq(t.cat(), tensor([1,2,3,4]))", "_____no_output_____" ] ], [ [ "## Utility functions", "_____no_output_____" ], [ "### Basics", "_____no_output_____" ] ], [ [ "# export\ndef ifnone(a, b):\n \"`b` if `a` is None else `a`\"\n return b if a is None else a", "_____no_output_____" ] ], [ [ "Since `b if a is None else a` is such a common pattern, we wrap it in a function. However, be careful, because python will evaluate *both* `a` and `b` when calling `ifnone` (which it doesn't do if using the `if` version directly).", "_____no_output_____" ] ], [ [ "test_eq(ifnone(None,1), 1)\ntest_eq(ifnone(2 ,1), 2)", "_____no_output_____" ], [ "#export\ndef get_class(nm, *fld_names, sup=None, doc=None, funcs=None, **flds):\n \"Dynamically create a class containing `fld_names`\"\n for f in fld_names: flds[f] = None\n for f in L(funcs): flds[f.__name__] = f\n sup = ifnone(sup, ())\n if not isinstance(sup, tuple): sup=(sup,)\n \n def _init(self, *args, **kwargs):\n for i,v in enumerate(args): setattr(self, fld_names[i], v)\n for k,v in kwargs.items(): setattr(self,k,v)\n \n def _repr(self):\n return '\\n'.join(f'{o}: {getattr(self,o)}' for o in set(dir(self))\n if not o.startswith('_') and not isinstance(getattr(self,o), types.MethodType))\n \n if not sup: flds['__repr__'] = _repr\n flds['__init__'] = _init\n res = type(nm, sup, flds)\n if doc is not None: res.__doc__ = doc\n return res", "_____no_output_____" ], [ "_t = get_class('_t', 'a')\nt = _t()\ntest_eq(t.a, None)", "_____no_output_____" ] ], [ [ "Most often you'll want to call `mk_class`, since it adds the class to your module. See `mk_class` for more details and examples of use (which also apply to `get_class`).", "_____no_output_____" ] ], [ [ "#export\ndef mk_class(nm, *fld_names, sup=None, doc=None, funcs=None, mod=None, **flds):\n \"Create a class using `get_class` and add to the caller's module\"\n if mod is None: mod = inspect.currentframe().f_back.f_locals\n res = get_class(nm, *fld_names, sup=sup, doc=doc, funcs=funcs, **flds)\n mod[nm] = res", "_____no_output_____" ], [ "sys.modules[ifnone.__module__]", "_____no_output_____" ] ], [ [ "Any `kwargs` will be added as class attributes, and `sup` is an optional (tuple of) base classes.", "_____no_output_____" ] ], [ [ "mk_class('_t', a=1, sup=GetAttr)\nt = _t()\ntest_eq(t.a, 1)\nassert(isinstance(t,GetAttr))", "_____no_output_____" ] ], [ [ "A `__init__` is provided that sets attrs for any `kwargs`, and for any `args` (matching by position to fields), along with a `__repr__` which prints all attrs. The docstring is set to `doc`. You can pass `funcs` which will be added as attrs with the function names.", "_____no_output_____" ] ], [ [ "def foo(self): return 1\nmk_class('_t', 'a', sup=GetAttr, doc='test doc', funcs=foo)\n\nt = _t(3, b=2)\ntest_eq(t.a, 3)\ntest_eq(t.b, 2)\ntest_eq(t.foo(), 1)\ntest_eq(t.__doc__, 'test doc')\nt", "_____no_output_____" ], [ "#export\ndef wrap_class(nm, *fld_names, sup=None, doc=None, funcs=None, **flds):\n \"Decorator: makes function a method of a new class `nm` passing parameters to `mk_class`\"\n def _inner(f):\n mk_class(nm, *fld_names, sup=sup, doc=doc, funcs=L(funcs)+f, mod=f.__globals__, **flds)\n return f\n return _inner", "_____no_output_____" ], [ "@wrap_class('_t', a=2)\ndef bar(self,x): return x+1\n\nt = _t()\ntest_eq(t.a, 2)\ntest_eq(t.bar(3), 4)\nt", "_____no_output_____" ], [ "# export\ndef noop (x=None, *args, **kwargs):\n \"Do nothing\"\n return x", "_____no_output_____" ], [ "noop()\ntest_eq(noop(1),1)", "_____no_output_____" ], [ "# export\ndef noops(self, x, *args, **kwargs):\n \"Do nothing (method)\"\n return x", "_____no_output_____" ], [ "mk_class('_t', foo=noops)\ntest_eq(_t().foo(1),1)", "_____no_output_____" ] ], [ [ "### Collection functions", "_____no_output_____" ] ], [ [ "#export\ndef tuplify(o, use_list=False, match=None):\n \"Make `o` a tuple\"\n return tuple(L(o, use_list=use_list, match=match))", "_____no_output_____" ], [ "test_eq(tuplify(None),())\ntest_eq(tuplify([1,2,3]),(1,2,3))\ntest_eq(tuplify(1,match=[1,2,3]),(1,1,1))", "_____no_output_____" ], [ "#export\ndef replicate(item,match):\n \"Create tuple of `item` copied `len(match)` times\"\n return (item,)*len(match)", "_____no_output_____" ], [ "t = [1,1]\ntest_eq(replicate([1,2], t),([1,2],[1,2]))\ntest_eq(replicate(1, t),(1,1))", "_____no_output_____" ], [ "#export\ndef uniqueify(x, sort=False, bidir=False, start=None):\n \"Return the unique elements in `x`, optionally `sort`-ed, optionally return the reverse correspondance.\"\n res = list(OrderedDict.fromkeys(x).keys())\n if start is not None: res = L(start)+res\n if sort: res.sort()\n if bidir: return res, {v:k for k,v in enumerate(res)}\n return res", "_____no_output_____" ], [ "# test\ntest_eq(set(uniqueify([1,1,0,5,0,3])),{0,1,3,5})\ntest_eq(uniqueify([1,1,0,5,0,3], sort=True),[0,1,3,5])\nv,o = uniqueify([1,1,0,5,0,3], bidir=True)\ntest_eq(v,[1,0,5,3])\ntest_eq(o,{1:0, 0: 1, 5: 2, 3: 3})\nv,o = uniqueify([1,1,0,5,0,3], sort=True, bidir=True)\ntest_eq(v,[0,1,3,5])\ntest_eq(o,{0:0, 1: 1, 3: 2, 5: 3})", "_____no_output_____" ], [ "# export\ndef setify(o): return o if isinstance(o,set) else set(L(o))", "_____no_output_____" ], [ "# test\ntest_eq(setify(None),set())\ntest_eq(setify('abc'),{'abc'})\ntest_eq(setify([1,2,2]),{1,2})\ntest_eq(setify(range(0,3)),{0,1,2})\ntest_eq(setify({1,2}),{1,2})", "_____no_output_____" ], [ "#export\ndef is_listy(x):\n \"`isinstance(x, (tuple,list,L))`\"\n return isinstance(x, (tuple,list,L,slice))", "_____no_output_____" ], [ "assert is_listy([1])\nassert is_listy(L([1]))\nassert is_listy(slice(2))\nassert not is_listy(torch.tensor([1]))", "_____no_output_____" ], [ "#export\ndef range_of(x):\n \"All indices of collection `x` (i.e. `list(range(len(x)))`)\"\n return list(range(len(x)))", "_____no_output_____" ], [ "test_eq(range_of([1,1,1,1]), [0,1,2,3])", "_____no_output_____" ], [ "# export\ndef mask2idxs(mask):\n \"Convert bool mask or index list to index `L`\"\n return L(_mask2idxs(mask))", "_____no_output_____" ], [ "test_eq(mask2idxs([False,True,False,True]), [1,3])\ntest_eq(mask2idxs(torch.tensor([1,2,3])), [1,2,3])", "_____no_output_____" ] ], [ [ "### File and network functions", "_____no_output_____" ] ], [ [ "def bunzip(fn):\n \"bunzip `fn`, raising exception if output already exists\"\n fn = Path(fn)\n assert fn.exists(), f\"{fn} doesn't exist\"\n out_fn = fn.with_suffix('')\n assert not out_fn.exists(), f\"{out_fn} already exists\"\n with bz2.BZ2File(fn, 'rb') as src, out_fn.open('wb') as dst:\n for d in iter(lambda: src.read(1024*1024), b''): dst.write(d)", "_____no_output_____" ], [ "f = Path('files/test.txt')\nif f.exists(): f.unlink()\nbunzip('files/test.txt.bz2')\nt = f.open().readlines()\ntest_eq(len(t),1)\ntest_eq(t[0], 'test\\n')\nf.unlink()", "_____no_output_____" ] ], [ [ "### Tensor functions", "_____no_output_____" ] ], [ [ "#export\ndef apply(func, x, *args, **kwargs):\n \"Apply `func` recursively to `x`, passing on args\"\n if is_listy(x): return [apply(func, o, *args, **kwargs) for o in x]\n if isinstance(x,dict): return {k: apply(func, v, *args, **kwargs) for k,v in x.items()}\n return func(x, *args, **kwargs)", "_____no_output_____" ], [ "#export\ndef to_detach(b, cpu=True):\n \"Recursively detach lists of tensors in `b `; put them on the CPU if `cpu=True`.\"\n def _inner(x, cpu=True):\n if not isinstance(x,Tensor): return x\n x = x.detach()\n return x.cpu() if cpu else x\n return apply(_inner, b, cpu=cpu)", "_____no_output_____" ], [ "#export\ndef to_half(b):\n \"Recursively map lists of tensors in `b ` to FP16.\"\n return apply(lambda x: x.half() if x.dtype not in [torch.int64, torch.int32, torch.int16] else x, b)", "_____no_output_____" ], [ "#export\ndef to_float(b):\n \"Recursively map lists of int tensors in `b ` to float.\"\n return apply(lambda x: x.float() if x.dtype not in [torch.int64, torch.int32, torch.int16] else x, b)", "_____no_output_____" ], [ "#export\ndefaults.device = torch.cuda.current_device() if torch.cuda.is_available() else torch.device('cpu')", "_____no_output_____" ], [ "#export\ndef to_device(b, device=defaults.device):\n \"Recursively put `b` on `device`.\"\n def _inner(o): return o.to(device, non_blocking=True) if isinstance(o,Tensor) else o\n return apply(_inner, b)", "_____no_output_____" ], [ "t1,(t2,t3) = to_device([3,[tensor(3),tensor(2)]])\ntest_eq((t1,t2,t3),(3,3,2))\ntest_eq(t2.type(), \"torch.cuda.LongTensor\")\ntest_eq(t3.type(), \"torch.cuda.LongTensor\")", "_____no_output_____" ], [ "#export\ndef to_cpu(b):\n \"Recursively map lists of tensors in `b ` to the cpu.\"\n return to_device(b,'cpu')", "_____no_output_____" ], [ "t3 = to_cpu(t3)\ntest_eq(t3.type(), \"torch.LongTensor\")\ntest_eq(t3, 2)", "_____no_output_____" ], [ "def to_np(x):\n \"Convert a tensor to a numpy array.\"\n return x.data.cpu().numpy()", "_____no_output_____" ], [ "t3 = to_np(t3)\ntest_eq(type(t3), np.ndarray)\ntest_eq(t3, 2)", "_____no_output_____" ], [ "#export\ndef item_find(x, idx=0):\n \"Recursively takes the `idx`-th element of `x`\"\n if is_listy(x): return item_find(x[idx])\n if isinstance(x,dict): \n key = list(x.keys())[idx] if isinstance(idx, int) else idx\n return item_find(x[key])\n return x", "_____no_output_____" ], [ "#export\ndef find_device(b):\n \"Recursively search the device of `b`.\"\n return item_find(b).device", "_____no_output_____" ], [ "test_eq(find_device(t2).index, defaults.device)\ntest_eq(find_device([t2,t2]).index, defaults.device)\ntest_eq(find_device({'a':t2,'b':t2}).index, defaults.device)\ntest_eq(find_device({'a':[[t2],[t2]],'b':t2}).index, defaults.device)", "_____no_output_____" ], [ "#export\ndef find_bs(b):\n \"Recursively search the batch size of `b`.\"\n return item_find(b).shape[0]", "_____no_output_____" ], [ "x = torch.randn(4,5)\ntest_eq(find_bs(x), 4)\ntest_eq(find_bs([x, x]), 4)\ntest_eq(find_bs({'a':x,'b':x}), 4)\ntest_eq(find_bs({'a':[[x],[x]],'b':x}), 4)", "_____no_output_____" ], [ "def np_func(f):\n \"Convert a function taking and returning numpy arrays to one taking and returning tensors\"\n def _inner(*args, **kwargs):\n nargs = [to_np(arg) if isinstance(arg,Tensor) else arg for arg in args]\n return tensor(f(*nargs, **kwargs))\n functools.update_wrapper(_inner, f)\n return _inner", "_____no_output_____" ] ], [ [ "This decorator is particularly useful for using numpy functions as fastai metrics, for instance:", "_____no_output_____" ] ], [ [ "from sklearn.metrics import f1_score\n\n@np_func\ndef f1(inp,targ): return f1_score(targ, inp)\n\na1,a2 = array([0,1,1]),array([1,0,1])\nt = f1(tensor(a1),tensor(a2))\ntest_eq(f1_score(a1,a2), t)\nassert isinstance(t,Tensor)", "_____no_output_____" ], [ "class Module(nn.Module, metaclass=PrePostInitMeta):\n \"Same as `nn.Module`, but no need for subclasses to call `super().__init__`\"\n def __pre_init__(self): super().__init__()\n def __init__(self): pass", "_____no_output_____" ], [ "show_doc(Module, title_level=3)", "_____no_output_____" ], [ "class _T(Module):\n def __init__(self): self.f = nn.Linear(1,1)\n def forward(self,x): return self.f(x)\n\nt = _T()\nt(tensor([1.]))", "_____no_output_____" ] ], [ [ "### Functions on functions", "_____no_output_____" ] ], [ [ "# export\n@chk\ndef compose(*funcs: Callable, order=None):\n \"Create a function that composes all functions in `funcs`, passing along remaining `*args` and `**kwargs` to all\"\n funcs = L(funcs)\n if order is not None: funcs = funcs.sorted(order)\n def _inner(x, *args, **kwargs):\n for f in L(funcs): x = f(x, *args, **kwargs)\n return x\n return _inner", "_____no_output_____" ], [ "f1 = lambda o,p=0: (o*2)+p\nf2 = lambda o,p=1: (o+1)/p\ntest_eq(f2(f1(3)), compose(f1,f2)(3))\ntest_eq(f2(f1(3,p=3),p=3), compose(f1,f2)(3,p=3))\ntest_eq(f2(f1(3, 3), 3), compose(f1,f2)(3, 3))", "_____no_output_____" ], [ "f1.order = 1\ntest_eq(f1(f2(3)), compose(f1,f2, order=\"order\")(3))", "_____no_output_____" ], [ "#export\ndef mapper(f):\n \"Create a function that maps `f` over an input collection\"\n return lambda o: [f(o_) for o_ in o]", "_____no_output_____" ], [ "func = mapper(lambda o:o*2)\ntest_eq(func(range(3)),[0,2,4])", "_____no_output_____" ], [ "#export\ndef partialler(f, *args, order=None, **kwargs):\n \"Like `functools.partial` but also copies over docstring\"\n fnew = partial(f,*args,**kwargs)\n fnew.__doc__ = f.__doc__\n if order is not None: fnew.order=order\n elif hasattr(f,'order'): fnew.order=f.order\n return fnew", "_____no_output_____" ], [ "def _f(x,a=1):\n \"test func\"\n return x+a\n_f.order=1\n\nf = partialler(_f, a=2)\ntest_eq(f.order, 1)\nf = partialler(_f, a=2, order=3)\ntest_eq(f.__doc__, \"test func\")\ntest_eq(f.order, 3)\ntest_eq(f(3), _f(3,2))", "_____no_output_____" ] ], [ [ "### Sorting objects from before/after", "_____no_output_____" ], [ "Transforms and callbacks will have run_after/run_before attributes, this function will sort them to respect those requirements (if it's possible). Also, sometimes we want a tranform/callback to be run at the end, but still be able to use run_after/run_before behaviors. For those, the function checks for a toward_end attribute (that needs to be True).", "_____no_output_____" ] ], [ [ "#export\ndef _is_instance(f, gs):\n tst = [g if type(g) in [type, 'function'] else g.__class__ for g in gs]\n for g in tst:\n if isinstance(f, g) or f==g: return True\n return False\n\ndef _is_first(f, gs):\n for o in L(getattr(f, 'run_after', None)): \n if _is_instance(o, gs): return False\n for g in gs:\n if _is_instance(f, L(getattr(g, 'run_before', None))): return False\n return True\n\ndef sort_by_run(fs):\n end = L(getattr(f, 'toward_end', False) for f in fs)\n inp,res = L(fs)[~end] + L(fs)[end], []\n while len(inp) > 0:\n for i,o in enumerate(inp):\n if _is_first(o, inp): \n res.append(inp.pop(i))\n break\n else: raise Exception(\"Impossible to sort\")\n return res", "_____no_output_____" ], [ "class Tst(): pass \nclass Tst1():\n run_before=[Tst]\nclass Tst2():\n run_before=Tst\n run_after=Tst1\n \ntsts = [Tst(), Tst1(), Tst2()]\ntest_eq(sort_by_run(tsts), [tsts[1], tsts[2], tsts[0]])\n\nTst2.run_before,Tst2.run_after = Tst1,Tst\ntest_fail(lambda: sort_by_run([Tst(), Tst1(), Tst2()]))\n\ndef tst1(x): return x\ntst1.run_before = Tst\ntest_eq(sort_by_run([tsts[0], tst1]), [tst1, tsts[0]])\n \nclass Tst1():\n toward_end=True\nclass Tst2():\n toward_end=True\n run_before=Tst1\ntsts = [Tst(), Tst1(), Tst2()]\ntest_eq(sort_by_run(tsts), [tsts[0], tsts[2], tsts[1]])", "_____no_output_____" ] ], [ [ "### Other helpers", "_____no_output_____" ] ], [ [ "#export\ndef num_cpus():\n \"Get number of cpus\"\n try: return len(os.sched_getaffinity(0))\n except AttributeError: return os.cpu_count()\n \ndefaults.cpus = min(16, num_cpus())", "_____no_output_____" ], [ "#export\ndef add_props(f, n=2):\n \"Create properties passing each of `range(n)` to f\"\n return (property(partial(f,i)) for i in range(n))", "_____no_output_____" ], [ "class _T(): a,b = add_props(lambda i,x:i*2)\n\nt = _T()\ntest_eq(t.a,0)\ntest_eq(t.b,2)", "_____no_output_____" ] ], [ [ "This is a quick way to generate, for instance, *train* and *valid* versions of a property. See `DataBunch` definition for an example of this.", "_____no_output_____" ] ], [ [ "#export\ndef make_cross_image(bw=True):\n \"Create a tensor containing a cross image, either `bw` (True) or color\"\n if bw:\n im = torch.zeros(5,5)\n im[2,:] = 1.\n im[:,2] = 1.\n else:\n im = torch.zeros(3,5,5)\n im[0,2,:] = 1.\n im[1,:,2] = 1.\n return im", "_____no_output_____" ], [ "plt.imshow(make_cross_image(), cmap=\"Greys\");", "_____no_output_____" ], [ "plt.imshow(make_cross_image(False).permute(1,2,0));", "_____no_output_____" ] ], [ [ "## Export -", "_____no_output_____" ] ], [ [ "#hide\nfrom local.notebook.export import notebook2script\nnotebook2script(all_fs=True)", "Converted 00_test.ipynb.\nConverted 01_core.ipynb.\nConverted 02_data_pipeline.ipynb.\nConverted 02_data_pipeline_v2-multi.ipynb.\nConverted 02_data_pipeline_v2.ipynb.\nConverted 03_data_external.ipynb.\nConverted 04_data_core.ipynb.\nConverted 05_data_source.ipynb.\nConverted 06_vision_core.ipynb.\nConverted 07_pets_tutorial-meta.ipynb.\nConverted 07_pets_tutorial-oo.ipynb.\nConverted 07_pets_tutorial-oo1.ipynb.\nConverted 07_pets_tutorial-oo2-meta.ipynb.\nConverted 07_pets_tutorial.ipynb.\nConverted 08_augmentation.ipynb.\nConverted 10_layers.ipynb.\nConverted 11_optimizer.ipynb.\nConverted 12_learner.ipynb.\nConverted 13_callback_schedule.ipynb.\nConverted 14_callback_hook.ipynb.\nConverted 15_callback_progress.ipynb.\nConverted 16_callback_tracker.ipynb.\nConverted 17_callback_fp16.ipynb.\nConverted 90_notebook_core.ipynb.\nConverted 91_notebook_export.ipynb.\nConverted 92_notebook_showdoc.ipynb.\nConverted 93_notebook_export2html.ipynb.\nConverted 94_index.ipynb.\nConverted 95_synth_learner.ipynb.\nConverted tmp_tensor_inherit.ipynb.\n" ] ] ]

[ "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ] ]

4ad8b9b5c034e9614cf6b00db98b1f9fde49427d

236,447

ipynb

Jupyter Notebook

_notebooks/2020-01-28-python_for_ML.ipynb

itskarthicklakshmanan/kalanai

fe397c596745b39a7c5d4005bd11bb6a5fae94cf

[ "Apache-2.0" ]

null

_notebooks/2020-01-28-python_for_ML.ipynb

itskarthicklakshmanan/kalanai

fe397c596745b39a7c5d4005bd11bb6a5fae94cf

[ "Apache-2.0" ]

3

2021-03-30T11:17:09.000Z

2022-02-26T08:46:24.000Z

_notebooks/2020-01-28-python_for_ML.ipynb

itskarthicklakshmanan/kalanai

fe397c596745b39a7c5d4005bd11bb6a5fae94cf

[ "Apache-2.0" ]

null

24.107565

598

0.376355

[ [ [ "# Python for ML\n> Basic Python reference useful for ML\n\n- toc: true \n- badges: true\n- comments: true\n- categories: [Python, NumPy, Pandas]\n- image: images/py.png\n", "_____no_output_____" ], [ "----------------------------------------------------------------------------------------------------------------------------", "_____no_output_____" ], [ "## Python Collections", "_____no_output_____" ], [ "Collection Types:\n\n1) List is a collection which is ordered and changeable. Allows duplicate members\n\n2) Tuple is a collection which is ordered and unchangeable. Allows duplicate members\n\n3) Set is a collection which is unordered and unindexed. No duplicate members\n\n4) Dictionary is a collection which is unordered, changeable and indexed. No duplicate members.", "_____no_output_____" ], [ "### 1) List", "_____no_output_____" ] ], [ [ "list = [\"apple\", \"grapes\", \"banana\"]\nprint(list)", "['apple', 'grapes', 'banana']\n" ], [ "print(list[1]) #access the list items by referring to the index number", "grapes\n" ], [ "print(list[-1]) #Negative indexing means beginning from the end, -1 refers to the last item", "banana\n" ], [ "list2 = [\"apple\", \"banana\", \"cherry\", \"orange\", \"kiwi\", \"melon\", \"mango\"] \nprint(list2[:4]) #By leaving out the start value, the range will start at the first item", "['apple', 'banana', 'cherry', 'orange']\n" ], [ "print(list2[2:])", "['cherry', 'orange', 'kiwi', 'melon', 'mango']\n" ], [ "print(list2[-4:-1]) #range", "['orange', 'kiwi', 'melon']\n" ], [ "list3 = [\"A\", \"B\", \"C\"]\nlist3[1] = \"D\" #change the value of a specific item, by refering to the index number\nprint(list3)", "['A', 'D', 'C']\n" ], [ "# For loop\n\nlist4 = [\"apple\", \"banana\", \"cherry\"]\nfor x in list4:\n print(x)", "apple\nbanana\ncherry\n" ], [ "#To determine if a specified item is present in a list\n\nif \"apple\" in list4:\n print(\"Yes\")", "Yes\n" ], [ "#To determine how many items a list has\n\nprint(len(list4))", "3\n" ] ], [ [ "-------------------------------------------------------------------------------------------------------------------------", "_____no_output_____" ], [ "#### List Methods:\n\n- append()\t: Adds an element at the end of the list\n- clear()\t: Removes all the elements from the list\n- copy()\t: Returns a copy of the list\n- count()\t: Returns the number of elements with the specified value\n- extend()\t: Add the elements of a list (or any iterable), to the end of the current list\n- index()\t: Returns the index of the first element with the specified value\n- insert()\t: Adds an element at the specified position\n- pop()\t : Removes the element at the specified position\n- remove()\t: Removes the item with the specified value\n- reverse()\t: Reverses the order of the list\n- sort()\t: Sorts the list", "_____no_output_____" ] ], [ [ "#append() method to append an item\n\nlist4.append(\"orange\")\nprint(list4)", "['apple', 'banana', 'cherry', 'orange']\n" ], [ "#Insert an item as the second position\n\nlist4.insert(1, \"orange\")\nprint(list4)", "['apple', 'orange', 'banana', 'cherry', 'orange']\n" ], [ "#The remove() method removes the specified item\n\nlist4.remove(\"banana\")\nprint(list4)", "['apple', 'orange', 'cherry', 'orange']\n" ], [ "#pop() method removes the specified index\n#and the last item if index is not specified\n\nlist4.pop()\nprint(list4)", "['apple', 'orange', 'cherry']\n" ], [ "#The del keyword removes the specified index\n\ndel list4[0]\nprint(list4)", "['orange', 'cherry']\n" ], [ "#The del keyword can also delete the list completely\ndel list4\nptint(list4)", "_____no_output_____" ], [ "#The clear() method empties the list\n\nlist5 = [\"apple\", \"banana\", \"cherry\"]\nlist5.clear()\nprint(list5)", "[]\n" ], [ "#the copy() method to make a copy of a list \nlist5 = [\"apple\", \"banana\", \"cherry\"]\nmylist = list5.copy()\nprint(mylist)", "['apple', 'banana', 'cherry']\n" ], [ "#Join Two Lists\n\nlist1 = [\"a\", \"b\" , \"c\"]\nlist2 = [1, 2, 3]\n\nlist3 = list1 + list2\nprint(list3)", "['a', 'b', 'c', 1, 2, 3]\n" ], [ "#Append list2 into list1\n\nlist1 = [\"a\", \"b\" , \"c\"]\nlist2 = [1, 2, 3]\n\nfor x in list2:\n list1.append(x)\n\nprint(list1)", "['a', 'b', 'c', 1, 2, 3]\n" ], [ "#the extend() method to add list2 at the end of list1\n\nlist1 = [\"a\", \"b\" , \"c\"]\nlist2 = [1, 2, 3]\n\nlist1.extend(list2)\nprint(list1)", "['a', 'b', 'c', 1, 2, 3]\n" ] ], [ [ "### 2) Tuple\n\nA tuple is a collection which is ordered and unchangeable.", "_____no_output_____" ] ], [ [ "tuple1 = (\"apple\", \"banana\", \"cherry\")\nprint(tuple1)", "('apple', 'banana', 'cherry')\n" ], [ "#access tuple item\n\nprint(tuple1[1])", "banana\n" ], [ "#Negative indexing means beginning from the end, -1 refers to the last item\n\nprint(tuple1[-1])", "cherry\n" ], [ "#Range : Return the third, fourth, and fifth item\n\ntuple2 = (\"apple\", \"banana\", \"cherry\", \"orange\", \"kiwi\", \"melon\", \"mango\")\nprint(tuple2[2:5])", "('cherry', 'orange', 'kiwi')\n" ], [ "#Specify negative indexes if you want to start the search from the end of the tuple\n\nprint(tuple2[-4:-1])", "('orange', 'kiwi', 'melon')\n" ], [ "#loop through the tuple items by using a for loop\n\ntuple3 = (\"apple\", \"banana\", \"cherry\")\nfor x in tuple3:\n print(x)", "apple\nbanana\ncherry\n" ], [ "#Check if Item Exists\n\nif \"apple\" in tuple3:\n print(\"Yes\")", "Yes\n" ], [ "#Print the number of items in the tuple\n\nprint(len(tuple3))", "3\n" ], [ "# join two or more tuples you can use the + operator\n\ntuple1 = (\"a\", \"b\" , \"c\")\ntuple2 = (1, 2, 3)\n\ntuple3 = tuple1 + tuple2\nprint(tuple3)", "('a', 'b', 'c', 1, 2, 3)\n" ], [ "#Using the tuple() method to make a tuple\n\nthistuple = tuple((\"apple\", \"banana\", \"cherry\")) # note the double round-brackets\nprint(thistuple)", "('apple', 'banana', 'cherry')\n" ] ], [ [ "### 3) Set\nA set is a collection which is unordered and unindexed. Sets are written with curly brackets.", "_____no_output_____" ] ], [ [ "set1 = {\"apple\", \"banana\", \"cherry\"}\nprint(set1)", "{'banana', 'apple', 'cherry'}\n" ], [ "#Access items, Loop through the set, and print the values\n\nfor x in set1:\n print(x)", "banana\napple\ncherry\n" ], [ "if \"apple\" in set1:\n print(\"Yes\")", "Yes\n" ] ], [ [ "#### Set methods:\n\n- add()\tAdds an element to the set\n- clear()\tRemoves all the elements from the set\n- copy()\tReturns a copy of the set\n- difference()\tReturns a set containing the difference between two or more sets\n- difference_update()\tRemoves the items in this set that are also included in another, specified set\n- discard()\tRemove the specified item\n- intersection()\tReturns a set, that is the intersection of two other sets\n- intersection_update()\tRemoves the items in this set that are not present in other, specified set(s)\n- isdisjoint()\tReturns whether two sets have a intersection or not\n- issubset()\tReturns whether another set contains this set or not\n- issuperset()\tReturns whether this set contains another set or not\n- pop()\tRemoves an element from the set\n- remove()\tRemoves the specified element\n- symmetric_difference()\tReturns a set with the symmetric differences of two sets\n- symmetric_difference_update()\tinserts the symmetric differences from this set and another\n- union()\tReturn a set containing the union of sets\n- update()\tUpdate the set with the union of this set and others", "_____no_output_____" ] ], [ [ "# Adding new items \n\n\nset1.add(\"orange\")\nprint(set1)", "{'banana', 'apple', 'cherry', 'orange'}\n" ], [ "#Add multiple items to a set, using the update() method\n\nset1.update([\"orange\", \"mango\", \"grapes\"])\n\nprint(set1)", "{'banana', 'cherry', 'orange', 'apple', 'grapes', 'mango'}\n" ], [ "# length of the set\n\n\nprint(len(set1))", "6\n" ], [ "# remove item \n\nset1.remove(\"banana\")\n\nprint(set1)", "{'cherry', 'orange', 'apple', 'grapes', 'mango'}\n" ], [ "#Remove the last item by using the pop() method\n\nset2 = {\"apple\", \"banana\", \"cherry\"}\n\nx = set2.pop()\n\nprint(x)\nprint(set2)", "banana\n{'apple', 'cherry'}\n" ], [ "#clear() method empties the set\n\n\nthisset = {\"apple\", \"banana\", \"cherry\"}\n\nthisset.clear()\n\nprint(thisset)", "set()\n" ], [ "#del keyword will delete the set completely\n\nthisset = {\"apple\", \"banana\", \"cherry\"}\n\ndel thisset\n\nprint(thisset)", "_____no_output_____" ], [ "#use the union() method that returns a new set containing all items from both sets, \n#or the update() method that inserts all the items from one set into another\n\nset1 = {\"a\", \"b\" , \"c\"}\nset2 = {1, 2, 3}\n\nset3 = set1.union(set2)\nprint(set3)", "{'b', 1, 2, 3, 'a', 'c'}\n" ], [ "#update() method inserts the items in set2 into set1\n\nset1 = {\"a\", \"b\" , \"c\"}\nset2 = {1, 2, 3}\n\nset1.update(set2)\nprint(set1)", "{'b', 1, 2, 3, 'a', 'c'}\n" ] ], [ [ "#### 4) Dictionary\nA dictionary is a collection which is unordered, changeable and indexed.", "_____no_output_____" ] ], [ [ "dict = {\n \"brand\": \"Ford\",\n \"model\": \"Mustang\",\n \"year\": 1964\n}\nprint(dict)", "{'brand': 'Ford', 'model': 'Mustang', 'year': 1964}\n" ], [ "#access the items of a dictionary by referring to its key name, inside square brackets\n\ndict[\"model\"]", "_____no_output_____" ] ], [ [ "#### Dict methods\n\n- clear()\tRemoves all the elements from the dictionary\n- copy()\tReturns a copy of the dictionary\n- fromkeys()\tReturns a dictionary with the specified keys and value\n- get()\tReturns the value of the specified key\n- items()\tReturns a list containing a tuple for each key value pair\n- keys()\tReturns a list containing the dictionary's keys\n- pop()\tRemoves the element with the specified key\n- popitem()\tRemoves the last inserted key-value pair\n- setdefault()\tReturns the value of the specified key. If the key does not exist: insert the key, with the specified value\n- update()\tUpdates the dictionary with the specified key-value pairs\n- values()\tReturns a list of all the values in the dictionary", "_____no_output_____" ] ], [ [ "#use get() to get the same result\n\ndict.get(\"model\")", "_____no_output_____" ], [ "#change the value of a specific item by referring to its key name\n\ndict1 = {\n \"brand\": \"Ford\",\n \"model\": \"Mustang\",\n \"year\": 1964\n}\ndict1[\"year\"] = 2018\n\nprint(dict1)", "{'brand': 'Ford', 'model': 'Mustang', 'year': 2018}\n" ], [ "#loop through a dictionary by using a for loop\n\nfor x in dict1:\n print(x)", "brand\nmodel\nyear\n" ], [ "#Print all values in the dictionary, one by one\n\nfor x in dict1:\n print(dict1[x])", "Ford\nMustang\n2018\n" ], [ "#use the values() method to return values of a dictionary\n\nfor x in dict1.values():\n print(x)", "Ford\nMustang\n2018\n" ], [ "#Loop through both keys and values, by using the items() method\n\nfor x, y in dict1.items():\n print(x, y)", "brand Ford\nmodel Mustang\nyear 2018\n" ], [ "#Check if an item present in the dictionary\n\nif \"model\" in dict1:\n print(\"Yes\")", "Yes\n" ], [ "print(len(dict1))", "3\n" ], [ "#adding items\n\nthisdict = {\n \"brand\": \"Ford\",\n \"model\": \"Mustang\",\n \"year\": 1964\n}\nthisdict[\"color\"] = \"red\"\nprint(thisdict)", "{'brand': 'Ford', 'model': 'Mustang', 'year': 1964, 'color': 'red'}\n" ], [ "#pop() method removes the item with the specified key name\n\nthisdict = {\n \"brand\": \"Ford\",\n \"model\": \"Mustang\",\n \"year\": 1964\n}\nthisdict.pop(\"model\")\nprint(thisdict)", "{'brand': 'Ford', 'year': 1964}\n" ], [ "# popitem() method removes the last inserted item\n\nthisdict = {\n \"brand\": \"Ford\",\n \"model\": \"Mustang\",\n \"year\": 1964\n}\nthisdict.popitem()\nprint(thisdict)", "{'brand': 'Ford', 'model': 'Mustang'}\n" ], [ "#del keyword removes the item with the specified key name\n\nthisdict = {\n \"brand\": \"Ford\",\n \"model\": \"Mustang\",\n \"year\": 1964\n}\ndel thisdict[\"model\"]\nprint(thisdict)", "{'brand': 'Ford', 'year': 1964}\n" ], [ "#dictionary can also contain many dictionaries, this is called nested dictionaries\n\nmyfamily = {\n \"child1\" : {\n \"name\" : \"Emil\",\n \"year\" : 2004\n },\n \"child2\" : {\n \"name\" : \"Tobias\",\n \"year\" : 2007\n },\n \"child3\" : {\n \"name\" : \"Linus\",\n \"year\" : 2011\n }\n}", "_____no_output_____" ], [ "#Create three dictionaries, then create one dictionary that will contain the other three dictionaries\n\nchild1 = {\n \"name\" : \"Emil\",\n \"year\" : 2004\n}\nchild2 = {\n \"name\" : \"Tobias\",\n \"year\" : 2007\n}\nchild3 = {\n \"name\" : \"Linus\",\n \"year\" : 2011\n}\n\nmyfamily = {\n \"child1\" : child1,\n \"child2\" : child2,\n \"child3\" : child3\n}", "_____no_output_____" ] ], [ [ "## Python Conditions", "_____no_output_____" ], [ "### If statement", "_____no_output_____" ] ], [ [ "a = 100\nb = 200\nif b > a:\n print(\"b is greater than a\")", "b is greater than a\n" ], [ "#simplyfied:\n\na = 100\nb = 200\nif a < b: print(\"a is greater than b\")", "a is greater than b\n" ], [ "a = 20\nb = 20\nif b > a:\n print(\"b is greater than a\")\nelif a == b:\n print(\"a and b are equal\")", "a and b are equal\n" ], [ "a = 200\nb = 100\nif b > a:\n print(\"b is greater than a\")\nelif a == b:\n print(\"a and b are equal\")\nelse:\n print(\"a is greater than b\")", "a is greater than b\n" ], [ "# simplyfied:\n\na = 100\nb = 300\nprint(\"A\") if a > b else print(\"B\")", "B\n" ] ], [ [ "### AND and OR Statement", "_____no_output_____" ] ], [ [ "a = 200\nb = 33\nc = 500\nif a > b and c > a:\n print(\"Both conditions are True\")", "Both conditions are True\n" ], [ "a = 200\nb = 33\nc = 500\nif a > b or a > c:\n print(\"At least one of the conditions is True\")", "At least one of the conditions is True\n" ] ], [ [ "### Nested If", "_____no_output_____" ], [ "x = 41\n\nif x > 10:\n print(\"Above ten,\")\n if x > 20:\n print(\"and also above 20!\")\n else:\n print(\"but not above 20.\")", "_____no_output_____" ], [ "### Pass", "_____no_output_____" ] ], [ [ "#if statements cannot be empty, but if you for some reason have an if statement \n#with no content, put in the pass statement to avoid getting an error\n\na = 33\nb = 200\n\nif b > a:\n pass", "_____no_output_____" ] ], [ [ "### The while Loop", "_____no_output_____" ] ], [ [ "i = 1\nwhile i < 6:\n print(i)\n i += 1", "1\n2\n3\n4\n5\n" ] ], [ [ "### Break Statement", "_____no_output_____" ] ], [ [ "i = 1\nwhile i < 6:\n print(i)\n if i == 3:\n break\n i += 1", "1\n2\n3\n" ], [ "# with Continue\n\ni = 0\nwhile i < 6:\n i += 1\n if i == 3:\n continue\n print(i)", "1\n2\n4\n5\n6\n" ], [ "### Else statement\n\ni = 1\nwhile i < 6:\n print(i)\n i += 1\nelse:\n print(\"i is no longer less than 6\")", "1\n2\n3\n4\n5\ni is no longer less than 6\n" ] ], [ [ "### For Loops", "_____no_output_____" ] ], [ [ "# For loop for List\n\nfruits = [\"apple\", \"banana\", \"cherry\"]\nfor x in fruits:\n print(x)", "apple\nbanana\ncherry\n" ], [ "# strings\n\nfor x in \"banana\":\n print(x)", "b\na\nn\na\nn\na\n" ], [ "#break statement\n\nfruits = [\"apple\", \"banana\", \"cherry\"]\nfor x in fruits:\n print(x)\n if x == \"banana\":\n break", "apple\nbanana\n" ], [ "fruits = [\"apple\", \"banana\", \"cherry\"]\nfor x in fruits:\n if x == \"banana\":\n break\n print(x)", "apple\n" ], [ "#continue\n\nfruits = [\"apple\", \"banana\", \"cherry\"]\nfor x in fruits:\n if x == \"banana\":\n continue\n print(x)", "apple\ncherry\n" ], [ "# Range\n\nfor x in range(6):\n print(x)", "0\n1\n2\n3\n4\n5\n" ], [ "for x in range(2, 6):\n print(x)", "2\n3\n4\n5\n" ], [ "for x in range(2, 30, 3):\n print(x)", "2\n5\n8\n11\n14\n17\n20\n23\n26\n29\n" ], [ "for x in range(6):\n print(x)\nelse:\n print(\"Finally finished!\")", "0\n1\n2\n3\n4\n5\nFinally finished!\n" ], [ "adj = [\"red\", \"big\", \"tasty\"]\nfruits = [\"apple\", \"banana\", \"cherry\"]\n\nfor x in adj:\n for y in fruits:\n print(x, y)", "red apple\nred banana\nred cherry\nbig apple\nbig banana\nbig cherry\ntasty apple\ntasty banana\ntasty cherry\n" ], [ "for x in [0, 1, 2]:\n pass", "_____no_output_____" ] ], [ [ "### Creating a Function", "_____no_output_____" ] ], [ [ "def my_function():\n print(\"Hello\")\n\n\nmy_function()", "Hello\n" ], [ "def my_function(*kids):\n print(\"The youngest child is \" + kids[2])\n\nmy_function(\"Emil\", \"Tobias\", \"Linus\")", "The youngest child is Linus\n" ], [ "def my_function(child3, child2, child1):\n print(\"The youngest child is \" + child3)\n\nmy_function(child1 = \"Emil\", child2 = \"Tobias\", child3 = \"Linus\")", "The youngest child is Linus\n" ], [ "#Passing a List as an Argument\n\ndef my_function(food):\n for x in food:\n print(x)\n\nfruits = [\"apple\", \"banana\", \"cherry\"]\n\nmy_function(fruits)", "apple\nbanana\ncherry\n" ], [ "#return value\n\ndef my_function(x):\n return 5 * x\n\nprint(my_function(3))\n", "15\n" ], [ "#Recursion Example\n\ndef tri_recursion(k):\n if(k > 0):\n result = k + tri_recursion(k - 1)\n print(result)\n else:\n result = 0\n return result\n\nprint(\"\\n\\nRecursion Example Results\")\ntri_recursion(6)", "\n\nRecursion Example Results\n1\n3\n6\n10\n15\n21\n" ] ], [ [ "### lambda function", "_____no_output_____" ] ], [ [ "x = lambda a, b, c : a + b + c\nprint(x(5, 6, 2))", "13\n" ], [ "def myfunc(n):\n return lambda a : a * n", "_____no_output_____" ], [ "def myfunc(n):\n return lambda a : a * n\n\nmydoubler = myfunc(2)\n\nprint(mydoubler(11))", "22\n" ], [ "def myfunc(n):\n return lambda a : a * n\n\nmydoubler = myfunc(2)\nmytripler = myfunc(3)\n\nprint(mydoubler(11))\nprint(mytripler(11))", "22\n33\n" ] ], [ [ "### Open a File on the Server", "_____no_output_____" ], [ "### Reading files", "_____no_output_____" ] ], [ [ "#f = open(\"demofile.txt\", \"r\")\n#print(f.read())\n\n#f = open(\"D:\\\\myfiles\\welcome.txt\", \"r\")\n#print(f.read())", "_____no_output_____" ], [ "#Read one line of the file\n\n#f = open(\"demofile.txt\", \"r\")\n#print(f.readline())", "_____no_output_____" ], [ "#Loop through the file line by line\n\n#f = open(\"demofile.txt\", \"r\")\n#for x in f:\n# print(x)", "_____no_output_____" ], [ "#Close the file when you are finish with it\n\n#f = open(\"demofile.txt\", \"r\")\n#print(f.readline())\n#f.close()", "_____no_output_____" ] ], [ [ "### Writing files:", "_____no_output_____" ] ], [ [ "#Open the file \"demofile2.txt\" and append content to the file\n\n#f = open(\"demofile2.txt\", \"a\")\n#f.write(\"Now the file has more content!\")\n#f.close()\n\n#open and read the file after the appending:\n#f = open(\"demofile2.txt\", \"r\")\n#print(f.read())", "_____no_output_____" ], [ "#Open the file \"demofile3.txt\" and overwrite the content\n\n#f = open(\"demofile3.txt\", \"w\")\n#f.write(\"Woops! I have deleted the content!\")\n#f.close()\n\n#open and read the file after the appending:\n#f = open(\"demofile3.txt\", \"r\")\n#print(f.read())", "_____no_output_____" ], [ "#Create a file called \"myfile.txt\"\n\n#f = open(\"myfile.txt\", \"x\")", "_____no_output_____" ], [ "#Remove the file \"demofile.txt\"\n\n#import os\n#os.remove(\"demofile.txt\")", "_____no_output_____" ], [ "#Check if file exists, then delete it:\n\n#import os\n#if os.path.exists(\"demofile.txt\"):\n# os.remove(\"demofile.txt\")\n#else:\n# print(\"The file does not exist\")", "_____no_output_____" ], [ "#Try to open and write to a file that is not writable:\n\n#try:\n# f = open(\"demofile.txt\")\n# f.write(\"Lorum Ipsum\")\n#except:\n# print(\"Something went wrong when writing to the file\")\n#finally:\n# f.close()", "_____no_output_____" ], [ "#Raise an error and stop the program if x is lower than 0:\n\n#x = -1\n\n#if x < 0:\n# raise Exception(\"Sorry, no numbers below zero\")\n", "_____no_output_____" ], [ "#Raise a TypeError if x is not an integer:\n\n#x = \"hello\"\n\n#if not type(x) is int:\n# raise TypeError(\"Only integers are allowed\")\n", "_____no_output_____" ] ], [ [ "-----------------------------------------------------------------------------------------------------------------------------", "_____no_output_____" ], [ "# NumPy", "_____no_output_____" ] ], [ [ "import numpy as np", "_____no_output_____" ], [ "simple_list = [1,2,3]", "_____no_output_____" ], [ "np.array(simple_list)", "_____no_output_____" ], [ "list_of_lists = [[1,2,3], [4,5,6], [7,8,9]]", "_____no_output_____" ], [ "np.array(list_of_lists)", "_____no_output_____" ], [ "np.arange(0,10)", "_____no_output_____" ], [ "np.arange(0,21,5)", "_____no_output_____" ], [ "np.zeros(50)", "_____no_output_____" ], [ "np.ones((4,5))", "_____no_output_____" ], [ "np.linspace(0,20,10)", "_____no_output_____" ], [ "np.eye(5)", "_____no_output_____" ], [ "np.random.rand(3,2)", "_____no_output_____" ], [ "np.random.randint(5,20,10)", "_____no_output_____" ], [ "np.arange(30)", "_____no_output_____" ], [ "np.random.randint(0,100,20)", "_____no_output_____" ], [ "sample_array = np.arange(30)\nsample_array.reshape(5,6)", "_____no_output_____" ], [ "rand_array = np.random.randint(0,100,20)\nrand_array.argmin()", "_____no_output_____" ], [ "sample_array.shape", "_____no_output_____" ], [ "sample_array.reshape(1,30)", "_____no_output_____" ], [ "sample_array.reshape(30,1)", "_____no_output_____" ], [ "sample_array.dtype", "_____no_output_____" ], [ "a = np.random.randn(2,3)\na.T", "_____no_output_____" ], [ "sample_array = np.arange(10,21)", "_____no_output_____" ], [ "sample_array", "_____no_output_____" ], [ "sample_array[[2,5]]", "_____no_output_____" ], [ "sample_array[1:2] = 100", "_____no_output_____" ], [ "sample_array", "_____no_output_____" ], [ "sample_array = np.arange(10,21)", "_____no_output_____" ], [ "sample_array[0:7]", "_____no_output_____" ], [ "sample_array = np.arange(10,21)\n \nsample_array", "_____no_output_____" ], [ "subset_sample_array = sample_array[0:7]\n\nsubset_sample_array", "_____no_output_____" ], [ "subset_sample_array[:]=1001", "_____no_output_____" ], [ "subset_sample_array", "_____no_output_____" ], [ "sample_array", "_____no_output_____" ], [ "copy_sample_array = sample_array.copy()", "_____no_output_____" ], [ "copy_sample_array", "_____no_output_____" ], [ "copy_sample_array[:]=10\ncopy_sample_array", "_____no_output_____" ], [ "sample_array", "_____no_output_____" ], [ "sample_matrix = np.array(([50,20,1,23], [24,23,21,32], [76,54,32,12], [98,6,4,3]))", "_____no_output_____" ], [ "sample_matrix", "_____no_output_____" ], [ "sample_matrix[0][3]", "_____no_output_____" ], [ "sample_matrix[0,3]", "_____no_output_____" ], [ "sample_matrix[3,:]", "_____no_output_____" ], [ "sample_matrix[3]", "_____no_output_____" ], [ "sample_matrix = np.array(([50,20,1,23,34], [24,23,21,32,34], [76,54,32,12,98], [98,6,4,3,67], [12,23,34,56,67]))", "_____no_output_____" ], [ "sample_matrix", "_____no_output_____" ], [ "sample_matrix[:,[1,3]]", "_____no_output_____" ], [ "sample_matrix[:,(3,1)]", "_____no_output_____" ], [ "sample_array=np.arange(1,31)", "_____no_output_____" ], [ "sample_array", "_____no_output_____" ], [ "bool = sample_array < 10", "_____no_output_____" ], [ "sample_array[bool]", "_____no_output_____" ], [ "sample_array[sample_array <10]", "_____no_output_____" ], [ "a=11", "_____no_output_____" ], [ "sample_array[sample_array < a]", "_____no_output_____" ], [ "sample_array + sample_array", "_____no_output_____" ], [ "sample_array / sample_array", "_____no_output_____" ], [ "10/sample_array", "_____no_output_____" ], [ "sample_array + 1", "_____no_output_____" ], [ "np.var(sample_array)", "_____no_output_____" ], [ "array = np.random.randn(6,6)", "_____no_output_____" ], [ "array", "_____no_output_____" ], [ "np.std(array)", "_____no_output_____" ], [ "np.mean(array)", "_____no_output_____" ], [ "sports = np.array(['golf', 'cric', 'fball', 'cric', 'Cric', 'fooseball'])\n\nnp.unique(sports)", "_____no_output_____" ], [ "sample_array", "_____no_output_____" ], [ "simple_array = np.arange(0,20)", "_____no_output_____" ], [ "simple_array", "_____no_output_____" ], [ "np.save('sample_array', sample_array)", "_____no_output_____" ], [ "np.savez('2_arrays.npz', a=sample_array, b=simple_array)", "_____no_output_____" ], [ "np.load('sample_array.npy')", "_____no_output_____" ], [ "archive = np.load('2_arrays.npz')", "_____no_output_____" ], [ "archive['b']", "_____no_output_____" ], [ "np.savetxt('text_file.txt', sample_array,delimiter=',')", "_____no_output_____" ], [ "np.loadtxt('text_file.txt', delimiter=',')", "_____no_output_____" ], [ "data = {'prodID': ['101', '102', '103', '104', '104'],\n\n 'prodname': ['X', 'Y', 'Z', 'X', 'W'],\n\n 'profit': ['2738', '2727', '3497', '7347', '3743']}", "_____no_output_____" ] ], [ [ "--------------------------------------------------------------------------------------------------------------------------", "_____no_output_____" ], [ "# Pandas", "_____no_output_____" ] ], [ [ "import pandas as pd", "_____no_output_____" ], [ "score = [10, 15, 20, 25]", "_____no_output_____" ], [ "pd.Series(data=score, index = ['a','b','c','d'])", "_____no_output_____" ], [ "demo_matrix = np.array(([13,35,74,48], [23,37,37,38], [73,39,93,39]))", "_____no_output_____" ], [ "demo_matrix", "_____no_output_____" ], [ "demo_matrix[2,3]", "_____no_output_____" ], [ "np.arange(0,22,6)", "_____no_output_____" ], [ "demo_array=np.arange(0,10)", "_____no_output_____" ], [ "demo_array", "_____no_output_____" ], [ "demo_array <3", "_____no_output_____" ], [ "demo_array[demo_array <6]", "_____no_output_____" ], [ "np.max(demo_array)", "_____no_output_____" ], [ "s1 = pd.Series(['a', 'b'])", "_____no_output_____" ], [ "s2 = pd.Series(['c', 'd'])", "_____no_output_____" ], [ "pd.concat([s1+s2])", "_____no_output_____" ] ], [ [ "#### Creating a Series using Pandas\nYou could convert a list,numpy array, or dictionary to a Series in the following manner", "_____no_output_____" ] ], [ [ "labels = ['w','x','y','z']\nlist = [10,20,30,40]\narray = np.array([10,20,30,40])\ndict = {'w':10,'x':20,'y':30,'z':40}", "_____no_output_____" ], [ "pd.Series(data=list)", "_____no_output_____" ], [ "pd.Series(data=list,index=labels)", "_____no_output_____" ], [ "pd.Series(list,labels)", "_____no_output_____" ], [ "pd.Series(array)", "_____no_output_____" ], [ "pd.Series(array,labels)", "_____no_output_____" ], [ "pd.Series(dict)", "_____no_output_____" ] ], [ [ "#### Using an Index\nWe shall now see how to index in a Series using the following examples of 2 series", "_____no_output_____" ] ], [ [ "sports1 = pd.Series([1,2,3,4],index = ['Cricket', 'Football','Basketball', 'Golf'])", "_____no_output_____" ], [ "sports1", "_____no_output_____" ], [ "sports2 = pd.Series([1,2,5,4],index = ['Cricket', 'Football','Baseball', 'Golf'])", "_____no_output_____" ], [ "sports2", "_____no_output_____" ], [ "sports1 + sports2", "_____no_output_____" ] ], [ [ "#### DataFrames\nDataFrames concept in python is similar to that of R programming language. DataFrame is a collection of Series combined together to share the same index positions.\n\n", "_____no_output_____" ] ], [ [ "from numpy.random import randn\nnp.random.seed(1)", "_____no_output_____" ], [ "dataframe = pd.DataFrame(randn(10,5),index='A B C D E F G H I J'.split(),columns='Score1 Score2 Score3 Score4 Score5'.split())", "_____no_output_____" ], [ "dataframe", "_____no_output_____" ] ], [ [ "#### Selection and Indexing\nWays in which we can grab data from a DataFrame", "_____no_output_____" ] ], [ [ "dataframe['Score3']", "_____no_output_____" ], [ "# Pass a list of column names in any order necessary\ndataframe[['Score2','Score1']]", "_____no_output_____" ], [ "#DataFrame Columns are nothing but a Series each\ntype(dataframe['Score1'])", "_____no_output_____" ] ], [ [ "#### Adding a new column to the DataFrame", "_____no_output_____" ] ], [ [ "dataframe['Score6'] = dataframe['Score1'] + dataframe['Score2']", "_____no_output_____" ], [ "dataframe", "_____no_output_____" ] ], [ [ "#### Removing Columns from DataFrame", "_____no_output_____" ] ], [ [ "# Use axis=0 for dropping rows and axis=1 for dropping columns\n \ndataframe.drop('Score6',axis=1) ", "_____no_output_____" ], [ "# column is not dropped unless inplace input is TRUE\ndataframe", "_____no_output_____" ], [ "dataframe.drop('Score6',axis=1,inplace=True)\ndataframe", "_____no_output_____" ] ], [ [ "#### Dropping rows using axis=0", "_____no_output_____" ] ], [ [ "# Row will also be dropped only if inplace=TRUE is given as input\n\ndataframe.drop('A',axis=0) ", "_____no_output_____" ] ], [ [ "#### Selecting Rows", "_____no_output_____" ] ], [ [ "dataframe.loc['F']", "_____no_output_____" ] ], [ [ "#### select based off of index position instead of label - use iloc instead of loc function", "_____no_output_____" ] ], [ [ "dataframe.iloc[2]", "_____no_output_____" ] ], [ [ "#### Selecting subset of rows and columns using loc function", "_____no_output_____" ] ], [ [ "dataframe.loc['A','Score1']", "_____no_output_____" ], [ "dataframe.loc[['A','B'],['Score1','Score2']]", "_____no_output_____" ] ], [ [ "#### Conditional Selection\nSimilar to NumPy, we can make conditional selections using Brackets", "_____no_output_____" ] ], [ [ "dataframe>0.5", "_____no_output_____" ], [ "dataframe[dataframe>0.5]", "_____no_output_____" ], [ "dataframe[dataframe['Score1']>0.5]", "_____no_output_____" ], [ "dataframe[dataframe['Score1']>0.5]['Score2']", "_____no_output_____" ], [ "dataframe[dataframe['Score1']>0.5][['Score2','Score3']]", "_____no_output_____" ] ], [ [ "#### Some more features of indexing includes\n\n- resetting the index\n- setting a different value\n- index hierarchy", "_____no_output_____" ] ], [ [ "# Reset to default index value instead of A to J\ndataframe.reset_index()", "_____no_output_____" ], [ "# Setting new index value\nnewindex = 'IND JP CAN GE IT PL FY IU RT IP'.split()", "_____no_output_____" ], [ "dataframe['Countries'] = newindex\ndataframe", "_____no_output_____" ], [ "dataframe.set_index('Countries')", "_____no_output_____" ], [ "# Once again, ensure that you input inplace=TRUE\ndataframe", "_____no_output_____" ], [ "dataframe.set_index('Countries',inplace=True)", "_____no_output_____" ], [ "dataframe", "_____no_output_____" ] ], [ [ "#### Missing Data\nMethods to deal with missing data in Pandas", "_____no_output_____" ] ], [ [ "dataframe = pd.DataFrame({'Cricket':[1,2,np.nan,4,6,7,2,np.nan],\n 'Baseball':[5,np.nan,np.nan,5,7,2,4,5],\n 'Tennis':[1,2,3,4,5,6,7,8]})", "_____no_output_____" ], [ "dataframe", "_____no_output_____" ], [ "dataframe.dropna()", "_____no_output_____" ], [ "# Use axis=1 for dropping columns with nan values\n\ndataframe.dropna(axis=1) ", "_____no_output_____" ], [ "dataframe.dropna(thresh=2)", "_____no_output_____" ], [ "dataframe.fillna(value=0)", "_____no_output_____" ], [ "dataframe['Baseball'].fillna(value=dataframe['Baseball'].mean())", "_____no_output_____" ] ], [ [ "#### Groupby\nThe groupby method is used to group rows together and perform aggregate functions", "_____no_output_____" ] ], [ [ "dat = {'CustID':['1001','1001','1002','1002','1003','1003'],\n 'CustName':['UIPat','DatRob','Goog','Chrysler','Ford','GM'],\n 'Profitinlakhs':[2005,3245,1245,8765,5463,3547]}", "_____no_output_____" ], [ "dataframe = pd.DataFrame(dat)", "_____no_output_____" ], [ "dataframe", "_____no_output_____" ] ], [ [ "We can now use the .groupby() method to group rows together based on a column name. \nFor example let's group based on CustID. This will create a DataFrameGroupBy object:", "_____no_output_____" ] ], [ [ "dataframe.groupby('CustID') #This object can be saved as a variable", "_____no_output_____" ], [ "CustID_grouped = dataframe.groupby(\"CustID\") #Now we can aggregate using the variable", "_____no_output_____" ], [ "CustID_grouped.mean()", "_____no_output_____" ] ], [ [ "#### groupby function for each aggregation", "_____no_output_____" ] ], [ [ "dataframe.groupby('CustID').mean()", "_____no_output_____" ], [ "CustID_grouped.std()", "_____no_output_____" ], [ "CustID_grouped.min()", "_____no_output_____" ], [ "CustID_grouped.max()", "_____no_output_____" ], [ "CustID_grouped.count()", "_____no_output_____" ], [ "CustID_grouped.describe()", "_____no_output_____" ], [ "CustID_grouped.describe().transpose()", "_____no_output_____" ], [ "CustID_grouped.describe().transpose()['1001']", "_____no_output_____" ] ], [ [ "#### combining DataFrames together:\n\n- Merging\n- Joining\n- Concatenating", "_____no_output_____" ] ], [ [ "dafa1 = pd.DataFrame({'CustID': ['101', '102', '103', '104'],\n 'Sales': [13456, 45321, 54385, 53212],\n 'Priority': ['CAT0', 'CAT1', 'CAT2', 'CAT3'],\n 'Prime': ['yes', 'no', 'no', 'yes']},\n index=[0, 1, 2, 3])\n\ndafa2 = pd.DataFrame({'CustID': ['101', '103', '104', '105'],\n 'Sales': [13456, 54385, 53212, 4534],\n 'Payback': ['CAT4', 'CAT5', 'CAT6', 'CAT7'],\n 'Imp': ['yes', 'no', 'no', 'no']},\n index=[4, 5, 6, 7]) \n\ndafa3 = pd.DataFrame({'CustID': ['101', '104', '105', '106'],\n 'Sales': [13456, 53212, 4534, 3241],\n 'Pol': ['CAT8', 'CAT9', 'CAT10', 'CAT11'],\n 'Level': ['yes', 'no', 'no', 'yes']},\n index=[8, 9, 10, 11])", "_____no_output_____" ] ], [ [ "#### Concatenation\nConcatenation joins DataFrames basically either by rows or colums(axis=0 or 1).\n\nWe also need to ensure dimension sizes of dataframes are the same", "_____no_output_____" ] ], [ [ "pd.concat([dafa1,dafa2])", "D:\\Anaconda3\\lib\\site-packages\\ipykernel_launcher.py:1: FutureWarning: Sorting because non-concatenation axis is not aligned. A future version\nof pandas will change to not sort by default.\n\nTo accept the future behavior, pass 'sort=False'.\n\nTo retain the current behavior and silence the warning, pass 'sort=True'.\n\n \"\"\"Entry point for launching an IPython kernel.\n" ], [ "pd.concat([dafa1,dafa2,dafa3],axis=1)", "_____no_output_____" ] ], [ [ "#### Merging\nJust like SQL tables, merge function in python allows us to merge dataframes", "_____no_output_____" ] ], [ [ "pd.merge(dafa1,dafa2,how='outer',on='CustID')", "_____no_output_____" ] ], [ [ "#### Operations\nLet us discuss some useful Operations using Pandas", "_____no_output_____" ] ], [ [ "dataframe = pd.DataFrame({'custID':[1,2,3,4],'SaleType':['big','small','medium','big'],'SalesCode':['121','131','141','151']})\ndataframe.head()", "_____no_output_____" ] ], [ [ "Info on Unique Values", "_____no_output_____" ] ], [ [ "dataframe['SaleType'].unique()", "_____no_output_____" ], [ "dataframe['SaleType'].nunique()", "_____no_output_____" ], [ "dataframe['SaleType'].value_counts()", "_____no_output_____" ] ], [ [ "Selecting Data\n", "_____no_output_____" ] ], [ [ "#Select from DataFrame using criteria from multiple columns\nnewdataframe = dataframe[(dataframe['custID']!=3) & (dataframe['SaleType']=='big')]\nnewdataframe", "_____no_output_____" ] ], [ [ "Applying Functions", "_____no_output_____" ] ], [ [ "def profit(a):\n return a*4", "_____no_output_____" ], [ "dataframe['custID'].apply(profit)", "_____no_output_____" ], [ "dataframe['SaleType'].apply(len)", "_____no_output_____" ], [ "dataframe['custID'].sum()", "_____no_output_____" ] ], [ [ "#### Permanently Removing a Column", "_____no_output_____" ] ], [ [ "dataframe", "_____no_output_____" ], [ "del dataframe['custID']\ndataframe", "_____no_output_____" ] ], [ [ "#### Get column and index names", "_____no_output_____" ] ], [ [ "dataframe.columns", "_____no_output_____" ], [ "dataframe.index", "_____no_output_____" ] ], [ [ "#### Sorting and Ordering a DataFrame", "_____no_output_____" ] ], [ [ "dataframe.sort_values(by='SaleType') #inplace=False by default", "_____no_output_____" ] ], [ [ "#### Find Null Values or Check for Null Values", "_____no_output_____" ] ], [ [ "dataframe.isnull()", "_____no_output_____" ], [ "# Drop rows with NaN Values\ndataframe.dropna()", "_____no_output_____" ] ], [ [ "#### Filling in NaN values with something else", "_____no_output_____" ] ], [ [ "dataframe = pd.DataFrame({'Sale1':[5,np.nan,10,np.nan],\n 'Sale2':[np.nan,121,np.nan,141],\n 'Sale3':['XUI','VYU','NMA','IUY']})\ndataframe.head()", "_____no_output_____" ], [ "dataframe.fillna('Not nan')", "_____no_output_____" ] ], [ [ "### Data Input and Output\nReading DataFrames from external sources using pd.read functions", "_____no_output_____" ], [ "CSV Input", "_____no_output_____" ] ], [ [ "# dataframe = pd.read_csv('filename.csv')", "_____no_output_____" ] ], [ [ "CSV output", "_____no_output_____" ] ], [ [ "#If index=FALSE then csv does not store index values\n\n# dataframe.to_csv('filename.csv',index=False) ", "_____no_output_____" ] ], [ [ "Excel Input", "_____no_output_____" ] ], [ [ "# pd.read_excel('filename.xlsx',sheet_name='Data1')", "_____no_output_____" ] ], [ [ "Excel Output", "_____no_output_____" ] ], [ [ "# dataframe.to_excel('Consumer2.xlsx',sheet_name='Sheet1')", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown", "markdown", "markdown", "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad8d77d3cbf3b76ae667a7947feaf9d830cd968

3,778

ipynb

Jupyter Notebook

downloaded_kernels/loan_data/kernel_169.ipynb

josepablocam/common-code-extraction

a6978fae73eee8ece6f1db09f2f38cf92f03b3ad

[ "MIT" ]

null

downloaded_kernels/loan_data/kernel_169.ipynb

josepablocam/common-code-extraction

a6978fae73eee8ece6f1db09f2f38cf92f03b3ad

[ "MIT" ]

null

downloaded_kernels/loan_data/kernel_169.ipynb

josepablocam/common-code-extraction

a6978fae73eee8ece6f1db09f2f38cf92f03b3ad

[ "MIT" ]

2

2021-07-12T00:48:08.000Z

2021-08-11T12:53:05.000Z

3,778

0.676284

[ [ [ "I tried to make this generic, so you can cut and paste whatever query you want, **after replacing the input files with the appropriate data source**. For this example, I'm just working my way down the list of most popular SQLite kernels, and the next in line is the one identified below.", "_____no_output_____" ], [ "**Source of query:** \n[Total loans by state barplot](https://www.kaggle.com/bruessow/total-loans-by-state-histogram) by SvenBrüssow", "_____no_output_____" ] ], [ [ "dbname = 'database.sqlite'", "_____no_output_____" ], [ "import numpy as np \nimport pandas as pd\nimport sqlite3\n\nfrom subprocess import check_output\nprint(check_output([\"ls\", \"../input\"]).decode(\"utf8\"))", "_____no_output_____" ], [ "path = \"../input/\" #Insert path here\ndatabase = path + dbname\n\nconn = sqlite3.connect(database)\n\ntables = pd.read_sql(\"\"\"SELECT *\n FROM sqlite_master\n WHERE type='table';\"\"\", conn)\ntables", "_____no_output_____" ] ], [ [ "### PASTE QUERY IN THE BLOCK BELOW", "_____no_output_____" ] ], [ [ "query = '''\nSELECT addr_state,\nREPLACE(SUBSTR(QUOTE(\n ZEROBLOB((COUNT(*)/1000+1)/2)\n), 3, COUNT(*)/1000), '0', '*')\nAS total_loans\nFROM loan\nGROUP BY addr_state\nORDER BY COUNT(*) DESC;\n'''", "_____no_output_____" ], [ "result = pd.read_sql( query, conn )", "_____no_output_____" ], [ "result", "_____no_output_____" ], [ "result.to_csv(\"result.csv\", index=False)", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code" ] ]

4ad8e821940dacd7b9922aa10bf586af7d261845

5,875

ipynb

Jupyter Notebook

preprocessing/preprocess_tweetsum.ipynb

sarahaman/CIS6930_TweetSum_Summarization

c98117f9beffb7e65f7d4046df307d9fb2966065

[ "MIT" ]

2

2022-01-20T20:17:24.000Z

2022-02-08T02:39:37.000Z

preprocessing/preprocess_tweetsum.ipynb

sarahaman/CIS6930_TweetSum_Summarization

c98117f9beffb7e65f7d4046df307d9fb2966065

[ "MIT" ]

null

preprocessing/preprocess_tweetsum.ipynb

sarahaman/CIS6930_TweetSum_Summarization

c98117f9beffb7e65f7d4046df307d9fb2966065

[ "MIT" ]

null

5,875

0.602043

[ [ [ "# ------------------------- # \n# SET - UP # \n# ------------------------- # \n\n# ---- Requirements ----- # \n\n#!pip install datasets\n#!pip install sentencepiece\n#!pip install transformers\n#!pip install jsonlines\n\nimport csv\nimport datasets\nfrom google.colab import drive\nimport huggingface_hub\nimport jsonlines\nimport json\nimport pandas as pd\nimport re\nimport sys\n\n# ----- Check if GPU is connected ----- # \ngpu_info = !nvidia-smi -L\ngpu_info = \"\\n\".join(gpu_info)\nif gpu_info.find(\"failed\") >= 0:\n print(\"Not connected to a GPU\")\nelse:\n print(gpu_info)\n\n# ----- Mounting Google Drive ----- # \n\ndrive.mount('/content/drive')\nsys.path.append('/content/drive/MyDrive/CIS6930_final')\n\n# ----- Importing TweetSum processing module ----- #\nfrom tweet_sum_processor import TweetSumProcessor\n\n# ----------------------------------------------------------------------\n\n# ------------------------- # \n# PRE-PROCESSING FUNCTIONS # \n# ------------------------- # \n\n\ndef get_inputs(json_format):\n '''\n ---------------------\n Input: Dictionary containing the metadata for one tweet conversation\n Output: Concatenated string containing the content of one conversation. \n \n Notes: \n Special characters inserted for links and transitions between speaker. \n Anonymized usernames are removed, as they do not add value to the text, \n as they are usually just located at the beginning of the tweet by default \n (feature of threads). Whereas usernames containing the name of the business \n are retained for contextual purpose. \n ---------------------\n '''\n dialogue = json_format['dialog']['turns']\n full_text = []\n for i in dialogue:\n string = ' '.join(i['sentences'])\n full_text.append(string + \" \")\n conversation = ' '.join(full_text)\n by_word = conversation.split(' ')\n for i in range(0, len(by_word)):\n if \"https\" in by_word[i]:\n by_word[i] = \"<LINK>\"\n if \"@\" in by_word[i]:\n if by_word[i][1:].isnumeric():\n by_word[i] = ''\n text = ' '.join(by_word)\n text = re.sub(r'[^a-zA-Z0-9,!.?<> ]', '', text)\n text = re.sub(r'(\\W)(?=\\1)', '', text)\n return text\n \ndef get_summary(json_format): \n '''\n ---------------------\n Input: Dictionary containing the metadata for one tweet conversation\n Output: The text of a single human-generated summary for that one tweet\n ---------------------\n '''\n temp = json_format['summaries']['abstractive_summaries'][0]\n summary = ' '.join(temp)\n return summary\n\ndef prepare_data(file_name, processor):\n '''\n Processing the TweetSum dataset so that it can be read as a HuggingFace dataset. \n ---------------------\n Input: Path to a dataset file and the TweetSum processor\n Output: The inputs and summaries for the given data\n ---------------------\n '''\n inputs = []\n summaries = []\n with open('/content/drive/MyDrive/CIS6930_final/' + file_name) as f:\n dialog_with_summaries = processor.get_dialog_with_summaries(f.readlines())\n for dialog_with_summary in dialog_with_summaries:\n try:\n json_format = json.loads(dialog_with_summary.get_json())\n inputs.append(get_inputs(json_format))\n summaries.append(get_summary(json_format))\n except TypeError:\n pass\n return inputs, summaries\n\n# ----------------------------------------------------------------------\n\n# ------------------------- # \n# \"MAIN\" # \n# ------------------------- # \n\n# --- \"Fake\" main function because this is a notebook and not a script :)\n\n# ------ Process data \n\nprocessor = TweetSumProcessor('/content/drive/MyDrive/CIS6930_final/kaggle_files/twcs.csv')\ntrain_inputs, train_summs = prepare_data('final_train_tweetsum.jsonl', processor)\nvalid_inputs, valid_summs = prepare_data('final_valid_tweetsum.jsonl', processor)\ntest_inputs, test_summs = prepare_data('final_test_tweetsum.jsonl', processor)\n\n# ----- Save as CSVs\n\ntrain = pd.DataFrame({\"inputs\": train_inputs, \"summaries\": train_summs})\ntrain.to_csv('/content/drive/MyDrive/CIS6930_final/tweetsum_train.csv', index=False)\n\nvalid = pd.DataFrame({\"inputs\": valid_inputs, \"summaries\": valid_summs})\nvalid.to_csv('/content/drive/MyDrive/CIS6930_final/tweetsum_valid.csv', index=False)\n\ntest = pd.DataFrame({\"inputs\": test_inputs, \"summaries\": test_summs})\ntest.to_csv('/content/drive/MyDrive/CIS6930_final/tweetsum_test.csv', index=False)\n", "Not connected to a GPU\nDrive already mounted at /content/drive; to attempt to forcibly remount, call drive.mount(\"/content/drive\", force_remount=True).\n" ] ] ]

[ "code" ]

[ [ "code" ] ]

4ad8f3b4b48fbf17cc0a3ad2c8c843e127aa1b24

19,269

ipynb

Jupyter Notebook

notebooks/03_categorical_pipeline.ipynb

lesteve/scikit-learn-mooc

b822586b98e71dbbf003bde86be57412cb170291

[ "CC-BY-4.0" ]

1

2022-01-25T19:20:21.000Z

notebooks/03_categorical_pipeline.ipynb

lesteve/scikit-learn-mooc

b822586b98e71dbbf003bde86be57412cb170291

[ "CC-BY-4.0" ]

null

notebooks/03_categorical_pipeline.ipynb

lesteve/scikit-learn-mooc

b822586b98e71dbbf003bde86be57412cb170291

[ "CC-BY-4.0" ]

null

32.826235

155

0.627485

[ [ [ "# Encoding of categorical variables\n\nIn this notebook, we will present typical ways of dealing with\n**categorical variables** by encoding them, namely **ordinal encoding** and\n**one-hot encoding**.", "_____no_output_____" ], [ "Let's first load the entire adult dataset containing both numerical and\ncategorical data.", "_____no_output_____" ] ], [ [ "import pandas as pd\n\nadult_census = pd.read_csv(\"../datasets/adult-census.csv\")\n# drop the duplicated column `\"education-num\"` as stated in the first notebook\nadult_census = adult_census.drop(columns=\"education-num\")\n\ntarget_name = \"class\"\ntarget = adult_census[target_name]\n\ndata = adult_census.drop(columns=[target_name])", "_____no_output_____" ] ], [ [ "\n## Identify categorical variables\n\nAs we saw in the previous section, a numerical variable is a\nquantity represented by a real or integer number. These variables can be\nnaturally handled by machine learning algorithms that are typically composed\nof a sequence of arithmetic instructions such as additions and\nmultiplications.\n\nIn contrast, categorical variables have discrete values, typically\nrepresented by string labels (but not only) taken from a finite list of\npossible choices. For instance, the variable `native-country` in our dataset\nis a categorical variable because it encodes the data using a finite list of\npossible countries (along with the `?` symbol when this information is\nmissing):", "_____no_output_____" ] ], [ [ "data[\"native-country\"].value_counts().sort_index()", "_____no_output_____" ] ], [ [ "How can we easily recognize categorical columns among the dataset? Part of\nthe answer lies in the columns' data type:", "_____no_output_____" ] ], [ [ "data.dtypes", "_____no_output_____" ] ], [ [ "If we look at the `\"native-country\"` column, we observe its data type is\n`object`, meaning it contains string values.\n\n## Select features based on their data type\n\nIn the previous notebook, we manually defined the numerical columns. We could\ndo a similar approach. Instead, we will use the scikit-learn helper function\n`make_column_selector`, which allows us to select columns based on\ntheir data type. We will illustrate how to use this helper.", "_____no_output_____" ] ], [ [ "from sklearn.compose import make_column_selector as selector\n\ncategorical_columns_selector = selector(dtype_include=object)\ncategorical_columns = categorical_columns_selector(data)\ncategorical_columns", "_____no_output_____" ] ], [ [ "Here, we created the selector by passing the data type to include; we then\npassed the input dataset to the selector object, which returned a list of\ncolumn names that have the requested data type. We can now filter out the\nunwanted columns:", "_____no_output_____" ] ], [ [ "data_categorical = data[categorical_columns]\ndata_categorical.head()", "_____no_output_____" ], [ "print(f\"The dataset is composed of {data_categorical.shape[1]} features\")", "_____no_output_____" ] ], [ [ "In the remainder of this section, we will present different strategies to\nencode categorical data into numerical data which can be used by a\nmachine-learning algorithm.", "_____no_output_____" ], [ "## Strategies to encode categories\n\n### Encoding ordinal categories\n\nThe most intuitive strategy is to encode each category with a different\nnumber. The `OrdinalEncoder` will transform the data in such manner.\nWe will start by encoding a single column to understand how the encoding\nworks.", "_____no_output_____" ] ], [ [ "from sklearn.preprocessing import OrdinalEncoder\n\neducation_column = data_categorical[[\"education\"]]\n\nencoder = OrdinalEncoder()\neducation_encoded = encoder.fit_transform(education_column)\neducation_encoded", "_____no_output_____" ] ], [ [ "We see that each category in `\"education\"` has been replaced by a numeric\nvalue. We could check the mapping between the categories and the numerical\nvalues by checking the fitted attribute `categories_`.", "_____no_output_____" ] ], [ [ "encoder.categories_", "_____no_output_____" ] ], [ [ "Now, we can check the encoding applied on all categorical features.", "_____no_output_____" ] ], [ [ "data_encoded = encoder.fit_transform(data_categorical)\ndata_encoded[:5]", "_____no_output_____" ], [ "print(\n f\"The dataset encoded contains {data_encoded.shape[1]} features\")", "_____no_output_____" ] ], [ [ "We see that the categories have been encoded for each feature (column)\nindependently. We also note that the number of features before and after the\nencoding is the same.\n\nHowever, be careful when applying this encoding strategy:\nusing this integer representation leads downstream predictive models\nto assume that the values are ordered (0 < 1 < 2 < 3... for instance).\n\nBy default, `OrdinalEncoder` uses a lexicographical strategy to map string\ncategory labels to integers. This strategy is arbitrary and often\nmeaningless. For instance, suppose the dataset has a categorical variable\nnamed `\"size\"` with categories such as \"S\", \"M\", \"L\", \"XL\". We would like the\ninteger representation to respect the meaning of the sizes by mapping them to\nincreasing integers such as `0, 1, 2, 3`.\nHowever, the lexicographical strategy used by default would map the labels\n\"S\", \"M\", \"L\", \"XL\" to 2, 1, 0, 3, by following the alphabetical order.\n\nThe `OrdinalEncoder` class accepts a `categories` constructor argument to\npass categories in the expected ordering explicitly. You can find more\ninformation in the\n[scikit-learn documentation](https://scikit-learn.org/stable/modules/preprocessing.html#encoding-categorical-features)\nif needed.\n\nIf a categorical variable does not carry any meaningful order information\nthen this encoding might be misleading to downstream statistical models and\nyou might consider using one-hot encoding instead (see below).\n\n### Encoding nominal categories (without assuming any order)\n\n`OneHotEncoder` is an alternative encoder that prevents the downstream\nmodels to make a false assumption about the ordering of categories. For a\ngiven feature, it will create as many new columns as there are possible\ncategories. For a given sample, the value of the column corresponding to the\ncategory will be set to `1` while all the columns of the other categories\nwill be set to `0`.\n\nWe will start by encoding a single feature (e.g. `\"education\"`) to illustrate\nhow the encoding works.", "_____no_output_____" ] ], [ [ "from sklearn.preprocessing import OneHotEncoder\n\nencoder = OneHotEncoder(sparse=False)\neducation_encoded = encoder.fit_transform(education_column)\neducation_encoded", "_____no_output_____" ] ], [ [ "<div class=\"admonition note alert alert-info\">\nNote\n<tt class=\"docutils literal\">sparse=False</tt> is used in the <tt class=\"docutils literal\">OneHotEncoder</tt> for didactic purposes, namely\neasier visualization of the data.\nSparse matrices are efficient data structures when most of your matrix\nelements are zero. They won't be covered in detail in this course. If you\nwant more details about them, you can look at\n<a class=\"reference external\" href=\"https://scipy-lectures.org/advanced/scipy_sparse/introduction.html#why-sparse-matrices\">this</a>.\n</div>", "_____no_output_____" ], [ "We see that encoding a single feature will give a NumPy array full of zeros\nand ones. We can get a better understanding using the associated feature\nnames resulting from the transformation.", "_____no_output_____" ] ], [ [ "feature_names = encoder.get_feature_names_out(input_features=[\"education\"])\neducation_encoded = pd.DataFrame(education_encoded, columns=feature_names)\neducation_encoded", "_____no_output_____" ] ], [ [ "As we can see, each category (unique value) became a column; the encoding\nreturned, for each sample, a 1 to specify which category it belongs to.\n\nLet's apply this encoding on the full dataset.", "_____no_output_____" ] ], [ [ "print(\n f\"The dataset is composed of {data_categorical.shape[1]} features\")\ndata_categorical.head()", "_____no_output_____" ], [ "data_encoded = encoder.fit_transform(data_categorical)\ndata_encoded[:5]", "_____no_output_____" ], [ "print(\n f\"The encoded dataset contains {data_encoded.shape[1]} features\")", "_____no_output_____" ] ], [ [ "Let's wrap this NumPy array in a dataframe with informative column names as\nprovided by the encoder object:", "_____no_output_____" ] ], [ [ "columns_encoded = encoder.get_feature_names_out(data_categorical.columns)\npd.DataFrame(data_encoded, columns=columns_encoded).head()", "_____no_output_____" ] ], [ [ "Look at how the `\"workclass\"` variable of the 3 first records has been\nencoded and compare this to the original string representation.\n\nThe number of features after the encoding is more than 10 times larger than\nin the original data because some variables such as `occupation` and\n`native-country` have many possible categories.", "_____no_output_____" ], [ "### Choosing an encoding strategy\n\nChoosing an encoding strategy will depend on the underlying models and the\ntype of categories (i.e. ordinal vs. nominal).", "_____no_output_____" ], [ "<div class=\"admonition note alert alert-info\">\nNote\nIn general <tt class=\"docutils literal\">OneHotEncoder</tt> is the encoding strategy used when the\ndownstream models are linear models while <tt class=\"docutils literal\">OrdinalEncoder</tt> is often a\ngood strategy with tree-based models.\n</div>", "_____no_output_____" ], [ "\nUsing an `OrdinalEncoder` will output ordinal categories. This means\nthat there is an order in the resulting categories (e.g. `0 < 1 < 2`). The\nimpact of violating this ordering assumption is really dependent on the\ndownstream models. Linear models will be impacted by misordered categories\nwhile tree-based models will not.\n\nYou can still use an `OrdinalEncoder` with linear models but you need to be\nsure that:\n- the original categories (before encoding) have an ordering;\n- the encoded categories follow the same ordering than the original\n categories.\nThe **next exercise** highlights the issue of misusing `OrdinalEncoder` with\na linear model.\n\nOne-hot encoding categorical variables with high cardinality can cause \ncomputational inefficiency in tree-based models. Because of this, it is not recommended\nto use `OneHotEncoder` in such cases even if the original categories do not \nhave a given order. We will show this in the **final exercise** of this sequence.", "_____no_output_____" ], [ "## Evaluate our predictive pipeline\n\nWe can now integrate this encoder inside a machine learning pipeline like we\ndid with numerical data: let's train a linear classifier on the encoded data\nand check the generalization performance of this machine learning pipeline using\ncross-validation.\n\nBefore we create the pipeline, we have to linger on the `native-country`.\nLet's recall some statistics regarding this column.", "_____no_output_____" ] ], [ [ "data[\"native-country\"].value_counts()", "_____no_output_____" ] ], [ [ "We see that the `Holand-Netherlands` category is occurring rarely. This will\nbe a problem during cross-validation: if the sample ends up in the test set\nduring splitting then the classifier would not have seen the category during\ntraining and will not be able to encode it.\n\nIn scikit-learn, there are two solutions to bypass this issue:\n\n* list all the possible categories and provide it to the encoder via the\n keyword argument `categories`;\n* use the parameter `handle_unknown`.\n\nHere, we will use the latter solution for simplicity.", "_____no_output_____" ], [ "<div class=\"admonition tip alert alert-warning\">\nTip\nBe aware the <tt class=\"docutils literal\">OrdinalEncoder</tt> exposes as well a parameter\n<tt class=\"docutils literal\">handle_unknown</tt>. It can be set to <tt class=\"docutils literal\">use_encoded_value</tt> and by setting\n<tt class=\"docutils literal\">unknown_value</tt> to handle rare categories. You are going to use these\nparameters in the next exercise.\n</div>", "_____no_output_____" ], [ "We can now create our machine learning pipeline.", "_____no_output_____" ] ], [ [ "from sklearn.pipeline import make_pipeline\nfrom sklearn.linear_model import LogisticRegression\n\nmodel = make_pipeline(\n OneHotEncoder(handle_unknown=\"ignore\"), LogisticRegression(max_iter=500)\n)", "_____no_output_____" ] ], [ [ "<div class=\"admonition note alert alert-info\">\nNote\nHere, we need to increase the maximum number of iterations to obtain a fully\nconverged <tt class=\"docutils literal\">LogisticRegression</tt> and silence a <tt class=\"docutils literal\">ConvergenceWarning</tt>. Contrary\nto the numerical features, the one-hot encoded categorical features are all\non the same scale (values are 0 or 1), so they would not benefit from\nscaling. In this case, increasing <tt class=\"docutils literal\">max_iter</tt> is the right thing to do.\n</div>", "_____no_output_____" ], [ "Finally, we can check the model's generalization performance only using the\ncategorical columns.", "_____no_output_____" ] ], [ [ "from sklearn.model_selection import cross_validate\ncv_results = cross_validate(model, data_categorical, target)\ncv_results", "_____no_output_____" ], [ "scores = cv_results[\"test_score\"]\nprint(f\"The accuracy is: {scores.mean():.3f} +/- {scores.std():.3f}\")", "_____no_output_____" ] ], [ [ "As you can see, this representation of the categorical variables is\nslightly more predictive of the revenue than the numerical variables\nthat we used previously.", "_____no_output_____" ], [ "\nIn this notebook we have:\n* seen two common strategies for encoding categorical features: **ordinal\n encoding** and **one-hot encoding**;\n* used a **pipeline** to use a **one-hot encoder** before fitting a logistic\n regression.", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown" ]

[ [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown", "markdown" ], [ "code", "code" ], [ "markdown", "markdown" ] ]

4ad906a2f4a692b6ea36dd34c67e0f3246560bd7

9,102

ipynb

Jupyter Notebook

notebooks/ex_cosmology/pywt_swt.ipynb

csinva/local-vae

a28a3e58369cb18053c27a994a60d66b9467829d

[ "MIT" ]

2

2021-11-28T06:36:22.000Z

2021-12-12T08:21:58.000Z

notebooks/ex_cosmology/pywt_swt.ipynb

csinva/local-vae

a28a3e58369cb18053c27a994a60d66b9467829d

[ "MIT" ]

null

notebooks/ex_cosmology/pywt_swt.ipynb

csinva/local-vae

a28a3e58369cb18053c27a994a60d66b9467829d

[ "MIT" ]

null

32.507143

101

0.504395

[ [ [ "%load_ext autoreload\n%autoreload 2\n%matplotlib inline\nimport matplotlib.pyplot as plt\nimport numpy as np\nimport torch\ndevice = 'cuda' if torch.cuda.is_available() else 'cpu'\nimport os,sys\nopj = os.path.join\nfrom copy import deepcopy\nimport pickle as pkl\n\nsys.path.append('../../src')\nsys.path.append('../../src/dsets/cosmology')\nfrom dset import get_dataloader\nfrom viz import viz_im_r, cshow, viz_filters\nfrom sim_cosmology import p, load_dataloader_and_pretrained_model\nfrom losses import get_loss_f\nfrom train import Trainer, Validator\n\n# wt modules\nfrom wavelet_transform import Wavelet_Transform, Attributer, get_2dfilts, initialize_filters\nfrom utils import tuple_L1Loss, tuple_L2Loss, thresh_attrs, viz_list", "_____no_output_____" ], [ "import pywt\nimport warnings\nfrom itertools import product\n\nimport numpy as np\n\nfrom pywt._c99_config import _have_c99_complex\nfrom pywt._extensions._dwt import idwt_single\nfrom pywt._extensions._swt import swt_max_level, swt as _swt, swt_axis as _swt_axis\nfrom pywt._extensions._pywt import Wavelet, Modes, _check_dtype\nfrom pywt._multidim import idwt2, idwtn\nfrom pywt._utils import _as_wavelet, _wavelets_per_axis", "_____no_output_____" ], [ "# get dataloader and model\n(train_loader, test_loader), model = load_dataloader_and_pretrained_model(p, img_size=256)\ntorch.manual_seed(p.seed)\nim = iter(test_loader).next()[0][0:1].numpy().squeeze()", "_____no_output_____" ], [ "db2 = pywt.Wavelet('db2')", "_____no_output_____" ], [ "start_level = 0\naxes = (-2, -1)\ntrim_approx = False\nnorm = False", "_____no_output_____" ], [ "data = np.asarray(im)\n# coefs = swtn(data, wavelet, level, start_level, axes, trim_approx, norm)\naxes = [a + data.ndim if a < 0 else a for a in axes]\nnum_axes = len(axes)\nprint(axes)\nwavelets = _wavelets_per_axis(db2, axes)\nret = []\nlevel = 1", "[0, 1]\n" ], [ "i = 0\ncoeffs = [('', data)]\naxis = axes[0]\nwavelet = wavelets[0]\nnew_coeffs = []\nsubband, x = coeffs[0]\ncA, cD = _swt_axis(x, wavelet, level=1, start_level=i,\n axis=axis)[0]\nnew_coeffs.extend([(subband + 'a', cA),\n (subband + 'd', cD)])\ncoeffs = new_coeffs", "_____no_output_____" ], [ "coeffs", "_____no_output_____" ], [ "def swtn(data, wavelet, level, start_level=0, axes=None, trim_approx=False,\n norm=False):\n\n wavelets = _wavelets_per_axis(wavelet, axes)\n if norm:\n if not np.all([wav.orthogonal for wav in wavelets]):\n warnings.warn(\n \"norm=True, but the wavelets used are not orthogonal: \\n\"\n \"\\tThe conditions for energy preservation are not satisfied.\")\n wavelets = [_rescale_wavelet_filterbank(wav, 1/np.sqrt(2))\n for wav in wavelets]\n ret = []\n for i in range(start_level, start_level + level):\n coeffs = [('', data)]\n for axis, wavelet in zip(axes, wavelets):\n new_coeffs = []\n for subband, x in coeffs:\n cA, cD = _swt_axis(x, wavelet, level=1, start_level=i,\n axis=axis)[0]\n new_coeffs.extend([(subband + 'a', cA),\n (subband + 'd', cD)])\n coeffs = new_coeffs\n\n coeffs = dict(coeffs)\n ret.append(coeffs)\n\n # data for the next level is the approximation coeffs from this level\n data = coeffs['a' * num_axes]\n if trim_approx:\n coeffs.pop('a' * num_axes)\n if trim_approx:\n ret.append(data)\n ret.reverse()\n return ret", "_____no_output_____" ], [ "def swt2(data, wavelet, level, start_level=0, axes=(-2, -1),\n trim_approx=False, norm=False):\n axes = tuple(axes)\n data = np.asarray(data)\n if len(axes) != 2:\n raise ValueError(\"Expected 2 axes\")\n if len(axes) != len(set(axes)):\n raise ValueError(\"The axes passed to swt2 must be unique.\")\n if data.ndim < len(np.unique(axes)):\n raise ValueError(\"Input array has fewer dimensions than the specified \"\n \"axes\")\n\n coefs = swtn(data, wavelet, level, start_level, axes, trim_approx, norm)\n ret = []\n if trim_approx:\n ret.append(coefs[0])\n coefs = coefs[1:]\n for c in coefs:\n if trim_approx:\n ret.append((c['da'], c['ad'], c['dd']))\n else:\n ret.append((c['aa'], (c['da'], c['ad'], c['dd'])))\n return ret", "_____no_output_____" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad90992fd68124563edc1c3c66f6f3bcd23c7ea

5,007

ipynb

Jupyter Notebook

doc/point_groups.ipynb

bm424/texpy

8d78b568209a6da36fc831c6bc9e2b0cb4c740c8

[ "MIT" ]

6

2018-02-05T18:37:10.000Z

2018-10-07T22:07:26.000Z

doc/point_groups.ipynb

bm424/texpy

8d78b568209a6da36fc831c6bc9e2b0cb4c740c8

[ "MIT" ]

5

2018-11-04T18:06:28.000Z

2019-09-13T11:22:43.000Z

doc/point_groups.ipynb

bm424/texpy

8d78b568209a6da36fc831c6bc9e2b0cb4c740c8

[ "MIT" ]

3

2019-04-27T09:24:28.000Z

2019-09-13T10:24:57.000Z

31.89172

343

0.536249

[ [ [ "This notebook is part of the orix documentation https://orix.readthedocs.io. Links to the documentation won’t work from the notebook.", "_____no_output_____" ], [ "## Visualizing point groups\n\nPoint group symmetry operations are shown here in the stereographic projection.\n\nVectors located on the upper (`z >= 0`) hemisphere are displayed as points (`o`), whereas vectors on the lower hemisphere are reprojected onto the upper hemisphere and shown as crosses (`+`) by default. For more information about plot formatting and visualization, see [Vector3d.scatter()](reference.rst#orix.vector.Vector3d.scatter).\n\nMore explanation of these figures is provided at http://xrayweb.chem.ou.edu/notes/symmetry.html#point.", "_____no_output_____" ] ], [ [ "%matplotlib inline\nfrom matplotlib import pyplot as plt\nimport numpy as np\n\nfrom orix import plot\nfrom orix.quaternion import Rotation, symmetry\nfrom orix.vector import Vector3d\n\nplt.rcParams.update({\"font.size\": 15})", "_____no_output_____" ] ], [ [ "For example, the `O (432)` point group:", "_____no_output_____" ] ], [ [ "symmetry.O.plot()", "_____no_output_____" ] ], [ [ "The stereographic projection of all point groups is shown below:", "_____no_output_____" ] ], [ [ "# fmt: off\nschoenflies = [\n \"C1\", \"Ci\", # triclinic,\n \"C2x\", \"C2y\", \"C2z\", \"Csx\", \"Csy\", \"Csz\", \"C2h\", # monoclinic\n \"D2\", \"C2v\", \"D2h\", # orthorhombic\n \"C4\", \"S4\", \"C4h\", \"D4\", \"C4v\", \"D2d\", \"D4h\", # tetragonal\n \"C3\", \"S6\", \"D3x\", \"D3y\", \"D3\", \"C3v\", \"D3d\", \"C6\", # trigonal\n \"C3h\", \"C6h\", \"D6\", \"C6v\", \"D3h\", \"D6h\", # hexagonal\n \"T\", \"Th\", \"O\", \"Td\", \"Oh\", # cubic\n]\n# fmt: on\n\nassert len(symmetry._groups) == len(schoenflies)\n\nschoenflies = [s for s in schoenflies if not (s.endswith(\"x\") or s.endswith(\"y\"))]\n\nassert len(schoenflies) == 32\n\norientation = Rotation.from_axes_angles((-1, 8, 1), np.deg2rad(65))\n\nfig, ax = plt.subplots(\n nrows=8, ncols=4, figsize=(10, 20), subplot_kw=dict(projection=\"stereographic\")\n)\nax = ax.ravel()\n\nfor i, s in enumerate(schoenflies):\n sym = getattr(symmetry, s)\n\n ori_sym = sym.outer(orientation)\n v = ori_sym * Vector3d.zvector()\n\n # reflection in the projection plane (x-y) is performed internally in\n # Symmetry.plot() or when using the `reproject=True` argument for\n # Vector3d.scatter()\n v_reproject = Vector3d(v.data.copy())\n v_reproject.z *= -1\n\n # the Symmetry marker formatting for vectors on the upper and lower hemisphere\n # can be set using `kwargs` and `reproject_scatter_kwargs`, respectively, for\n # Symmetry.plot()\n\n # vectors on the upper hemisphere are shown as open circles\n ax[i].scatter(v, marker=\"o\", fc=\"None\", ec=\"k\", s=150)\n # vectors on the lower hemisphere are reprojected onto the upper hemisphere and\n # shown as crosses\n ax[i].scatter(v_reproject, marker=\"+\", ec=\"C0\", s=150)\n\n ax[i].set_title(f\"${s}$ $({sym.name})$\")\n ax[i].set_labels(\"a\", \"b\", None)\n\nfig.tight_layout()", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad90a3d7bad802b1d2c1f8edcb00f1ec0cbbb34

15,059

ipynb

Jupyter Notebook

09. Cloud/AV/XGB_no tuning_working code.ipynb

vm1729/Openthink

63b3ae65c85b88bdffb71f5789a7799ae389fc82

[ "MIT" ]

1

2021-01-14T14:32:20.000Z

09. Cloud/AV/XGB_no tuning_working code.ipynb

vm1729/Openthink

63b3ae65c85b88bdffb71f5789a7799ae389fc82

[ "MIT" ]

2

2020-09-12T10:45:18.000Z

09. Cloud/AV/XGB_no tuning_working code.ipynb

vm1729/Openthink

63b3ae65c85b88bdffb71f5789a7799ae389fc82

[ "MIT" ]

2

2020-04-21T15:09:00.000Z

2020-12-01T07:27:10.000Z

53.782143

7,172

0.745866

[ [ [ "import pandas as pd\n\nfrom sklearn import model_selection\nfrom sklearn.metrics import accuracy_score\nfrom sklearn.preprocessing import LabelEncoder\n\nimport xgboost as xgb\nfrom xgboost.sklearn import XGBClassifier\nfrom sklearn import metrics #Additional scklearn functions\n#from sklearn import cross_validation\nfrom sklearn.model_selection import GridSearchCV", "_____no_output_____" ], [ "target='Disbursed'\nIDcol = 'ID'", "_____no_output_____" ], [ "train = pd.read_csv('train_modified.csv')\n\npredictors = [x for x in train.columns if x not in [target, IDcol]]", "_____no_output_____" ], [ "train.columns,test", "_____no_output_____" ], [ "model = xgb.XGBClassifier()\n# one hot encoding should be done as preparatory steps\n# predictors is the name of columns for X\n# Disbursed is the target column name", "_____no_output_____" ], [ "model.fit(train[predictors], train['Disbursed'],eval_metric='auc')\nprint(model)\n", "XGBClassifier(base_score=0.5, booster='gbtree', colsample_bylevel=1,\n colsample_bytree=1, gamma=0, learning_rate=0.1, max_delta_step=0,\n max_depth=3, min_child_weight=1, missing=None, n_estimators=100,\n n_jobs=1, nthread=None, objective='binary:logistic', random_state=0,\n reg_alpha=0, reg_lambda=1, scale_pos_weight=1, seed=None,\n silent=True, subsample=1)\n" ], [ "# make predictions for test data\ny_pred = model.predict(train[predictors])\ny_predprob = model.predict_proba(train[predictors])[:,1]\n\n#predictions = [round(value) for value in y_pred]\n# evaluate predictions\n#accuracy = accuracy_score(y_test, predictions)\n#print(\"Accuracy: %.2f%%\" % (accuracy * 100.0))", "_____no_output_____" ], [ " print (\"Accuracy : %.4g\" % metrics.accuracy_score(train['Disbursed'].values, y_pred))\n print (\"AUC Score (Train): %f\" % metrics.roc_auc_score(train['Disbursed'], y_predprob))\n", "Accuracy : 0.9854\nAUC Score (Train): 0.851058\n" ], [ "import numpy as np\nimport matplotlib.pyplot as plt\n#fscore = model.get_booster().get_score(importance_type='weight')\nfscore = pd.Series(model.get_booster().get_score(importance_type='weight'))\nfscore.plot(kind='bar', title='Feature Importances')\nplt.ylabel('Feature Importance Score')\n", "_____no_output_____" ], [ "test = pd.read_csv('test_modified.csv')\n\ntest_predictors = [x for x in train.columns if x not in [target, IDcol]]", "_____no_output_____" ], [ "import matplotlib.pyplot as plt\nplt.hist(model.predict(test[test_predictors]))", "_____no_output_____" ], [ "def modelpred(alg, dtest, predictors):\n \n #Predict training set:\n dtest_predictions = alg.predict(dtest[predictors])\n dtest_predprob = alg.predict_proba(dtest[predictors])[:,1]\n return dtest_predictions,dtest_predprob", "_____no_output_____" ], [ "val1,val2=modelpred(model, test, test_predictors)", "_____no_output_____" ], [ "val1.value_counts()", "_____no_output_____" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad90efa8951ae570fc64f4d2be543dc37bfbbe6

3,956

ipynb

Jupyter Notebook

facemask_detection_cnn_video.ipynb

arryaaas/Realtime-Facemask-Detection-CNN

6942c1c68646f510a749e598174058d1d614eb7c

[ "MIT" ]

null

facemask_detection_cnn_video.ipynb

arryaaas/Realtime-Facemask-Detection-CNN

6942c1c68646f510a749e598174058d1d614eb7c

[ "MIT" ]

null

facemask_detection_cnn_video.ipynb

arryaaas/Realtime-Facemask-Detection-CNN

6942c1c68646f510a749e598174058d1d614eb7c

[ "MIT" ]

null

28.666667

129

0.500506

[ [ [ "# Pengenalan Wajah Bermasker Dan Tidak Menggunakan CNN\nMochammad Arya Salsabila ( 19081010001 ) / Nadhif Mahardika Awandi ( 19081010064 ) / Kecerdasan Buatan Paralel D Tahun 2021", "_____no_output_____" ], [ "# Klasifikasi Menggunakan Video Realtime", "_____no_output_____" ] ], [ [ "import cv2\nimport datetime\nimport numpy as np\nfrom keras.preprocessing import image", "_____no_output_____" ], [ "from keras.models import load_model\n\nmodel = load_model(\"D:/Code/Notebook/realtime-facemask-detection-cnn/model_without_opt.h5\", compile = False)", "_____no_output_____" ], [ "rect_size = 4\ncap = cv2.VideoCapture(0)\n\nhaarcascade = cv2.CascadeClassifier(\"haarcascade_frontalface_default.xml\")\n\nwhile True:\n (_, img) = cap.read()\n img = cv2.flip(img, 1, 1)\n \n rerect_size = cv2.resize(img, (img.shape[1] // rect_size, img.shape[0] // rect_size))\n faces = haarcascade.detectMultiScale(rerect_size)\n \n for f in faces:\n (x, y, w, h) = [v * rect_size for v in f]\n \n face_img = img[y : y + h, x : x + w]\n \n cv2.imwrite(\"temp.jpg\", face_img)\n test_image = image.load_img(\"temp.jpg\", target_size = (150, 150, 3))\n test_image = image.img_to_array(test_image)\n test_image = np.expand_dims(test_image, axis=0)\n \n classes = model.predict(test_image)\n \n if classes == 0:\n cv2.rectangle(img, (x, y), (x + w, y + h), (0, 255, 0), 2)\n cv2.rectangle(img, (x, y - 40), (x + w, y), (0, 255, 0), -1)\n cv2.putText(img, \"With Mask\", (x, y - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.75, (255, 255, 255), 2)\n else:\n cv2.rectangle(img, (x, y), (x + w, y + h), (0, 0, 255), 2)\n cv2.rectangle(img, (x, y - 40), (x + w, y), (0, 0, 255), -1)\n cv2.putText(img, \"Without Mask\", (x, y - 10), cv2.FONT_HERSHEY_SIMPLEX, 0.75, (255, 255, 255), 2)\n \n now = str(datetime.datetime.now())\n cv2.putText(img, now, (400,450), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (255,255,255), 1)\n \n cv2.imshow(\"LIVE\", img)\n \n if cv2.waitKey(1) & 0xFF == ord('q'):\n break\n \ncap.release()\ncv2.destroyAllWindows()", "_____no_output_____" ] ], [ [ "<div class=\"alert alert-block alert-info\">\n Info: Tekan tombol q pada keyboard untuk keluar dari OpenCv.\n</div>", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown" ]

[ [ "markdown", "markdown" ], [ "code", "code", "code" ], [ "markdown" ] ]

4ad90fb045a3ff836e2098d1454476116425ab5d

177,287

ipynb

Jupyter Notebook

LoanInterestPrediction/Loan_Interest_Prediction.ipynb

MayukhSobo/AnalyticsVidya_Contests

a21079f8d217a35e88e72e88233c7ef0b8dd348b

[ "BSD-3-Clause" ]

11

2017-09-22T08:12:04.000Z

2021-10-30T14:30:44.000Z

LoanInterestPrediction/Loan_Interest_Prediction.ipynb

Manish041997/AnalyticsVidya_Contests

a21079f8d217a35e88e72e88233c7ef0b8dd348b

[ "BSD-3-Clause" ]

null

LoanInterestPrediction/Loan_Interest_Prediction.ipynb

Manish041997/AnalyticsVidya_Contests

a21079f8d217a35e88e72e88233c7ef0b8dd348b

[ "BSD-3-Clause" ]

5

2017-10-03T12:09:11.000Z

2019-08-03T14:05:47.000Z

48.505335

14,140

0.61546

[ [ [ "## Loan EDA", "_____no_output_____" ] ], [ [ "import pandas as pd\nimport numpy as np", "_____no_output_____" ], [ "dtrain = pd.read_csv('data/train.csv')\ntest = pd.read_csv('data/test.csv')", "_____no_output_____" ] ], [ [ "## Data Cleaning", "_____no_output_____" ] ], [ [ "dtrain.head()", "_____no_output_____" ], [ "dtrain.shape", "_____no_output_____" ], [ "# Removing the commas form `Loan_Amount_Requested`\ndtrain['Loan_Amount_Requested'] = dtrain.Loan_Amount_Requested.str.replace(',', '').astype(int)\ntest['Loan_Amount_Requested'] = test.Loan_Amount_Requested.str.replace(',', '').astype(int)", "_____no_output_____" ], [ "# Filling 0 for `Annual_Income` column\ndtrain['Annual_Income'] = dtrain['Annual_Income'].fillna(0).astype(int)\ntest['Annual_Income'] = test['Annual_Income'].fillna(0).astype(int)", "_____no_output_____" ], [ "# Showing the different types of values for `Home_Owner`\ndtrain['Home_Owner'] = dtrain['Home_Owner'].fillna('NA')\ntest['Home_Owner'] = test['Home_Owner'].fillna('NA')\nprint(dtrain.Home_Owner.value_counts())", "Mortgage 70345\nRent 56031\nNA 25349\nOwn 12525\nOther 49\nNone 10\nName: Home_Owner, dtype: int64\n" ] ], [ [ "We converted the ```NaN``` in ```Home_Owner``` into ```NA```. We are going to calculate the hash value of the string and we don't know how ```NaN``` should be replced. Now we see that there are almost **25349** rows which was NaN. Dropping these would cause loosing a lot of data. Hence we replaced these with sting \"NA\" and then converted these into hash values", "_____no_output_____" ] ], [ [ "# Filling 0 for missing `Months_Since_Deliquency`\ndtrain['Months_Since_Deliquency'] = dtrain['Months_Since_Deliquency'].fillna(0)\ntest['Months_Since_Deliquency'] = test['Months_Since_Deliquency'].fillna(0)", "_____no_output_____" ], [ "dtrain.isnull().values.any()", "_____no_output_____" ], [ "dtrain['Length_Employed'] = dtrain['Length_Employed'].fillna('0 year')\ntest['Length_Employed'] = test['Length_Employed'].fillna('0 year')", "_____no_output_____" ], [ "def convert_length_employed(elem):\n if elem[0] == '<':\n return 0.5 # because mean of 0 to 1 is 0.5\n elif str(elem[2]) == '+':\n return 15.0 # because mean of 10 to 20 is 15\n elif str(elem) == '0 year':\n return 0.0\n else:\n return float(str(elem).split()[0])\ndtrain['Length_Employed'] = dtrain['Length_Employed'].apply(convert_length_employed)\ntest['Length_Employed'] = test['Length_Employed'].apply(convert_length_employed)", "_____no_output_____" ], [ "dtrain['Loan_Grade'] = dtrain['Loan_Grade'].fillna('NA')\ntest['Loan_Grade'] = test['Loan_Grade'].fillna('NA')\ndtrain.Loan_Grade.value_counts()", "_____no_output_____" ], [ "# dtrain[(dtrain.Annual_Income == 0) & (dtrain.Income_Verified == 'not verified')]", "_____no_output_____" ], [ "from sklearn.preprocessing import LabelEncoder\nnumber = LabelEncoder()\ndtrain['Loan_Grade'] = number.fit_transform(dtrain.Loan_Grade.astype('str'))\ndtrain['Income_Verified'] = number.fit_transform(dtrain.Income_Verified.astype('str'))\ndtrain['Area_Type'] = number.fit_transform(dtrain.Area_Type.astype('str'))\ndtrain['Gender'] = number.fit_transform(dtrain.Gender.astype('str'))\n\ntest['Loan_Grade'] = number.fit_transform(test.Loan_Grade.astype('str'))\ntest['Income_Verified'] = number.fit_transform(test.Income_Verified.astype('str'))\ntest['Area_Type'] = number.fit_transform(test.Area_Type.astype('str'))\ntest['Gender'] = number.fit_transform(test.Gender.astype('str'))", "_____no_output_____" ], [ "dtrain.head()", "_____no_output_____" ], [ "# Converting `Purpose_Of_Loan` and `Home_Owner` into hash\nimport hashlib\n\ndef convert_to_hashes(elem):\n return round(int(hashlib.md5(elem.encode('utf-8')).hexdigest(), 16) / 1e35, 5)\n\ndtrain['Purpose_Of_Loan'] = dtrain['Purpose_Of_Loan'].apply(convert_to_hashes)\ndtrain['Home_Owner'] = dtrain['Home_Owner'].apply(convert_to_hashes)\n\ntest['Purpose_Of_Loan'] = test['Purpose_Of_Loan'].apply(convert_to_hashes)\ntest['Home_Owner'] = test['Home_Owner'].apply(convert_to_hashes)", "_____no_output_____" ], [ "import xgboost as xgb\n\nfeatures = np.array(dtrain.iloc[:, 1:-1])\nlabels = np.array(dtrain.iloc[:, -1])", "_____no_output_____" ], [ "import operator\nfrom xgboost import plot_importance\nfrom matplotlib import pylab as plt\nfrom collections import OrderedDict\ndef xgb_feature_importance(features, labels, num_rounds, fnames, plot=False):\n param = {}\n param['objective'] = 'multi:softmax'\n param['eta'] = 0.1\n param['max_depth'] = 6\n param['silent'] = 1\n param['num_class'] = 4\n param['eval_metric'] = \"merror\"\n param['min_child_weight'] = 1\n param['subsample'] = 0.7\n param['colsample_bytree'] = 0.7\n param['seed'] = 42\n nrounds = num_rounds\n xgtrain = xgb.DMatrix(features, label=labels)\n xgb_params = list(param.items())\n gbdt = xgb.train(xgb_params, xgtrain, nrounds)\n importance = sorted(gbdt.get_fscore().items(), key=operator.itemgetter(1), reverse=True)\n if plot:\n df = pd.DataFrame(importance, columns=['feature', 'fscore'])\n df['fscore'] = df['fscore'] / df['fscore'].sum()\n plt.figure()\n df.plot(kind='bar', x='feature', y='fscore', legend=False, figsize=(8, 8))\n plt.title('XGBoost Feature Importance')\n plt.xlabel('relative importance')\n plt.show()\n else:\n# fnames = dtrain.columns.values[1:-1].tolist()\n imp_features = OrderedDict()\n imps = dict(importance)\n for each in list(imps.keys()):\n index = int(each.split('f')[-1])\n imp_features[fnames[index]] = imps[each]\n return imp_features\n\nxgb_feature_importance(features, labels, num_rounds=1000, fnames=dtrain.columns.values[1:-1].tolist(), plot=True)", "_____no_output_____" ] ], [ [ "## Features Scores", "_____no_output_____" ], [ "```python\nOrderedDict([('Debt_To_Income', 29377),\n ('Loan_Amount_Requested', 22157),\n ('Annual_Income', 21378),\n ('Total_Accounts', 16675),\n ('Months_Since_Deliquency', 13287),\n ('Number_Open_Accounts', 13016),\n ('Length_Employed', 9140),\n ('Loan_Grade', 7906),\n ('Purpose_Of_Loan', 7284),\n ('Inquiries_Last_6Mo', 5691),\n ('Home_Owner', 4946),\n ('Income_Verified', 4434),\n ('Area_Type', 3755),\n ('Gender', 2027)])\n```", "_____no_output_____" ], [ "## Feature Creation", "_____no_output_____" ], [ "### 1. RatioOfLoanAndIncome", "_____no_output_____" ] ], [ [ "dtrain['RatioOfLoanAndIncome'] = dtrain.Loan_Amount_Requested / (dtrain.Annual_Income + 1)\ntest['RatioOfLoanAndIncome'] = test.Loan_Amount_Requested / (test.Annual_Income + 1)", "_____no_output_____" ] ], [ [ "### 2. RatioOfOpenAccToTotalAcc", "_____no_output_____" ] ], [ [ "dtrain['RatioOfOpenAccToTotalAcc'] = dtrain.Number_Open_Accounts / (dtrain.Total_Accounts + 0.001)\ntest['RatioOfOpenAccToTotalAcc'] = test.Number_Open_Accounts / (test.Total_Accounts + 0.001)", "_____no_output_____" ], [ "dtrain.drop(['Interest_Rate', 'Loan_Amount_Requested',\n 'Annual_Income', 'Number_Open_Accounts', 'Total_Accounts'], inplace=True, axis=1)\ndtrain['Interest_Rate'] = labels\n\ntest.drop(['Loan_Amount_Requested',\n 'Annual_Income', 'Number_Open_Accounts', 'Total_Accounts'], inplace=True, axis=1)", "_____no_output_____" ], [ "features = np.array(dtrain.iloc[:, 1:-1])", "_____no_output_____" ], [ "testFeatures = np.array(test.iloc[:, 1:])", "_____no_output_____" ], [ "xgb_feature_importance(features, labels, num_rounds=1000, fnames=dtrain.columns.values[1:-1].tolist(), plot=True)", "_____no_output_____" ] ], [ [ "## Feature Score with new features", "_____no_output_____" ], [ "```python\nOrderedDict([('Debt_To_Income', 23402),\n ('RatioOfLoanAndIncome', 20254),\n ('RatioOfOpenAccToTotalAcc', 17868),\n ('Loan_Amount_Requested', 16661),\n ('Annual_Income', 14228),\n ('Total_Accounts', 11892),\n ('Months_Since_Deliquency', 10293),\n ('Number_Open_Accounts', 8553),\n ('Length_Employed', 7614),\n ('Loan_Grade', 6938),\n ('Purpose_Of_Loan', 6013),\n ('Inquiries_Last_6Mo', 4284),\n ('Home_Owner', 3760),\n ('Income_Verified', 3451),\n ('Area_Type', 2892),\n ('Gender', 1708)])\n```", "_____no_output_____" ] ], [ [ "from sklearn.model_selection import train_test_split # for splitting the training and testing set\nfrom sklearn.decomposition import PCA # For possible dimentionality reduction\nfrom sklearn.feature_selection import SelectKBest # For feature selction\nfrom sklearn.model_selection import StratifiedShuffleSplit # For unbalanced class cross-validation\nfrom sklearn.preprocessing import MaxAbsScaler, StandardScaler, MinMaxScaler # Different scalars\nfrom sklearn.pipeline import Pipeline # For putting tasks in Pipeline\nfrom sklearn.model_selection import GridSearchCV # For fine tuning the classifiers", "_____no_output_____" ], [ "from sklearn.naive_bayes import BernoulliNB # For Naive Bayes\nfrom sklearn.neighbors import NearestCentroid # From modified KNN\nfrom sklearn.svm import SVC # For SVM Classifier\nfrom sklearn.tree import DecisionTreeClassifier # From decision Tree Classifier", "_____no_output_____" ], [ "X_train, X_test, y_train, y_test = train_test_split(features, labels, test_size=0.4, random_state=42)", "_____no_output_____" ], [ "from sklearn.metrics import accuracy_score", "_____no_output_____" ] ], [ [ "### Naive Bayes", "_____no_output_____" ] ], [ [ "clf = BernoulliNB()\nclf.fit(X_train, y_train)\npredictions = clf.predict(X_test)\nprint(accuracy_score(y_test, predictions, normalize=True))", "0.558867384821\n" ], [ "X_test.shape", "_____no_output_____" ] ], [ [ "### XGBoost Cassification", "_____no_output_____" ] ], [ [ "def xgb_classification(X_train, X_test, y_train, y_test, num_rounds, fnames='*'):\n if fnames == '*':\n # All the features are being used\n pass\n else:\n # Feature selection is being performed\n fnames.append('Interest_Rate')\n dataset = dtrain[fnames]\n features = np.array(dataset.iloc[:, 0:-1])\n labels = np.array(dataset.iloc[:, -1])\n X_train, X_test, y_train, y_test = train_test_split(features, labels, \n test_size=0.4, \n random_state=42)\n param = {}\n param['objective'] = 'multi:softmax'\n param['eta'] = 0.01\n param['max_depth'] = 6\n param['silent'] = 1\n param['num_class'] = 4\n param['nthread'] = -1\n param['eval_metric'] = \"merror\"\n param['min_child_weight'] = 1\n param['subsample'] = 0.8\n param['colsample_bytree'] = 0.5\n param['seed'] = 42\n xg_train = xgb.DMatrix(X_train, label=y_train)\n xg_test = xgb.DMatrix(X_test, label=y_test)\n watchlist = [(xg_train, 'train'), (xg_test, 'test')]\n bst = xgb.train(param, xg_train, num_rounds, watchlist)\n pred = bst.predict(xg_test)\n return accuracy_score(y_test, pred, normalize=True)", "_____no_output_____" ], [ "# 0.70447629480859353 <-- Boosting Rounds 1000\n# 0.70506968535086123 <--- Boosting Rounds 862\n# 0.70520662162984604 <-- Boosting Rounds 846\nxgb_classification(X_train, X_test, y_train, y_test, num_rounds=1000, fnames='*')", "[0]\ttrain-merror:0.318345\ttest-merror:0.323154\n[1]\ttrain-merror:0.309753\ttest-merror:0.31261\n[2]\ttrain-merror:0.310199\ttest-merror:0.312899\n[3]\ttrain-merror:0.304925\ttest-merror:0.306783\n[4]\ttrain-merror:0.305381\ttest-merror:0.308015\n[5]\ttrain-merror:0.303636\ttest-merror:0.307011\n[6]\ttrain-merror:0.301161\ttest-merror:0.305094\n[7]\ttrain-merror:0.301699\ttest-merror:0.306037\n[8]\ttrain-merror:0.302287\ttest-merror:0.30622\n[9]\ttrain-merror:0.303119\ttest-merror:0.307194\n[10]\ttrain-merror:0.301932\ttest-merror:0.304972\n[11]\ttrain-merror:0.30245\ttest-merror:0.305064\n[12]\ttrain-merror:0.301963\ttest-merror:0.305185\n[13]\ttrain-merror:0.302135\ttest-merror:0.305535\n[14]\ttrain-merror:0.302511\ttest-merror:0.306022\n[15]\ttrain-merror:0.301942\ttest-merror:0.305855\n[16]\ttrain-merror:0.302429\ttest-merror:0.305474\n[17]\ttrain-merror:0.301709\ttest-merror:0.30447\n[18]\ttrain-merror:0.301141\ttest-merror:0.304394\n[19]\ttrain-merror:0.301557\ttest-merror:0.304501\n[20]\ttrain-merror:0.300593\ttest-merror:0.303664\n[21]\ttrain-merror:0.299792\ttest-merror:0.30339\n[22]\ttrain-merror:0.300208\ttest-merror:0.303709\n[23]\ttrain-merror:0.299975\ttest-merror:0.303375\n[24]\ttrain-merror:0.299305\ttest-merror:0.302812\n[25]\ttrain-merror:0.29965\ttest-merror:0.303405\n[26]\ttrain-merror:0.299772\ttest-merror:0.30339\n[27]\ttrain-merror:0.299356\ttest-merror:0.302797\n[28]\ttrain-merror:0.299254\ttest-merror:0.302507\n[29]\ttrain-merror:0.298697\ttest-merror:0.30196\n[30]\ttrain-merror:0.298808\ttest-merror:0.302158\n[31]\ttrain-merror:0.29897\ttest-merror:0.302051\n[32]\ttrain-merror:0.299265\ttest-merror:0.302447\n[33]\ttrain-merror:0.299234\ttest-merror:0.302827\n[34]\ttrain-merror:0.299214\ttest-merror:0.302629\n[35]\ttrain-merror:0.299356\ttest-merror:0.302751\n[36]\ttrain-merror:0.299325\ttest-merror:0.30266\n[37]\ttrain-merror:0.299386\ttest-merror:0.302386\n[38]\ttrain-merror:0.299244\ttest-merror:0.302401\n[39]\ttrain-merror:0.299244\ttest-merror:0.302553\n[40]\ttrain-merror:0.299214\ttest-merror:0.302462\n[41]\ttrain-merror:0.299488\ttest-merror:0.302584\n[42]\ttrain-merror:0.299843\ttest-merror:0.302599\n[43]\ttrain-merror:0.300269\ttest-merror:0.302857\n[44]\ttrain-merror:0.299772\ttest-merror:0.302644\n[45]\ttrain-merror:0.29966\ttest-merror:0.302797\n[46]\ttrain-merror:0.300025\ttest-merror:0.302629\n[47]\ttrain-merror:0.300299\ttest-merror:0.302933\n[48]\ttrain-merror:0.300644\ttest-merror:0.303314\n[49]\ttrain-merror:0.300462\ttest-merror:0.302857\n[50]\ttrain-merror:0.300309\ttest-merror:0.302827\n[51]\ttrain-merror:0.300299\ttest-merror:0.302918\n[52]\ttrain-merror:0.300117\ttest-merror:0.303055\n[53]\ttrain-merror:0.300167\ttest-merror:0.303162\n[54]\ttrain-merror:0.300218\ttest-merror:0.303192\n[55]\ttrain-merror:0.299893\ttest-merror:0.303101\n[56]\ttrain-merror:0.299944\ttest-merror:0.30307\n[57]\ttrain-merror:0.299782\ttest-merror:0.302979\n[58]\ttrain-merror:0.29966\ttest-merror:0.302797\n[59]\ttrain-merror:0.29967\ttest-merror:0.30272\n[60]\ttrain-merror:0.299589\ttest-merror:0.302629\n[61]\ttrain-merror:0.299427\ttest-merror:0.302477\n[62]\ttrain-merror:0.299325\ttest-merror:0.302675\n[63]\ttrain-merror:0.299143\ttest-merror:0.302477\n[64]\ttrain-merror:0.299386\ttest-merror:0.30272\n[65]\ttrain-merror:0.299315\ttest-merror:0.30266\n[66]\ttrain-merror:0.299183\ttest-merror:0.302629\n[67]\ttrain-merror:0.299407\ttest-merror:0.302644\n[68]\ttrain-merror:0.299488\ttest-merror:0.302705\n[69]\ttrain-merror:0.299407\ttest-merror:0.302492\n[70]\ttrain-merror:0.299599\ttest-merror:0.302766\n[71]\ttrain-merror:0.29968\ttest-merror:0.302903\n[72]\ttrain-merror:0.299508\ttest-merror:0.302812\n[73]\ttrain-merror:0.29965\ttest-merror:0.302766\n[74]\ttrain-merror:0.299833\ttest-merror:0.302736\n[75]\ttrain-merror:0.29964\ttest-merror:0.302949\n[76]\ttrain-merror:0.299528\ttest-merror:0.302812\n[77]\ttrain-merror:0.29963\ttest-merror:0.302903\n[78]\ttrain-merror:0.299569\ttest-merror:0.302553\n[79]\ttrain-merror:0.299498\ttest-merror:0.30272\n[80]\ttrain-merror:0.299265\ttest-merror:0.302599\n[81]\ttrain-merror:0.299173\ttest-merror:0.302401\n[82]\ttrain-merror:0.299427\ttest-merror:0.302584\n[83]\ttrain-merror:0.299508\ttest-merror:0.302736\n[84]\ttrain-merror:0.299305\ttest-merror:0.302325\n[85]\ttrain-merror:0.299153\ttest-merror:0.302279\n[86]\ttrain-merror:0.299102\ttest-merror:0.302249\n[87]\ttrain-merror:0.299011\ttest-merror:0.30234\n[88]\ttrain-merror:0.299082\ttest-merror:0.302264\n[89]\ttrain-merror:0.299102\ttest-merror:0.302264\n[90]\ttrain-merror:0.298798\ttest-merror:0.302234\n[91]\ttrain-merror:0.298778\ttest-merror:0.302264\n[92]\ttrain-merror:0.298818\ttest-merror:0.302234\n[93]\ttrain-merror:0.298859\ttest-merror:0.302218\n[94]\ttrain-merror:0.29892\ttest-merror:0.302203\n[95]\ttrain-merror:0.29892\ttest-merror:0.302036\n[96]\ttrain-merror:0.298879\ttest-merror:0.30199\n[97]\ttrain-merror:0.298849\ttest-merror:0.301884\n[98]\ttrain-merror:0.299052\ttest-merror:0.301929\n[99]\ttrain-merror:0.299041\ttest-merror:0.302005\n[100]\ttrain-merror:0.298899\ttest-merror:0.301914\n[101]\ttrain-merror:0.29894\ttest-merror:0.301823\n[102]\ttrain-merror:0.298707\ttest-merror:0.30161\n[103]\ttrain-merror:0.298534\ttest-merror:0.301595\n[104]\ttrain-merror:0.298544\ttest-merror:0.30161\n[105]\ttrain-merror:0.298453\ttest-merror:0.301412\n[106]\ttrain-merror:0.298565\ttest-merror:0.301549\n[107]\ttrain-merror:0.298636\ttest-merror:0.301777\n[108]\ttrain-merror:0.298656\ttest-merror:0.301762\n[109]\ttrain-merror:0.298392\ttest-merror:0.301488\n[110]\ttrain-merror:0.298331\ttest-merror:0.30129\n[111]\ttrain-merror:0.2982\ttest-merror:0.301275\n[112]\ttrain-merror:0.298402\ttest-merror:0.301351\n[113]\ttrain-merror:0.29821\ttest-merror:0.301397\n[114]\ttrain-merror:0.298179\ttest-merror:0.301245\n[115]\ttrain-merror:0.29823\ttest-merror:0.301321\n[116]\ttrain-merror:0.298189\ttest-merror:0.301336\n[117]\ttrain-merror:0.297936\ttest-merror:0.301229\n[118]\ttrain-merror:0.297946\ttest-merror:0.301062\n[119]\ttrain-merror:0.297824\ttest-merror:0.301077\n[120]\ttrain-merror:0.297753\ttest-merror:0.301169\n[121]\ttrain-merror:0.297905\ttest-merror:0.301275\n[122]\ttrain-merror:0.297844\ttest-merror:0.30126\n[123]\ttrain-merror:0.297662\ttest-merror:0.301275\n[124]\ttrain-merror:0.297601\ttest-merror:0.301321\n[125]\ttrain-merror:0.297733\ttest-merror:0.301305\n[126]\ttrain-merror:0.297834\ttest-merror:0.301275\n[127]\ttrain-merror:0.298007\ttest-merror:0.301321\n[128]\ttrain-merror:0.297794\ttest-merror:0.301366\n[129]\ttrain-merror:0.297642\ttest-merror:0.301366\n[130]\ttrain-merror:0.297378\ttest-merror:0.30126\n[131]\ttrain-merror:0.29755\ttest-merror:0.301442\n[132]\ttrain-merror:0.297337\ttest-merror:0.301336\n[133]\ttrain-merror:0.297205\ttest-merror:0.301275\n[134]\ttrain-merror:0.297205\ttest-merror:0.301214\n[135]\ttrain-merror:0.297266\ttest-merror:0.30126\n[136]\ttrain-merror:0.297165\ttest-merror:0.301123\n[137]\ttrain-merror:0.297134\ttest-merror:0.301214\n[138]\ttrain-merror:0.297063\ttest-merror:0.301169\n[139]\ttrain-merror:0.297013\ttest-merror:0.300956\n[140]\ttrain-merror:0.297195\ttest-merror:0.300895\n[141]\ttrain-merror:0.297175\ttest-merror:0.300849\n[142]\ttrain-merror:0.297226\ttest-merror:0.300803\n[143]\ttrain-merror:0.297104\ttest-merror:0.300697\n[144]\ttrain-merror:0.296992\ttest-merror:0.300636\n[145]\ttrain-merror:0.297134\ttest-merror:0.300773\n[146]\ttrain-merror:0.297236\ttest-merror:0.30091\n[147]\ttrain-merror:0.297053\ttest-merror:0.300788\n[148]\ttrain-merror:0.296891\ttest-merror:0.300834\n[149]\ttrain-merror:0.296901\ttest-merror:0.300621\n[150]\ttrain-merror:0.297003\ttest-merror:0.300651\n[151]\ttrain-merror:0.296881\ttest-merror:0.300651\n[152]\ttrain-merror:0.297003\ttest-merror:0.300529\n[153]\ttrain-merror:0.296952\ttest-merror:0.300529\n[154]\ttrain-merror:0.29682\ttest-merror:0.300545\n[155]\ttrain-merror:0.296779\ttest-merror:0.300438\n[156]\ttrain-merror:0.296597\ttest-merror:0.300438\n[157]\ttrain-merror:0.296648\ttest-merror:0.300529\n[158]\ttrain-merror:0.296445\ttest-merror:0.300484\n[159]\ttrain-merror:0.296475\ttest-merror:0.300408\n[160]\ttrain-merror:0.296495\ttest-merror:0.300484\n[161]\ttrain-merror:0.296495\ttest-merror:0.300393\n[162]\ttrain-merror:0.296404\ttest-merror:0.300438\n[163]\ttrain-merror:0.296445\ttest-merror:0.300499\n[164]\ttrain-merror:0.296607\ttest-merror:0.300514\n[165]\ttrain-merror:0.296495\ttest-merror:0.300666\n[166]\ttrain-merror:0.296475\ttest-merror:0.300529\n[167]\ttrain-merror:0.296424\ttest-merror:0.300408\n[168]\ttrain-merror:0.296394\ttest-merror:0.300332\n[169]\ttrain-merror:0.296293\ttest-merror:0.300438\n[170]\ttrain-merror:0.296222\ttest-merror:0.300484\n" ] ], [ [ "## Submission", "_____no_output_____" ] ], [ [ "param = {}\nparam['objective'] = 'multi:softmax'\nparam['eta'] = 0.01\nparam['max_depth'] = 10\nparam['silent'] = 1\nparam['num_class'] = 4\nparam['nthread'] = -1\nparam['eval_metric'] = \"merror\"\nparam['min_child_weight'] = 1\nparam['subsample'] = 0.8\nparam['colsample_bytree'] = 0.5\nparam['seed'] = 42\nxg_train = xgb.DMatrix(features,label=labels)\nxg_test = xgb.DMatrix(testFeatures)\nwatchlist = [(xg_train, 'train')]\ngbm = xgb.train(param,xg_train, 846, watchlist)\ntest_pred = gbm.predict(xg_test)\ntest['Interest_Rate'] = test_pred", "[0]\ttrain-merror:0.356298\n[1]\ttrain-merror:0.313136\n[2]\ttrain-merror:0.300343\n[3]\ttrain-merror:0.298535\n[4]\ttrain-merror:0.297251\n[5]\ttrain-merror:0.296642\n[6]\ttrain-merror:0.297853\n[7]\ttrain-merror:0.296801\n[8]\ttrain-merror:0.296679\n[9]\ttrain-merror:0.295638\n[10]\ttrain-merror:0.294603\n[11]\ttrain-merror:0.294579\n[12]\ttrain-merror:0.293605\n[13]\ttrain-merror:0.292632\n[14]\ttrain-merror:0.292041\n[15]\ttrain-merror:0.291853\n[16]\ttrain-merror:0.291579\n[17]\ttrain-merror:0.291353\n[18]\ttrain-merror:0.291615\n[19]\ttrain-merror:0.291177\n[20]\ttrain-merror:0.290806\n[21]\ttrain-merror:0.290148\n[22]\ttrain-merror:0.290307\n[23]\ttrain-merror:0.290489\n[24]\ttrain-merror:0.290246\n[25]\ttrain-merror:0.290398\n[26]\ttrain-merror:0.290264\n[27]\ttrain-merror:0.290282\n[28]\ttrain-merror:0.29027\n[29]\ttrain-merror:0.290301\n[30]\ttrain-merror:0.290374\n[31]\ttrain-merror:0.290325\n[32]\ttrain-merror:0.290422\n[33]\ttrain-merror:0.290702\n[34]\ttrain-merror:0.29052\n[35]\ttrain-merror:0.290325\n[36]\ttrain-merror:0.290258\n[37]\ttrain-merror:0.290008\n[38]\ttrain-merror:0.289862\n[39]\ttrain-merror:0.290021\n[40]\ttrain-merror:0.289698\n[41]\ttrain-merror:0.28982\n[42]\ttrain-merror:0.289388\n[43]\ttrain-merror:0.289449\n[44]\ttrain-merror:0.28926\n[45]\ttrain-merror:0.289108\n[46]\ttrain-merror:0.289035\n[47]\ttrain-merror:0.289187\n[48]\ttrain-merror:0.289248\n[49]\ttrain-merror:0.289169\n[50]\ttrain-merror:0.288755\n[51]\ttrain-merror:0.288846\n[52]\ttrain-merror:0.288962\n[53]\ttrain-merror:0.289023\n[54]\ttrain-merror:0.288694\n[55]\ttrain-merror:0.288822\n[56]\ttrain-merror:0.289156\n[57]\ttrain-merror:0.289248\n[58]\ttrain-merror:0.289339\n[59]\ttrain-merror:0.289114\n[60]\ttrain-merror:0.28881\n[61]\ttrain-merror:0.28881\n[62]\ttrain-merror:0.288797\n[63]\ttrain-merror:0.288889\n[64]\ttrain-merror:0.288785\n[65]\ttrain-merror:0.288645\n[66]\ttrain-merror:0.288572\n[67]\ttrain-merror:0.288603\n[68]\ttrain-merror:0.288676\n[69]\ttrain-merror:0.288578\n[70]\ttrain-merror:0.288603\n[71]\ttrain-merror:0.288481\n[72]\ttrain-merror:0.288493\n[73]\ttrain-merror:0.288615\n[74]\ttrain-merror:0.288609\n[75]\ttrain-merror:0.288225\n[76]\ttrain-merror:0.28845\n[77]\ttrain-merror:0.288505\n[78]\ttrain-merror:0.288487\n[79]\ttrain-merror:0.288566\n[80]\ttrain-merror:0.288505\n[81]\ttrain-merror:0.288499\n[82]\ttrain-merror:0.28856\n[83]\ttrain-merror:0.288402\n[84]\ttrain-merror:0.288359\n[85]\ttrain-merror:0.288213\n[86]\ttrain-merror:0.288262\n[87]\ttrain-merror:0.288207\n[88]\ttrain-merror:0.288383\n[89]\ttrain-merror:0.28842\n[90]\ttrain-merror:0.288402\n[91]\ttrain-merror:0.288237\n[92]\ttrain-merror:0.288444\n[93]\ttrain-merror:0.28845\n[94]\ttrain-merror:0.288317\n[95]\ttrain-merror:0.28828\n[96]\ttrain-merror:0.288231\n[97]\ttrain-merror:0.28842\n[98]\ttrain-merror:0.288456\n[99]\ttrain-merror:0.288487\n[100]\ttrain-merror:0.288414\n[101]\ttrain-merror:0.288475\n[102]\ttrain-merror:0.288542\n[103]\ttrain-merror:0.288456\n[104]\ttrain-merror:0.288523\n[105]\ttrain-merror:0.28867\n[106]\ttrain-merror:0.288408\n[107]\ttrain-merror:0.288195\n[108]\ttrain-merror:0.288268\n[109]\ttrain-merror:0.288231\n[110]\ttrain-merror:0.288037\n[111]\ttrain-merror:0.288018\n[112]\ttrain-merror:0.287915\n[113]\ttrain-merror:0.28789\n[114]\ttrain-merror:0.287994\n[115]\ttrain-merror:0.287921\n[116]\ttrain-merror:0.28797\n[117]\ttrain-merror:0.287884\n[118]\ttrain-merror:0.287957\n[119]\ttrain-merror:0.287951\n[120]\ttrain-merror:0.287775\n[121]\ttrain-merror:0.28786\n[122]\ttrain-merror:0.287927\n[123]\ttrain-merror:0.288006\n[124]\ttrain-merror:0.288018\n[125]\ttrain-merror:0.288\n[126]\ttrain-merror:0.287884\n[127]\ttrain-merror:0.287884\n[128]\ttrain-merror:0.287945\n[129]\ttrain-merror:0.287763\n[130]\ttrain-merror:0.287848\n[131]\ttrain-merror:0.287903\n[132]\ttrain-merror:0.287787\n[133]\ttrain-merror:0.287641\n[134]\ttrain-merror:0.287647\n[135]\ttrain-merror:0.287623\n[136]\ttrain-merror:0.287464\n[137]\ttrain-merror:0.287556\n[138]\ttrain-merror:0.28772\n[139]\ttrain-merror:0.287635\n[140]\ttrain-merror:0.28769\n[141]\ttrain-merror:0.287659\n[142]\ttrain-merror:0.287556\n[143]\ttrain-merror:0.28744\n[144]\ttrain-merror:0.287531\n[145]\ttrain-merror:0.287458\n[146]\ttrain-merror:0.287361\n[147]\ttrain-merror:0.287331\n[148]\ttrain-merror:0.287349\n[149]\ttrain-merror:0.287343\n[150]\ttrain-merror:0.287282\n[151]\ttrain-merror:0.287197\n[152]\ttrain-merror:0.287191\n[153]\ttrain-merror:0.287142\n[154]\ttrain-merror:0.287111\n[155]\ttrain-merror:0.287032\n[156]\ttrain-merror:0.287014\n[157]\ttrain-merror:0.287008\n[158]\ttrain-merror:0.286984\n[159]\ttrain-merror:0.287099\n[160]\ttrain-merror:0.286996\n[161]\ttrain-merror:0.287136\n[162]\ttrain-merror:0.28699\n[163]\ttrain-merror:0.286965\n[164]\ttrain-merror:0.286947\n[165]\ttrain-merror:0.286905\n[166]\ttrain-merror:0.286941\n[167]\ttrain-merror:0.28685\n[168]\ttrain-merror:0.286838\n[169]\ttrain-merror:0.28671\n[170]\ttrain-merror:0.286667\n[171]\ttrain-merror:0.286728\n[172]\ttrain-merror:0.286661\n[173]\ttrain-merror:0.286667\n[174]\ttrain-merror:0.28657\n[175]\ttrain-merror:0.286594\n[176]\ttrain-merror:0.286533\n[177]\ttrain-merror:0.28646\n[178]\ttrain-merror:0.286454\n[179]\ttrain-merror:0.286472\n[180]\ttrain-merror:0.28643\n[181]\ttrain-merror:0.286357\n[182]\ttrain-merror:0.286412\n[183]\ttrain-merror:0.286472\n[184]\ttrain-merror:0.286381\n[185]\ttrain-merror:0.286381\n[186]\ttrain-merror:0.286363\n[187]\ttrain-merror:0.286296\n[188]\ttrain-merror:0.28632\n[189]\ttrain-merror:0.286247\n[190]\ttrain-merror:0.286247\n[191]\ttrain-merror:0.28629\n[192]\ttrain-merror:0.286199\n[193]\ttrain-merror:0.286101\n[194]\ttrain-merror:0.286223\n[195]\ttrain-merror:0.286211\n[196]\ttrain-merror:0.286223\n[197]\ttrain-merror:0.286089\n[198]\ttrain-merror:0.286077\n[199]\ttrain-merror:0.285961\n[200]\ttrain-merror:0.285894\n[201]\ttrain-merror:0.2859\n[202]\ttrain-merror:0.285894\n[203]\ttrain-merror:0.285803\n[204]\ttrain-merror:0.285773\n[205]\ttrain-merror:0.285718\n[206]\ttrain-merror:0.285712\n[207]\ttrain-merror:0.285633\n[208]\ttrain-merror:0.285639\n[209]\ttrain-merror:0.285639\n[210]\ttrain-merror:0.285547\n[211]\ttrain-merror:0.285523\n[212]\ttrain-merror:0.285511\n[213]\ttrain-merror:0.28545\n[214]\ttrain-merror:0.285432\n[215]\ttrain-merror:0.285401\n[216]\ttrain-merror:0.28542\n[217]\ttrain-merror:0.28545\n[218]\ttrain-merror:0.285334\n[219]\ttrain-merror:0.285377\n[220]\ttrain-merror:0.285353\n[221]\ttrain-merror:0.285231\n[222]\ttrain-merror:0.285237\n[223]\ttrain-merror:0.285237\n[224]\ttrain-merror:0.2852\n[225]\ttrain-merror:0.285261\n[226]\ttrain-merror:0.285188\n[227]\ttrain-merror:0.2852\n[228]\ttrain-merror:0.285152\n[229]\ttrain-merror:0.28514\n[230]\ttrain-merror:0.285164\n[231]\ttrain-merror:0.28506\n[232]\ttrain-merror:0.284957\n[233]\ttrain-merror:0.284939\n[234]\ttrain-merror:0.284963\n[235]\ttrain-merror:0.284878\n[236]\ttrain-merror:0.284793\n[237]\ttrain-merror:0.284799\n[238]\ttrain-merror:0.284805\n[239]\ttrain-merror:0.284768\n[240]\ttrain-merror:0.284701\n[241]\ttrain-merror:0.284781\n[242]\ttrain-merror:0.28472\n[243]\ttrain-merror:0.284659\n[244]\ttrain-merror:0.284647\n[245]\ttrain-merror:0.284598\n[246]\ttrain-merror:0.284616\n[247]\ttrain-merror:0.284549\n[248]\ttrain-merror:0.284561\n[249]\ttrain-merror:0.284555\n[250]\ttrain-merror:0.284507\n[251]\ttrain-merror:0.284428\n[252]\ttrain-merror:0.284409\n[253]\ttrain-merror:0.284391\n[254]\ttrain-merror:0.284409\n[255]\ttrain-merror:0.284397\n[256]\ttrain-merror:0.284464\n[257]\ttrain-merror:0.284354\n[258]\ttrain-merror:0.284245\n[259]\ttrain-merror:0.28419\n[260]\ttrain-merror:0.284154\n[261]\ttrain-merror:0.284166\n[262]\ttrain-merror:0.284038\n[263]\ttrain-merror:0.283995\n[264]\ttrain-merror:0.284001\n[265]\ttrain-merror:0.284014\n[266]\ttrain-merror:0.28402\n[267]\ttrain-merror:0.283935\n[268]\ttrain-merror:0.283892\n[269]\ttrain-merror:0.283862\n[270]\ttrain-merror:0.283862\n[271]\ttrain-merror:0.283776\n[272]\ttrain-merror:0.283667\n[273]\ttrain-merror:0.28363\n[274]\ttrain-merror:0.283655\n[275]\ttrain-merror:0.283533\n[276]\ttrain-merror:0.283588\n[277]\ttrain-merror:0.283551\n[278]\ttrain-merror:0.283496\n[279]\ttrain-merror:0.283423\n[280]\ttrain-merror:0.283375\n[281]\ttrain-merror:0.283356\n[282]\ttrain-merror:0.283332\n[283]\ttrain-merror:0.283332\n[284]\ttrain-merror:0.283277\n[285]\ttrain-merror:0.283314\n[286]\ttrain-merror:0.283186\n[287]\ttrain-merror:0.28318\n[288]\ttrain-merror:0.283247\n[289]\ttrain-merror:0.283125\n[290]\ttrain-merror:0.283046\n[291]\ttrain-merror:0.283028\n[292]\ttrain-merror:0.283016\n[293]\ttrain-merror:0.282991\n[294]\ttrain-merror:0.283016\n[295]\ttrain-merror:0.282955\n[296]\ttrain-merror:0.282869\n[297]\ttrain-merror:0.282863\n[298]\ttrain-merror:0.282869\n" ], [ "test['Interest_Rate'] = test.Interest_Rate.astype(int)", "_____no_output_____" ], [ "test[['Loan_ID', 'Interest_Rate']].to_csv('submission5.csv', index=False)", "_____no_output_____" ], [ "pd.read_csv('submission5.csv')\n# test[['Loan_ID', 'Interest_Rate']]", "_____no_output_____" ], [ "testFeatures.shape", "_____no_output_____" ], [ "features.shape", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code" ] ]

4ad9226e74b65b6b829198a54f3856102000512e

3,570

ipynb

Jupyter Notebook

00_core.ipynb

striderw/nbdev_sample

a8ce5118841f302fb0c3ebf2b3db5abfd1d9b27e

[ "Apache-2.0" ]

null

00_core.ipynb

striderw/nbdev_sample

a8ce5118841f302fb0c3ebf2b3db5abfd1d9b27e

[ "Apache-2.0" ]

2

2021-09-28T01:12:52.000Z

2022-02-26T06:51:23.000Z

00_core.ipynb

striderw/nbdev_sample

a8ce5118841f302fb0c3ebf2b3db5abfd1d9b27e

[ "Apache-2.0" ]

null

19.615385

175

0.445098

[ [ [ "# default_exp nbdev_sample", "_____no_output_____" ] ], [ [ "# module nbdev_sample\n\n> API details.", "_____no_output_____" ] ], [ [ "#export\ndef say_hello(to):\n \"Say Hello to Somebody\"\n return f'Hello {to}!'", "_____no_output_____" ], [ "say_hello(\"liuliu\")", "_____no_output_____" ], [ "assert say_hello('liuliu') == 'Hello liuliu!'", "_____no_output_____" ], [ "#export\ndef say_goodbye(c):\n \"Say goodbye to Somebody\"\n return f\"Goodbye {c}!\"", "_____no_output_____" ], [ "say_goodbye(\"liuliu\")", "_____no_output_____" ], [ "assert say_goodbye(\"liuliu\") == \"Goodbye liuliu!\"", "_____no_output_____" ], [ "#export\nclass HelloSayer:\n \"Say hello to `to` using `say_hello`\"\n def __init__(self, to): self.to = to\n \n def say(self):\n \"Do the saying\"\n print(say_hello(self.to))", "_____no_output_____" ], [ "from nbdev.showdoc import *\nshow_doc(HelloSayer.say)", "_____no_output_____" ], [ "o = HelloSayer(\"Alexis\")\no.say()", "Hello Alexis!\n" ] ] ]

[ "code", "markdown", "code" ]

[ [ "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad94db482580dcd2728709c312a89cf12f203a1

89,276

ipynb

Jupyter Notebook

notebooks/training_validation.ipynb

sergekomarov/object-detection-s2ds

db7c9a43881790c21d62ae553ec1c8764cb38f64

[ "MIT" ]

null

notebooks/training_validation.ipynb

sergekomarov/object-detection-s2ds

db7c9a43881790c21d62ae553ec1c8764cb38f64

[ "MIT" ]

null

notebooks/training_validation.ipynb

sergekomarov/object-detection-s2ds

db7c9a43881790c21d62ae553ec1c8764cb38f64

[ "MIT" ]

null

83.984948

1,843

0.648663

[ [ [ "# Validation notebook to accompany EPRI_ObjectDetection_Training.ipynb ", "_____no_output_____" ], [ "This notebook needs to be run in parallel with training in order to evaluate model performance on the validation data. This is done by tracking the checkpoint files saved by the training script and calculating validation metrics for each incoming checkpoint. *Needs to be run from the notebooks folder if run locally.*", "_____no_output_____" ], [ "### Specify if the model being trained is stored locally or on GCS", "_____no_output_____" ] ], [ [ "MODEL_LOCATION = 'gcs' # 'loc' or 'gcs'\nif MODEL_LOCATION == 'gcs':\n # GCS bucket (change as appropriate)\n GS_ROOT = \"gs://s2ds-aug2021-cv/\"", "_____no_output_____" ] ], [ [ "## Colab Initialization (skip if running locally)", "_____no_output_____" ] ], [ [ "from pathlib import Path\nMAIN_DIR = Path(\"/content/gdrive/MyDrive/object-detection-s2ds\")", "_____no_output_____" ] ], [ [ "Mount Google Drive", "_____no_output_____" ] ], [ [ "from google.colab import drive\ndrive.mount('/content/gdrive')", "Mounted at /content/gdrive\n" ] ], [ [ "Init Google Cloud Storage", "_____no_output_____" ] ], [ [ "if MODEL_LOCATION=='gcs':\n from google.colab import auth\n auth.authenticate_user()\n !gcloud init", "_____no_output_____" ] ], [ [ "Update Python PATH with the previously installed (to Google Drive) packages and go to the notebooks directory", "_____no_output_____" ] ], [ [ "%cd {MAIN_DIR}\nfrom src.utils import add_pkgs_to_path\nadd_pkgs_to_path(str(MAIN_DIR/\"colab-install/lib/python3.7/site-packages\"))\n%cd notebooks", "/content/gdrive/MyDrive/s2ds-aug2021-cv\n/content/gdrive/My Drive/s2ds-aug2021-cv/notebooks\n" ] ], [ [ "## Specify the model, set paths", "_____no_output_____" ] ], [ [ "import os\nNOTEBOOK_DIR = Path(os.getcwd().replace('My Drive','MyDrive'))\nMAIN_DIR = NOTEBOOK_DIR.parent\n\n# change the Object Detection API location as appropriate\nTF_DIR = os.path.join(MAIN_DIR, \"tf-models\")\n\n# specify the name of the model being trained\nMY_MODEL_NAME = \"efficient_det_1024_rand_aug_1_4_demo\"\n\n# specify the model's directory\nif MODEL_LOCATION=='loc': \n my_model_dir = os.path.join(MAIN_DIR, \"models/fine-tuned\", MY_MODEL_NAME)\nelif MODEL_LOCATION=='gcs':\n my_model_dir = os.path.join(GS_ROOT, \"models/fine-tuned\", MY_MODEL_NAME)\nconfig_path = os.path.join(my_model_dir, \"pipeline.config\")", "_____no_output_____" ] ], [ [ "## Launch validation", "_____no_output_____" ] ], [ [ "%cd {os.path.join(TF_DIR, \"research/object_detection\")}\n!python model_main_tf2.py \\\n--model_dir=$my_model_dir \\\n--pipeline_config_path=$config_path \\\n--checkpoint_dir=$my_model_dir\n%cd {NOTEBOOK_DIR}", "/content/gdrive/MyDrive/s2ds-aug2021-cv/tf-models/research/object_detection\n2021-09-20 22:21:19.862618: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2021-09-20 22:21:20.201066: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2021-09-20 22:21:20.201749: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\nWARNING:tensorflow:Forced number of epochs for all eval validations to be 1.\nW0920 22:21:21.608826 139884754093952 model_lib_v2.py:1082] Forced number of epochs for all eval validations to be 1.\nINFO:tensorflow:Maybe overwriting sample_1_of_n_eval_examples: None\nI0920 22:21:21.609061 139884754093952 config_util.py:552] Maybe overwriting sample_1_of_n_eval_examples: None\nINFO:tensorflow:Maybe overwriting use_bfloat16: False\nI0920 22:21:21.609152 139884754093952 config_util.py:552] Maybe overwriting use_bfloat16: False\nINFO:tensorflow:Maybe overwriting eval_num_epochs: 1\nI0920 22:21:21.609234 139884754093952 config_util.py:552] Maybe overwriting eval_num_epochs: 1\nWARNING:tensorflow:Expected number of evaluation epochs is 1, but instead encountered `eval_on_train_input_config.num_epochs` = 0. Overwriting `num_epochs` to 1.\nW0920 22:21:21.609369 139884754093952 model_lib_v2.py:1103] Expected number of evaluation epochs is 1, but instead encountered `eval_on_train_input_config.num_epochs` = 0. Overwriting `num_epochs` to 1.\n2021-09-20 22:21:21.618617: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2021-09-20 22:21:21.619287: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2021-09-20 22:21:21.619855: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2021-09-20 22:21:24.544823: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2021-09-20 22:21:24.545523: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2021-09-20 22:21:24.546102: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:937] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero\n2021-09-20 22:21:24.546710: W tensorflow/core/common_runtime/gpu/gpu_bfc_allocator.cc:39] Overriding allow_growth setting because the TF_FORCE_GPU_ALLOW_GROWTH environment variable is set. Original config value was 0.\n2021-09-20 22:21:24.546777: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1510] Created device /job:localhost/replica:0/task:0/device:GPU:0 with 13839 MB memory: -> device: 0, name: Tesla T4, pci bus id: 0000:00:04.0, compute capability: 7.5\nI0920 22:21:24.573896 139884754093952 ssd_efficientnet_bifpn_feature_extractor.py:143] EfficientDet EfficientNet backbone version: efficientnet-b4\nI0920 22:21:24.574060 139884754093952 ssd_efficientnet_bifpn_feature_extractor.py:144] EfficientDet BiFPN num filters: 224\nI0920 22:21:24.574142 139884754093952 ssd_efficientnet_bifpn_feature_extractor.py:146] EfficientDet BiFPN num iterations: 7\nI0920 22:21:24.577589 139884754093952 efficientnet_model.py:147] round_filter input=32 output=48\nI0920 22:21:24.703068 139884754093952 efficientnet_model.py:147] round_filter input=32 output=48\nI0920 22:21:24.703257 139884754093952 efficientnet_model.py:147] round_filter input=16 output=24\nI0920 22:21:24.826859 139884754093952 efficientnet_model.py:147] round_filter input=16 output=24\nI0920 22:21:24.827011 139884754093952 efficientnet_model.py:147] round_filter input=24 output=32\nI0920 22:21:25.112632 139884754093952 efficientnet_model.py:147] round_filter input=24 output=32\nI0920 22:21:25.112778 139884754093952 efficientnet_model.py:147] round_filter input=40 output=56\nI0920 22:21:25.367774 139884754093952 efficientnet_model.py:147] round_filter input=40 output=56\nI0920 22:21:25.367927 139884754093952 efficientnet_model.py:147] round_filter input=80 output=112\nI0920 22:21:25.754755 139884754093952 efficientnet_model.py:147] round_filter input=80 output=112\nI0920 22:21:25.754912 139884754093952 efficientnet_model.py:147] round_filter input=112 output=160\nI0920 22:21:26.148336 139884754093952 efficientnet_model.py:147] round_filter input=112 output=160\nI0920 22:21:26.148493 139884754093952 efficientnet_model.py:147] round_filter input=192 output=272\nI0920 22:21:26.671197 139884754093952 efficientnet_model.py:147] round_filter input=192 output=272\nI0920 22:21:26.671374 139884754093952 efficientnet_model.py:147] round_filter input=320 output=448\nI0920 22:21:26.799449 139884754093952 efficientnet_model.py:147] round_filter input=1280 output=1792\nI0920 22:21:26.831659 139884754093952 efficientnet_model.py:458] Building model efficientnet with params ModelConfig(width_coefficient=1.4, depth_coefficient=1.8, resolution=380, dropout_rate=0.4, blocks=(BlockConfig(input_filters=32, output_filters=16, kernel_size=3, num_repeat=1, expand_ratio=1, strides=(1, 1), se_ratio=0.25, id_skip=True, fused_conv=False, conv_type='depthwise'), BlockConfig(input_filters=16, output_filters=24, kernel_size=3, num_repeat=2, expand_ratio=6, strides=(2, 2), se_ratio=0.25, id_skip=True, fused_conv=False, conv_type='depthwise'), BlockConfig(input_filters=24, output_filters=40, kernel_size=5, num_repeat=2, expand_ratio=6, strides=(2, 2), se_ratio=0.25, id_skip=True, fused_conv=False, conv_type='depthwise'), BlockConfig(input_filters=40, output_filters=80, kernel_size=3, num_repeat=3, expand_ratio=6, strides=(2, 2), se_ratio=0.25, id_skip=True, fused_conv=False, conv_type='depthwise'), BlockConfig(input_filters=80, output_filters=112, kernel_size=5, num_repeat=3, expand_ratio=6, strides=(1, 1), se_ratio=0.25, id_skip=True, fused_conv=False, conv_type='depthwise'), BlockConfig(input_filters=112, output_filters=192, kernel_size=5, num_repeat=4, expand_ratio=6, strides=(2, 2), se_ratio=0.25, id_skip=True, fused_conv=False, conv_type='depthwise'), BlockConfig(input_filters=192, output_filters=320, kernel_size=3, num_repeat=1, expand_ratio=6, strides=(1, 1), se_ratio=0.25, id_skip=True, fused_conv=False, conv_type='depthwise')), stem_base_filters=32, top_base_filters=1280, activation='simple_swish', batch_norm='default', bn_momentum=0.99, bn_epsilon=0.001, weight_decay=5e-06, drop_connect_rate=0.2, depth_divisor=8, min_depth=None, use_se=True, input_channels=3, num_classes=1000, model_name='efficientnet', rescale_input=False, data_format='channels_last', dtype='float32')\nINFO:tensorflow:Reading unweighted datasets: ['gs://s2ds-aug2021-cv/data/train-valid-split/valid.record']\nI0920 22:21:27.670289 139884754093952 dataset_builder.py:163] Reading unweighted datasets: ['gs://s2ds-aug2021-cv/data/train-valid-split/valid.record']\nINFO:tensorflow:Reading record datasets for input file: ['gs://s2ds-aug2021-cv/data/train-valid-split/valid.record']\nI0920 22:21:27.788917 139884754093952 dataset_builder.py:80] Reading record datasets for input file: ['gs://s2ds-aug2021-cv/data/train-valid-split/valid.record']\nINFO:tensorflow:Number of filenames to read: 1\nI0920 22:21:27.789126 139884754093952 dataset_builder.py:81] Number of filenames to read: 1\nWARNING:tensorflow:num_readers has been reduced to 1 to match input file shards.\nW0920 22:21:27.789216 139884754093952 dataset_builder.py:88] num_readers has been reduced to 1 to match input file shards.\nWARNING:tensorflow:From /content/gdrive/MyDrive/s2ds-aug2021-cv/colab-install/lib/python3.7/site-packages/object_detection-0.1-py3.7.egg/object_detection/builders/dataset_builder.py:105: parallel_interleave (from tensorflow.python.data.experimental.ops.interleave_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.data.Dataset.interleave(map_func, cycle_length, block_length, num_parallel_calls=tf.data.AUTOTUNE)` instead. If sloppy execution is desired, use `tf.data.Options.experimental_deterministic`.\nW0920 22:21:27.796844 139884754093952 deprecation.py:345] From /content/gdrive/MyDrive/s2ds-aug2021-cv/colab-install/lib/python3.7/site-packages/object_detection-0.1-py3.7.egg/object_detection/builders/dataset_builder.py:105: parallel_interleave (from tensorflow.python.data.experimental.ops.interleave_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.data.Dataset.interleave(map_func, cycle_length, block_length, num_parallel_calls=tf.data.AUTOTUNE)` instead. If sloppy execution is desired, use `tf.data.Options.experimental_deterministic`.\nWARNING:tensorflow:From /content/gdrive/MyDrive/s2ds-aug2021-cv/colab-install/lib/python3.7/site-packages/object_detection-0.1-py3.7.egg/object_detection/builders/dataset_builder.py:237: DatasetV1.map_with_legacy_function (from tensorflow.python.data.ops.dataset_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.data.Dataset.map()\nW0920 22:21:27.824324 139884754093952 deprecation.py:345] From /content/gdrive/MyDrive/s2ds-aug2021-cv/colab-install/lib/python3.7/site-packages/object_detection-0.1-py3.7.egg/object_detection/builders/dataset_builder.py:237: DatasetV1.map_with_legacy_function (from tensorflow.python.data.ops.dataset_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.data.Dataset.map()\nWARNING:tensorflow:From /usr/local/lib/python3.7/dist-packages/tensorflow/python/util/dispatch.py:206: sparse_to_dense (from tensorflow.python.ops.sparse_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nCreate a `tf.sparse.SparseTensor` and use `tf.sparse.to_dense` instead.\nW0920 22:21:32.332355 139884754093952 deprecation.py:345] From /usr/local/lib/python3.7/dist-packages/tensorflow/python/util/dispatch.py:206: sparse_to_dense (from tensorflow.python.ops.sparse_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nCreate a `tf.sparse.SparseTensor` and use `tf.sparse.to_dense` instead.\nWARNING:tensorflow:From /usr/local/lib/python3.7/dist-packages/tensorflow/python/autograph/impl/api.py:464: to_float (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.cast` instead.\nW0920 22:21:34.361688 139884754093952 deprecation.py:345] From /usr/local/lib/python3.7/dist-packages/tensorflow/python/autograph/impl/api.py:464: to_float (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.cast` instead.\nINFO:tensorflow:Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nI0920 22:21:37.577253 139884754093952 checkpoint_utils.py:140] Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nINFO:tensorflow:Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-1\nI0920 22:23:02.372278 139884754093952 checkpoint_utils.py:149] Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-1\n/usr/local/lib/python3.7/dist-packages/keras/backend.py:401: UserWarning: `tf.keras.backend.set_learning_phase` is deprecated and will be removed after 2020-10-11. To update it, simply pass a True/False value to the `training` argument of the `__call__` method of your layer or model.\n warnings.warn('`tf.keras.backend.set_learning_phase` is deprecated and '\n2021-09-20 22:23:22.169929: I tensorflow/compiler/mlir/mlir_graph_optimization_pass.cc:185] None of the MLIR Optimization Passes are enabled (registered 2)\n2021-09-20 22:31:01.494910: I tensorflow/stream_executor/cuda/cuda_dnn.cc:369] Loaded cuDNN version 8005\nWARNING:tensorflow:From /content/gdrive/MyDrive/s2ds-aug2021-cv/colab-install/lib/python3.7/site-packages/object_detection-0.1-py3.7.egg/object_detection/eval_util.py:929: to_int64 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.cast` instead.\nW0920 22:31:18.431624 139884754093952 deprecation.py:345] From /content/gdrive/MyDrive/s2ds-aug2021-cv/colab-install/lib/python3.7/site-packages/object_detection-0.1-py3.7.egg/object_detection/eval_util.py:929: to_int64 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\nUse `tf.cast` instead.\nINFO:tensorflow:Finished eval step 0\nI0920 22:31:18.442065 139884754093952 model_lib_v2.py:958] Finished eval step 0\nWARNING:tensorflow:From /usr/local/lib/python3.7/dist-packages/tensorflow/python/autograph/impl/api.py:464: py_func (from tensorflow.python.ops.script_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\ntf.py_func is deprecated in TF V2. Instead, there are two\n options available in V2.\n - tf.py_function takes a python function which manipulates tf eager\n tensors instead of numpy arrays. It's easy to convert a tf eager tensor to\n an ndarray (just call tensor.numpy()) but having access to eager tensors\n means `tf.py_function`s can use accelerators such as GPUs as well as\n being differentiable using a gradient tape.\n - tf.numpy_function maintains the semantics of the deprecated tf.py_func\n (it is not differentiable, and manipulates numpy arrays). It drops the\n stateful argument making all functions stateful.\n \nW0920 22:31:18.704968 139884754093952 deprecation.py:345] From /usr/local/lib/python3.7/dist-packages/tensorflow/python/autograph/impl/api.py:464: py_func (from tensorflow.python.ops.script_ops) is deprecated and will be removed in a future version.\nInstructions for updating:\ntf.py_func is deprecated in TF V2. Instead, there are two\n options available in V2.\n - tf.py_function takes a python function which manipulates tf eager\n tensors instead of numpy arrays. It's easy to convert a tf eager tensor to\n an ndarray (just call tensor.numpy()) but having access to eager tensors\n means `tf.py_function`s can use accelerators such as GPUs as well as\n being differentiable using a gradient tape.\n - tf.numpy_function maintains the semantics of the deprecated tf.py_func\n (it is not differentiable, and manipulates numpy arrays). It drops the\n stateful argument making all functions stateful.\n \nINFO:tensorflow:Performing evaluation on 57 images.\nI0920 22:31:44.330371 139884754093952 coco_evaluation.py:293] Performing evaluation on 57 images.\ncreating index...\nindex created!\nINFO:tensorflow:Loading and preparing annotation results...\nI0920 22:31:44.330815 139884754093952 coco_tools.py:116] Loading and preparing annotation results...\nINFO:tensorflow:DONE (t=0.00s)\nI0920 22:31:44.333792 139884754093952 coco_tools.py:138] DONE (t=0.00s)\ncreating index...\nindex created!\nRunning per image evaluation...\nEvaluate annotation type *bbox*\nDONE (t=0.30s).\nAccumulating evaluation results...\nDONE (t=0.04s).\n Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.000\n Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.000\n Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.000\n Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000\n Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.000\nINFO:tensorflow:Eval metrics at step 0\nI0920 22:31:44.674082 139884754093952 model_lib_v2.py:1007] Eval metrics at step 0\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP: 0.000000\nI0920 22:31:44.676001 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP: 0.000000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.000000\nI0920 22:31:45.661247 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.000000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.000000\nI0920 22:31:45.662970 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.000000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (small): -1.000000\nI0920 22:31:45.665035 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (medium): 0.000000\nI0920 22:31:45.666218 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (medium): 0.000000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (large): 0.000000\nI0920 22:31:45.667353 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (large): 0.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@1: 0.000000\nI0920 22:31:45.668439 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@1: 0.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@10: 0.000000\nI0920 22:31:45.669486 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@10: 0.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100: 0.000000\nI0920 22:31:45.670572 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100: 0.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nI0920 22:31:45.671632 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (medium): 0.000000\nI0920 22:31:45.672683 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (medium): 0.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (large): 0.000000\nI0920 22:31:45.673737 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (large): 0.000000\nINFO:tensorflow:\t+ Loss/localization_loss: 0.023383\nI0920 22:31:46.661346 139884754093952 model_lib_v2.py:1010] \t+ Loss/localization_loss: 0.023383\nINFO:tensorflow:\t+ Loss/classification_loss: 1.184951\nI0920 22:31:46.662968 139884754093952 model_lib_v2.py:1010] \t+ Loss/classification_loss: 1.184951\nINFO:tensorflow:\t+ Loss/regularization_loss: 0.048917\nI0920 22:31:46.664098 139884754093952 model_lib_v2.py:1010] \t+ Loss/regularization_loss: 0.048917\nINFO:tensorflow:\t+ Loss/total_loss: 1.257250\nI0920 22:31:46.665160 139884754093952 model_lib_v2.py:1010] \t+ Loss/total_loss: 1.257250\nINFO:tensorflow:Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nI0920 22:31:47.955268 139884754093952 checkpoint_utils.py:140] Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nWARNING:tensorflow:ParseError: 5:1 : Message type \"tensorflow.CheckpointState\" has no field named \"all_mode\".\nW0920 22:40:05.656394 139884754093952 checkpoint_management.py:300] ParseError: 5:1 : Message type \"tensorflow.CheckpointState\" has no field named \"all_mode\".\nWARNING:tensorflow:gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/checkpoint: Checkpoint ignored\nW0920 22:40:05.656609 139884754093952 checkpoint_management.py:301] gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/checkpoint: Checkpoint ignored\nWARNING:tensorflow:ParseError: 5:1 : Message type \"tensorflow.CheckpointState\" has no field named \"all_mode\".\nW0920 22:40:06.864014 139884754093952 checkpoint_management.py:300] ParseError: 5:1 : Message type \"tensorflow.CheckpointState\" has no field named \"all_mode\".\nWARNING:tensorflow:gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/checkpoint: Checkpoint ignored\nW0920 22:40:06.864216 139884754093952 checkpoint_management.py:301] gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/checkpoint: Checkpoint ignored\nWARNING:tensorflow:ParseError: 5:1 : Message type \"tensorflow.CheckpointState\" has no field named \"all_mode\".\nW0920 22:40:07.982741 139884754093952 checkpoint_management.py:300] ParseError: 5:1 : Message type \"tensorflow.CheckpointState\" has no field named \"all_mode\".\nWARNING:tensorflow:gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/checkpoint: Checkpoint ignored\nW0920 22:40:07.982934 139884754093952 checkpoint_management.py:301] gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/checkpoint: Checkpoint ignored\nWARNING:tensorflow:ParseError: 5:1 : Message type \"tensorflow.CheckpointState\" has no field named \"all_mode\".\nW0920 22:40:09.103401 139884754093952 checkpoint_management.py:300] ParseError: 5:1 : Message type \"tensorflow.CheckpointState\" has no field named \"all_mode\".\nWARNING:tensorflow:gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/checkpoint: Checkpoint ignored\nW0920 22:40:09.103611 139884754093952 checkpoint_management.py:301] gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/checkpoint: Checkpoint ignored\nINFO:tensorflow:Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-2\nI0920 22:40:10.441192 139884754093952 checkpoint_utils.py:149] Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-2\nINFO:tensorflow:Finished eval step 0\nI0920 22:49:58.894298 139884754093952 model_lib_v2.py:958] Finished eval step 0\nINFO:tensorflow:Performing evaluation on 57 images.\nI0920 22:50:29.609162 139884754093952 coco_evaluation.py:293] Performing evaluation on 57 images.\ncreating index...\nindex created!\nINFO:tensorflow:Loading and preparing annotation results...\nI0920 22:50:29.609595 139884754093952 coco_tools.py:116] Loading and preparing annotation results...\nINFO:tensorflow:DONE (t=0.00s)\nI0920 22:50:29.612813 139884754093952 coco_tools.py:138] DONE (t=0.00s)\ncreating index...\nindex created!\nRunning per image evaluation...\nEvaluate annotation type *bbox*\nDONE (t=0.29s).\nAccumulating evaluation results...\nDONE (t=0.04s).\n Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.141\n Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.548\n Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.029\n Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.190\n Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.222\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.188\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.305\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.339\n Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.333\n Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.500\nINFO:tensorflow:Eval metrics at step 500\nI0920 22:50:29.953565 139884754093952 model_lib_v2.py:1007] Eval metrics at step 500\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP: 0.140818\nI0920 22:50:30.952851 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP: 0.140818\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.547543\nI0920 22:50:30.954782 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.547543\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.028814\nI0920 22:50:30.956136 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.028814\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (small): -1.000000\nI0920 22:50:30.957229 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (medium): 0.190255\nI0920 22:50:30.958588 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (medium): 0.190255\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (large): 0.222082\nI0920 22:50:30.960462 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (large): 0.222082\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@1: 0.188299\nI0920 22:50:30.961747 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@1: 0.188299\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@10: 0.304899\nI0920 22:50:30.963006 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@10: 0.304899\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100: 0.338729\nI0920 22:50:30.964266 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100: 0.338729\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nI0920 22:50:30.965368 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (medium): 0.332585\nI0920 22:50:30.966652 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (medium): 0.332585\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (large): 0.500000\nI0920 22:50:31.909376 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (large): 0.500000\nINFO:tensorflow:\t+ Loss/localization_loss: 0.021594\nI0920 22:50:31.910993 139884754093952 model_lib_v2.py:1010] \t+ Loss/localization_loss: 0.021594\nINFO:tensorflow:\t+ Loss/classification_loss: 1.162642\nI0920 22:50:31.912172 139884754093952 model_lib_v2.py:1010] \t+ Loss/classification_loss: 1.162642\nINFO:tensorflow:\t+ Loss/regularization_loss: 0.048891\nI0920 22:50:31.913213 139884754093952 model_lib_v2.py:1010] \t+ Loss/regularization_loss: 0.048891\nINFO:tensorflow:\t+ Loss/total_loss: 1.233127\nI0920 22:50:31.914251 139884754093952 model_lib_v2.py:1010] \t+ Loss/total_loss: 1.233127\nINFO:tensorflow:Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nI0920 22:50:33.088662 139884754093952 checkpoint_utils.py:140] Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nINFO:tensorflow:Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-3\nI0920 22:50:33.616855 139884754093952 checkpoint_utils.py:149] Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-3\nINFO:tensorflow:Finished eval step 0\nI0920 23:00:30.958849 139884754093952 model_lib_v2.py:958] Finished eval step 0\nINFO:tensorflow:Performing evaluation on 57 images.\nI0920 23:01:01.103723 139884754093952 coco_evaluation.py:293] Performing evaluation on 57 images.\ncreating index...\nindex created!\nINFO:tensorflow:Loading and preparing annotation results...\nI0920 23:01:01.104118 139884754093952 coco_tools.py:116] Loading and preparing annotation results...\nINFO:tensorflow:DONE (t=0.00s)\nI0920 23:01:01.107059 139884754093952 coco_tools.py:138] DONE (t=0.00s)\ncreating index...\nindex created!\nRunning per image evaluation...\nEvaluate annotation type *bbox*\nDONE (t=0.28s).\nAccumulating evaluation results...\nDONE (t=0.05s).\n Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.424\n Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.764\n Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.431\n Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.428\n Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.433\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.465\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.497\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.515\n Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.515\n Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.525\nINFO:tensorflow:Eval metrics at step 1000\nI0920 23:01:01.449106 139884754093952 model_lib_v2.py:1007] Eval metrics at step 1000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP: 0.424283\nI0920 23:01:02.433341 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP: 0.424283\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.764151\nI0920 23:01:02.435082 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.764151\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.431268\nI0920 23:01:02.436369 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.431268\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (small): -1.000000\nI0920 23:01:02.437306 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (medium): 0.428042\nI0920 23:01:02.438416 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (medium): 0.428042\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (large): 0.433416\nI0920 23:01:02.439452 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (large): 0.433416\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@1: 0.465293\nI0920 23:01:02.440513 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@1: 0.465293\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@10: 0.496706\nI0920 23:01:02.441569 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@10: 0.496706\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100: 0.515376\nI0920 23:01:02.442606 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100: 0.515376\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nI0920 23:01:02.443517 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (medium): 0.514973\nI0920 23:01:02.444573 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (medium): 0.514973\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (large): 0.525000\nI0920 23:01:03.429160 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (large): 0.525000\nINFO:tensorflow:\t+ Loss/localization_loss: 0.008759\nI0920 23:01:03.430793 139884754093952 model_lib_v2.py:1010] \t+ Loss/localization_loss: 0.008759\nINFO:tensorflow:\t+ Loss/classification_loss: 0.492439\nI0920 23:01:03.431900 139884754093952 model_lib_v2.py:1010] \t+ Loss/classification_loss: 0.492439\nINFO:tensorflow:\t+ Loss/regularization_loss: 0.048829\nI0920 23:01:03.432928 139884754093952 model_lib_v2.py:1010] \t+ Loss/regularization_loss: 0.048829\nINFO:tensorflow:\t+ Loss/total_loss: 0.550027\nI0920 23:01:03.433842 139884754093952 model_lib_v2.py:1010] \t+ Loss/total_loss: 0.550027\nINFO:tensorflow:Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nI0920 23:01:04.604441 139884754093952 checkpoint_utils.py:140] Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nINFO:tensorflow:Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-4\nI0920 23:01:05.132462 139884754093952 checkpoint_utils.py:149] Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-4\nINFO:tensorflow:Finished eval step 0\nI0920 23:11:05.635410 139884754093952 model_lib_v2.py:958] Finished eval step 0\nINFO:tensorflow:Performing evaluation on 57 images.\nI0920 23:11:36.165417 139884754093952 coco_evaluation.py:293] Performing evaluation on 57 images.\ncreating index...\nindex created!\nINFO:tensorflow:Loading and preparing annotation results...\nI0920 23:11:36.165894 139884754093952 coco_tools.py:116] Loading and preparing annotation results...\nINFO:tensorflow:DONE (t=0.00s)\nI0920 23:11:36.168855 139884754093952 coco_tools.py:138] DONE (t=0.00s)\ncreating index...\nindex created!\nRunning per image evaluation...\nEvaluate annotation type *bbox*\nDONE (t=0.27s).\nAccumulating evaluation results...\nDONE (t=0.04s).\n Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.450\n Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.781\n Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.441\n Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.449\n Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.522\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.494\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.516\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.533\n Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.531\n Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.600\nINFO:tensorflow:Eval metrics at step 1500\nI0920 23:11:36.493798 139884754093952 model_lib_v2.py:1007] Eval metrics at step 1500\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP: 0.450424\nI0920 23:11:37.472051 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP: 0.450424\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.780831\nI0920 23:11:37.473753 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.780831\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.441213\nI0920 23:11:37.474962 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.441213\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (small): -1.000000\nI0920 23:11:37.475950 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (medium): 0.448881\nI0920 23:11:37.477076 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (medium): 0.448881\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (large): 0.521617\nI0920 23:11:37.478143 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (large): 0.521617\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@1: 0.494025\nI0920 23:11:37.479182 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@1: 0.494025\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@10: 0.516269\nI0920 23:11:37.480274 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@10: 0.516269\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100: 0.533251\nI0920 23:11:37.481346 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100: 0.533251\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nI0920 23:11:37.482259 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (medium): 0.530713\nI0920 23:11:37.483349 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (medium): 0.530713\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (large): 0.600000\nI0920 23:11:38.518153 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (large): 0.600000\nINFO:tensorflow:\t+ Loss/localization_loss: 0.007508\nI0920 23:11:38.519530 139884754093952 model_lib_v2.py:1010] \t+ Loss/localization_loss: 0.007508\nINFO:tensorflow:\t+ Loss/classification_loss: 0.476578\nI0920 23:11:38.520553 139884754093952 model_lib_v2.py:1010] \t+ Loss/classification_loss: 0.476578\nINFO:tensorflow:\t+ Loss/regularization_loss: 0.048768\nI0920 23:11:38.521472 139884754093952 model_lib_v2.py:1010] \t+ Loss/regularization_loss: 0.048768\nINFO:tensorflow:\t+ Loss/total_loss: 0.532855\nI0920 23:11:38.522381 139884754093952 model_lib_v2.py:1010] \t+ Loss/total_loss: 0.532855\nINFO:tensorflow:Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nI0920 23:11:39.687109 139884754093952 checkpoint_utils.py:140] Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nINFO:tensorflow:Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-6\nI0920 23:11:40.305709 139884754093952 checkpoint_utils.py:149] Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-6\nINFO:tensorflow:Finished eval step 0\nI0920 23:21:48.176845 139884754093952 model_lib_v2.py:958] Finished eval step 0\nINFO:tensorflow:Performing evaluation on 57 images.\nI0920 23:22:18.883420 139884754093952 coco_evaluation.py:293] Performing evaluation on 57 images.\ncreating index...\nindex created!\nINFO:tensorflow:Loading and preparing annotation results...\nI0920 23:22:18.883823 139884754093952 coco_tools.py:116] Loading and preparing annotation results...\nINFO:tensorflow:DONE (t=0.00s)\nI0920 23:22:18.886845 139884754093952 coco_tools.py:138] DONE (t=0.00s)\ncreating index...\nindex created!\nRunning per image evaluation...\nEvaluate annotation type *bbox*\nDONE (t=0.28s).\nAccumulating evaluation results...\nDONE (t=0.05s).\n Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.499\n Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.776\n Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.648\n Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.496\n Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.618\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.540\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.555\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.570\n Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.566\n Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.675\nINFO:tensorflow:Eval metrics at step 2500\nI0920 23:22:19.224351 139884754093952 model_lib_v2.py:1007] Eval metrics at step 2500\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP: 0.498654\nI0920 23:22:20.200912 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP: 0.498654\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.775729\nI0920 23:22:20.202599 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.775729\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.648030\nI0920 23:22:20.203763 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.648030\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (small): -1.000000\nI0920 23:22:20.204692 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (medium): 0.495650\nI0920 23:22:20.205787 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (medium): 0.495650\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (large): 0.617987\nI0920 23:22:20.206829 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (large): 0.617987\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@1: 0.539540\nI0920 23:22:20.207907 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@1: 0.539540\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@10: 0.555180\nI0920 23:22:20.208999 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@10: 0.555180\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100: 0.570275\nI0920 23:22:20.210047 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100: 0.570275\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nI0920 23:22:20.210967 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (medium): 0.566257\nI0920 23:22:20.212029 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (medium): 0.566257\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (large): 0.675000\nI0920 23:22:21.161121 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (large): 0.675000\nINFO:tensorflow:\t+ Loss/localization_loss: 0.006665\nI0920 23:22:21.162538 139884754093952 model_lib_v2.py:1010] \t+ Loss/localization_loss: 0.006665\nINFO:tensorflow:\t+ Loss/classification_loss: 0.507661\nI0920 23:22:21.163582 139884754093952 model_lib_v2.py:1010] \t+ Loss/classification_loss: 0.507661\nINFO:tensorflow:\t+ Loss/regularization_loss: 0.048663\nI0920 23:22:21.164533 139884754093952 model_lib_v2.py:1010] \t+ Loss/regularization_loss: 0.048663\nINFO:tensorflow:\t+ Loss/total_loss: 0.562989\nI0920 23:22:21.165450 139884754093952 model_lib_v2.py:1010] \t+ Loss/total_loss: 0.562989\nINFO:tensorflow:Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nI0920 23:22:22.333876 139884754093952 checkpoint_utils.py:140] Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nINFO:tensorflow:Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-7\nI0920 23:22:22.883119 139884754093952 checkpoint_utils.py:149] Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-7\nINFO:tensorflow:Finished eval step 0\nI0920 23:32:17.724228 139884754093952 model_lib_v2.py:958] Finished eval step 0\nINFO:tensorflow:Performing evaluation on 57 images.\nI0920 23:32:48.369152 139884754093952 coco_evaluation.py:293] Performing evaluation on 57 images.\ncreating index...\nindex created!\nINFO:tensorflow:Loading and preparing annotation results...\nI0920 23:32:48.369641 139884754093952 coco_tools.py:116] Loading and preparing annotation results...\nINFO:tensorflow:DONE (t=0.00s)\nI0920 23:32:48.372298 139884754093952 coco_tools.py:138] DONE (t=0.00s)\ncreating index...\nindex created!\nRunning per image evaluation...\nEvaluate annotation type *bbox*\nDONE (t=0.28s).\nAccumulating evaluation results...\nDONE (t=0.04s).\n Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.495\n Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.781\n Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.580\n Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.489\n Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.742\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.532\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.550\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.565\n Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.558\n Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.750\nINFO:tensorflow:Eval metrics at step 3000\nI0920 23:32:48.702328 139884754093952 model_lib_v2.py:1007] Eval metrics at step 3000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP: 0.495460\nI0920 23:32:49.773900 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP: 0.495460\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.781026\nI0920 23:32:49.775651 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.781026\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.580411\nI0920 23:32:49.776813 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.580411\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (small): -1.000000\nI0920 23:32:49.777751 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (medium): 0.488668\nI0920 23:32:49.778860 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (medium): 0.488668\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (large): 0.742244\nI0920 23:32:49.779949 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (large): 0.742244\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@1: 0.531662\nI0920 23:32:49.781088 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@1: 0.531662\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@10: 0.550132\nI0920 23:32:49.782185 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@10: 0.550132\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100: 0.565227\nI0920 23:32:49.783271 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100: 0.565227\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nI0920 23:32:49.784218 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (medium): 0.558128\nI0920 23:32:49.785280 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (medium): 0.558128\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (large): 0.750000\nI0920 23:32:50.734485 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (large): 0.750000\nINFO:tensorflow:\t+ Loss/localization_loss: 0.006609\nI0920 23:32:50.735879 139884754093952 model_lib_v2.py:1010] \t+ Loss/localization_loss: 0.006609\nINFO:tensorflow:\t+ Loss/classification_loss: 0.456218\nI0920 23:32:50.736867 139884754093952 model_lib_v2.py:1010] \t+ Loss/classification_loss: 0.456218\nINFO:tensorflow:\t+ Loss/regularization_loss: 0.048622\nI0920 23:32:50.737790 139884754093952 model_lib_v2.py:1010] \t+ Loss/regularization_loss: 0.048622\nINFO:tensorflow:\t+ Loss/total_loss: 0.511449\nI0920 23:32:50.738688 139884754093952 model_lib_v2.py:1010] \t+ Loss/total_loss: 0.511449\nINFO:tensorflow:Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nI0920 23:32:51.888205 139884754093952 checkpoint_utils.py:140] Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nINFO:tensorflow:Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-9\nI0920 23:32:52.427159 139884754093952 checkpoint_utils.py:149] Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-9\nINFO:tensorflow:Finished eval step 0\nI0920 23:42:46.962388 139884754093952 model_lib_v2.py:958] Finished eval step 0\nINFO:tensorflow:Performing evaluation on 57 images.\nI0920 23:43:05.734656 139884754093952 coco_evaluation.py:293] Performing evaluation on 57 images.\ncreating index...\nindex created!\nINFO:tensorflow:Loading and preparing annotation results...\nI0920 23:43:05.735040 139884754093952 coco_tools.py:116] Loading and preparing annotation results...\nINFO:tensorflow:DONE (t=0.00s)\nI0920 23:43:05.737743 139884754093952 coco_tools.py:138] DONE (t=0.00s)\ncreating index...\nindex created!\nRunning per image evaluation...\nEvaluate annotation type *bbox*\nDONE (t=0.28s).\nAccumulating evaluation results...\nDONE (t=0.05s).\n Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.508\n Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.783\n Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.640\n Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.503\n Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.742\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.541\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.559\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.579\n Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.572\n Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.750\nINFO:tensorflow:Eval metrics at step 4000\nI0920 23:43:06.074270 139884754093952 model_lib_v2.py:1007] Eval metrics at step 4000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP: 0.508466\nI0920 23:43:07.061945 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP: 0.508466\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.783376\nI0920 23:43:07.063683 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.783376\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.639705\nI0920 23:43:07.064864 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.639705\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (small): -1.000000\nI0920 23:43:07.065825 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (medium): 0.502934\nI0920 23:43:07.066890 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (medium): 0.502934\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (large): 0.742244\nI0920 23:43:07.067832 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (large): 0.742244\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@1: 0.540698\nI0920 23:43:07.068885 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@1: 0.540698\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@10: 0.559169\nI0920 23:43:07.069938 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@10: 0.559169\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100: 0.578980\nI0920 23:43:07.070971 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100: 0.578980\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nI0920 23:43:07.071917 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (medium): 0.572406\nI0920 23:43:07.072993 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (medium): 0.572406\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (large): 0.750000\nI0920 23:43:08.076527 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (large): 0.750000\nINFO:tensorflow:\t+ Loss/localization_loss: 0.006354\nI0920 23:43:08.078019 139884754093952 model_lib_v2.py:1010] \t+ Loss/localization_loss: 0.006354\nINFO:tensorflow:\t+ Loss/classification_loss: 0.474249\nI0920 23:43:08.079171 139884754093952 model_lib_v2.py:1010] \t+ Loss/classification_loss: 0.474249\nINFO:tensorflow:\t+ Loss/regularization_loss: 0.048564\nI0920 23:43:08.080197 139884754093952 model_lib_v2.py:1010] \t+ Loss/regularization_loss: 0.048564\nINFO:tensorflow:\t+ Loss/total_loss: 0.529167\nI0920 23:43:08.081217 139884754093952 model_lib_v2.py:1010] \t+ Loss/total_loss: 0.529167\nINFO:tensorflow:Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nI0920 23:43:09.291560 139884754093952 checkpoint_utils.py:140] Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nINFO:tensorflow:Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-10\nI0920 23:43:09.826003 139884754093952 checkpoint_utils.py:149] Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-10\nINFO:tensorflow:Finished eval step 0\nI0920 23:53:06.372712 139884754093952 model_lib_v2.py:958] Finished eval step 0\nINFO:tensorflow:Performing evaluation on 57 images.\nI0920 23:53:37.054055 139884754093952 coco_evaluation.py:293] Performing evaluation on 57 images.\ncreating index...\nindex created!\nINFO:tensorflow:Loading and preparing annotation results...\nI0920 23:53:37.054489 139884754093952 coco_tools.py:116] Loading and preparing annotation results...\nINFO:tensorflow:DONE (t=0.00s)\nI0920 23:53:37.057577 139884754093952 coco_tools.py:138] DONE (t=0.00s)\ncreating index...\nindex created!\nRunning per image evaluation...\nEvaluate annotation type *bbox*\nDONE (t=0.28s).\nAccumulating evaluation results...\nDONE (t=0.05s).\n Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.505\n Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.785\n Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.611\n Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.500\n Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.717\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.549\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.561\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.584\n Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.579\n Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.725\nINFO:tensorflow:Eval metrics at step 4500\nI0920 23:53:37.395220 139884754093952 model_lib_v2.py:1007] Eval metrics at step 4500\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP: 0.505186\nI0920 23:53:38.454051 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP: 0.505186\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.785156\nI0920 23:53:38.455716 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.785156\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.610924\nI0920 23:53:38.457030 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.610924\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (small): -1.000000\nI0920 23:53:38.458041 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (medium): 0.500335\nI0920 23:53:38.459143 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (medium): 0.500335\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (large): 0.716997\nI0920 23:53:38.460223 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (large): 0.716997\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@1: 0.548924\nI0920 23:53:38.461266 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@1: 0.548924\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@10: 0.560791\nI0920 23:53:38.462334 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@10: 0.560791\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100: 0.584376\nI0920 23:53:38.463374 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100: 0.584376\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nI0920 23:53:38.464262 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (medium): 0.578913\nI0920 23:53:38.465349 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (medium): 0.578913\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (large): 0.725000\nI0920 23:53:39.454225 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (large): 0.725000\nINFO:tensorflow:\t+ Loss/localization_loss: 0.006298\nI0920 23:53:39.455665 139884754093952 model_lib_v2.py:1010] \t+ Loss/localization_loss: 0.006298\nINFO:tensorflow:\t+ Loss/classification_loss: 0.472697\nI0920 23:53:39.456723 139884754093952 model_lib_v2.py:1010] \t+ Loss/classification_loss: 0.472697\nINFO:tensorflow:\t+ Loss/regularization_loss: 0.048548\nI0920 23:53:39.457654 139884754093952 model_lib_v2.py:1010] \t+ Loss/regularization_loss: 0.048548\nINFO:tensorflow:\t+ Loss/total_loss: 0.527543\nI0920 23:53:39.458603 139884754093952 model_lib_v2.py:1010] \t+ Loss/total_loss: 0.527543\nINFO:tensorflow:Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nI0920 23:53:40.653300 139884754093952 checkpoint_utils.py:140] Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nINFO:tensorflow:Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-12\nI0920 23:53:41.079544 139884754093952 checkpoint_utils.py:149] Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-12\nINFO:tensorflow:Finished eval step 0\nI0921 00:03:42.307473 139884754093952 model_lib_v2.py:958] Finished eval step 0\nINFO:tensorflow:Performing evaluation on 57 images.\nI0921 00:04:00.692014 139884754093952 coco_evaluation.py:293] Performing evaluation on 57 images.\ncreating index...\nindex created!\nINFO:tensorflow:Loading and preparing annotation results...\nI0921 00:04:00.692448 139884754093952 coco_tools.py:116] Loading and preparing annotation results...\nINFO:tensorflow:DONE (t=0.00s)\nI0921 00:04:00.695509 139884754093952 coco_tools.py:138] DONE (t=0.00s)\ncreating index...\nindex created!\nRunning per image evaluation...\nEvaluate annotation type *bbox*\nDONE (t=0.28s).\nAccumulating evaluation results...\nDONE (t=0.05s).\n Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.503\n Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.786\n Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.583\n Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.497\n Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.759\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.555\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.566\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.584\n Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.577\n Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.775\nINFO:tensorflow:Eval metrics at step 5500\nI0921 00:04:01.030855 139884754093952 model_lib_v2.py:1007] Eval metrics at step 5500\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP: 0.503479\nI0921 00:04:02.022570 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP: 0.503479\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.786221\nI0921 00:04:02.024234 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.786221\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.583225\nI0921 00:04:02.025425 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.583225\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (small): -1.000000\nI0921 00:04:02.026396 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (medium): 0.496839\nI0921 00:04:02.027456 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (medium): 0.496839\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (large): 0.758746\nI0921 00:04:02.029071 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (large): 0.758746\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@1: 0.555462\nI0921 00:04:02.030199 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@1: 0.555462\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@10: 0.566385\nI0921 00:04:02.031538 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@10: 0.566385\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100: 0.584310\nI0921 00:04:02.032771 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100: 0.584310\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nI0921 00:04:02.033848 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (medium): 0.577023\nI0921 00:04:02.035115 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (medium): 0.577023\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (large): 0.775000\nI0921 00:04:03.117431 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (large): 0.775000\nINFO:tensorflow:\t+ Loss/localization_loss: 0.006275\nI0921 00:04:03.118902 139884754093952 model_lib_v2.py:1010] \t+ Loss/localization_loss: 0.006275\nINFO:tensorflow:\t+ Loss/classification_loss: 0.481848\nI0921 00:04:03.119990 139884754093952 model_lib_v2.py:1010] \t+ Loss/classification_loss: 0.481848\nINFO:tensorflow:\t+ Loss/regularization_loss: 0.048536\nI0921 00:04:03.121022 139884754093952 model_lib_v2.py:1010] \t+ Loss/regularization_loss: 0.048536\nINFO:tensorflow:\t+ Loss/total_loss: 0.536659\nI0921 00:04:03.121978 139884754093952 model_lib_v2.py:1010] \t+ Loss/total_loss: 0.536659\nINFO:tensorflow:Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nI0921 00:04:04.421556 139884754093952 checkpoint_utils.py:140] Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nINFO:tensorflow:Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-13\nI0921 00:04:04.944048 139884754093952 checkpoint_utils.py:149] Found new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo/ckpt-13\nINFO:tensorflow:Finished eval step 0\nI0921 00:14:09.491717 139884754093952 model_lib_v2.py:958] Finished eval step 0\nINFO:tensorflow:Performing evaluation on 57 images.\nI0921 00:14:27.905162 139884754093952 coco_evaluation.py:293] Performing evaluation on 57 images.\ncreating index...\nindex created!\nINFO:tensorflow:Loading and preparing annotation results...\nI0921 00:14:27.905615 139884754093952 coco_tools.py:116] Loading and preparing annotation results...\nINFO:tensorflow:DONE (t=0.00s)\nI0921 00:14:27.908681 139884754093952 coco_tools.py:138] DONE (t=0.00s)\ncreating index...\nindex created!\nRunning per image evaluation...\nEvaluate annotation type *bbox*\nDONE (t=0.29s).\nAccumulating evaluation results...\nDONE (t=0.05s).\n Average Precision (AP) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.504\n Average Precision (AP) @[ IoU=0.50 | area= all | maxDets=100 ] = 0.785\n Average Precision (AP) @[ IoU=0.75 | area= all | maxDets=100 ] = 0.585\n Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.499\n Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.742\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 1 ] = 0.556\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets= 10 ] = 0.566\n Average Recall (AR) @[ IoU=0.50:0.95 | area= all | maxDets=100 ] = 0.582\n Average Recall (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000\n Average Recall (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.576\n Average Recall (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.750\nINFO:tensorflow:Eval metrics at step 6000\nI0921 00:14:28.250011 139884754093952 model_lib_v2.py:1007] Eval metrics at step 6000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP: 0.504456\nI0921 00:14:29.267917 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP: 0.504456\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.785415\nI0921 00:14:29.269681 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.785415\nINFO:tensorflow:\t+ DetectionBoxes_Precision/[email protected]: 0.585339\nI0921 00:14:29.270922 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/[email protected]: 0.585339\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (small): -1.000000\nI0921 00:14:29.271905 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (medium): 0.498571\nI0921 00:14:29.273053 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (medium): 0.498571\nINFO:tensorflow:\t+ DetectionBoxes_Precision/mAP (large): 0.742244\nI0921 00:14:29.274051 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Precision/mAP (large): 0.742244\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@1: 0.555528\nI0921 00:14:29.275205 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@1: 0.555528\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@10: 0.566452\nI0921 00:14:29.276360 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@10: 0.566452\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100: 0.582489\nI0921 00:14:29.277493 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100: 0.582489\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nI0921 00:14:29.278464 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (small): -1.000000\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (medium): 0.576043\nI0921 00:14:29.279589 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (medium): 0.576043\nINFO:tensorflow:\t+ DetectionBoxes_Recall/AR@100 (large): 0.750000\nI0921 00:14:30.306879 139884754093952 model_lib_v2.py:1010] \t+ DetectionBoxes_Recall/AR@100 (large): 0.750000\nINFO:tensorflow:\t+ Loss/localization_loss: 0.006291\nI0921 00:14:30.308382 139884754093952 model_lib_v2.py:1010] \t+ Loss/localization_loss: 0.006291\nINFO:tensorflow:\t+ Loss/classification_loss: 0.478905\nI0921 00:14:30.309785 139884754093952 model_lib_v2.py:1010] \t+ Loss/classification_loss: 0.478905\nINFO:tensorflow:\t+ Loss/regularization_loss: 0.048536\nI0921 00:14:30.310982 139884754093952 model_lib_v2.py:1010] \t+ Loss/regularization_loss: 0.048536\nINFO:tensorflow:\t+ Loss/total_loss: 0.533732\nI0921 00:14:30.312035 139884754093952 model_lib_v2.py:1010] \t+ Loss/total_loss: 0.533732\nINFO:tensorflow:Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nI0921 00:14:31.506707 139884754093952 checkpoint_utils.py:140] Waiting for new checkpoint at gs://s2ds-aug2021-cv/models/fine-tuned/efficient_det_1024_rand_aug_1_4_demo\nTraceback (most recent call last):\n File \"model_main_tf2.py\", line 115, in <module>\n tf.compat.v1.app.run()\n File \"/usr/local/lib/python3.7/dist-packages/tensorflow/python/platform/app.py\", line 40, in run\n _run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef)\n File \"/usr/local/lib/python3.7/dist-packages/absl/app.py\", line 303, in run\n _run_main(main, args)\n File \"/usr/local/lib/python3.7/dist-packages/absl/app.py\", line 251, in _run_main\n sys.exit(main(argv))\n File \"model_main_tf2.py\", line 90, in main\n wait_interval=300, timeout=FLAGS.eval_timeout)\n File \"/content/gdrive/MyDrive/s2ds-aug2021-cv/colab-install/lib/python3.7/site-packages/object_detection-0.1-py3.7.egg/object_detection/model_lib_v2.py\", line 1129, in eval_continuously\n checkpoint_dir, timeout=timeout, min_interval_secs=wait_interval):\n File \"/usr/local/lib/python3.7/dist-packages/tensorflow/python/training/checkpoint_utils.py\", line 199, in checkpoints_iterator\n checkpoint_dir, checkpoint_path, timeout=timeout)\n File \"/usr/local/lib/python3.7/dist-packages/tensorflow/python/training/checkpoint_utils.py\", line 147, in wait_for_new_checkpoint\n time.sleep(seconds_to_sleep)\nKeyboardInterrupt\n/content/gdrive/MyDrive/s2ds-aug2021-cv/notebooks\n" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown", "markdown", "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ] ]

4ad9516a46f55e5af3a1cedf3f7b333987e2ffe3

343,531

ipynb

Jupyter Notebook

intro to data science/Week3.ipynb

narottam90/umich_ds

154adeaf4a160c608a2f3dd2d539263eb266f48d

[ "MIT" ]

null

intro to data science/Week3.ipynb

narottam90/umich_ds

154adeaf4a160c608a2f3dd2d539263eb266f48d

[ "MIT" ]

null

intro to data science/Week3.ipynb

narottam90/umich_ds

154adeaf4a160c608a2f3dd2d539263eb266f48d

[ "MIT" ]

null

188.960946

23,746

0.545485

[ [ [ "# Merging Databases", "_____no_output_____" ] ], [ [ "import pandas as pd\n\nhelp(pd.merge)", "Help on function merge in module pandas.core.reshape.merge:\n\nmerge(left, right, how: str = 'inner', on=None, left_on=None, right_on=None, left_index: bool = False, right_index: bool = False, sort: bool = False, suffixes=('_x', '_y'), copy: bool = True, indicator: bool = False, validate=None) -> 'DataFrame'\n Merge DataFrame or named Series objects with a database-style join.\n \n The join is done on columns or indexes. If joining columns on\n columns, the DataFrame indexes *will be ignored*. Otherwise if joining indexes\n on indexes or indexes on a column or columns, the index will be passed on.\n \n Parameters\n ----------\n left : DataFrame\n right : DataFrame or named Series\n Object to merge with.\n how : {'left', 'right', 'outer', 'inner'}, default 'inner'\n Type of merge to be performed.\n \n * left: use only keys from left frame, similar to a SQL left outer join;\n preserve key order.\n * right: use only keys from right frame, similar to a SQL right outer join;\n preserve key order.\n * outer: use union of keys from both frames, similar to a SQL full outer\n join; sort keys lexicographically.\n * inner: use intersection of keys from both frames, similar to a SQL inner\n join; preserve the order of the left keys.\n on : label or list\n Column or index level names to join on. These must be found in both\n DataFrames. If `on` is None and not merging on indexes then this defaults\n to the intersection of the columns in both DataFrames.\n left_on : label or list, or array-like\n Column or index level names to join on in the left DataFrame. Can also\n be an array or list of arrays of the length of the left DataFrame.\n These arrays are treated as if they are columns.\n right_on : label or list, or array-like\n Column or index level names to join on in the right DataFrame. Can also\n be an array or list of arrays of the length of the right DataFrame.\n These arrays are treated as if they are columns.\n left_index : bool, default False\n Use the index from the left DataFrame as the join key(s). If it is a\n MultiIndex, the number of keys in the other DataFrame (either the index\n or a number of columns) must match the number of levels.\n right_index : bool, default False\n Use the index from the right DataFrame as the join key. Same caveats as\n left_index.\n sort : bool, default False\n Sort the join keys lexicographically in the result DataFrame. If False,\n the order of the join keys depends on the join type (how keyword).\n suffixes : tuple of (str, str), default ('_x', '_y')\n Suffix to apply to overlapping column names in the left and right\n side, respectively. To raise an exception on overlapping columns use\n (False, False).\n copy : bool, default True\n If False, avoid copy if possible.\n indicator : bool or str, default False\n If True, adds a column to output DataFrame called \"_merge\" with\n information on the source of each row.\n If string, column with information on source of each row will be added to\n output DataFrame, and column will be named value of string.\n Information column is Categorical-type and takes on a value of \"left_only\"\n for observations whose merge key only appears in 'left' DataFrame,\n \"right_only\" for observations whose merge key only appears in 'right'\n DataFrame, and \"both\" if the observation's merge key is found in both.\n \n validate : str, optional\n If specified, checks if merge is of specified type.\n \n * \"one_to_one\" or \"1:1\": check if merge keys are unique in both\n left and right datasets.\n * \"one_to_many\" or \"1:m\": check if merge keys are unique in left\n dataset.\n * \"many_to_one\" or \"m:1\": check if merge keys are unique in right\n dataset.\n * \"many_to_many\" or \"m:m\": allowed, but does not result in checks.\n \n .. versionadded:: 0.21.0\n \n Returns\n -------\n DataFrame\n A DataFrame of the two merged objects.\n \n See Also\n --------\n merge_ordered : Merge with optional filling/interpolation.\n merge_asof : Merge on nearest keys.\n DataFrame.join : Similar method using indices.\n \n Notes\n -----\n Support for specifying index levels as the `on`, `left_on`, and\n `right_on` parameters was added in version 0.23.0\n Support for merging named Series objects was added in version 0.24.0\n \n Examples\n --------\n \n >>> df1 = pd.DataFrame({'lkey': ['foo', 'bar', 'baz', 'foo'],\n ... 'value': [1, 2, 3, 5]})\n >>> df2 = pd.DataFrame({'rkey': ['foo', 'bar', 'baz', 'foo'],\n ... 'value': [5, 6, 7, 8]})\n >>> df1\n lkey value\n 0 foo 1\n 1 bar 2\n 2 baz 3\n 3 foo 5\n >>> df2\n rkey value\n 0 foo 5\n 1 bar 6\n 2 baz 7\n 3 foo 8\n \n Merge df1 and df2 on the lkey and rkey columns. The value columns have\n the default suffixes, _x and _y, appended.\n \n >>> df1.merge(df2, left_on='lkey', right_on='rkey')\n lkey value_x rkey value_y\n 0 foo 1 foo 5\n 1 foo 1 foo 8\n 2 foo 5 foo 5\n 3 foo 5 foo 8\n 4 bar 2 bar 6\n 5 baz 3 baz 7\n \n Merge DataFrames df1 and df2 with specified left and right suffixes\n appended to any overlapping columns.\n \n >>> df1.merge(df2, left_on='lkey', right_on='rkey',\n ... suffixes=('_left', '_right'))\n lkey value_left rkey value_right\n 0 foo 1 foo 5\n 1 foo 1 foo 8\n 2 foo 5 foo 5\n 3 foo 5 foo 8\n 4 bar 2 bar 6\n 5 baz 3 baz 7\n \n Merge DataFrames df1 and df2, but raise an exception if the DataFrames have\n any overlapping columns.\n \n >>> df1.merge(df2, left_on='lkey', right_on='rkey', suffixes=(False, False))\n Traceback (most recent call last):\n ...\n ValueError: columns overlap but no suffix specified:\n Index(['value'], dtype='object')\n\n" ], [ "df = pd.DataFrame([{'Name': 'Chris', 'Item Purchased': 'Sponge', 'Cost': 22.50},\n {'Name': 'Kevyn', 'Item Purchased': 'Kitty Litter', 'Cost': 2.50},\n {'Name': 'Filip', 'Item Purchased': 'Spoon', 'Cost': 5.00}],\n index=['Store 1', 'Store 1', 'Store 2'])\ndf", "_____no_output_____" ], [ "df['Date'] = ['December 1', 'January 1', 'mid-May'] # adding values using the square bracket index operator. index is shared\ndf", "_____no_output_____" ], [ "df['Delivered'] = True # adding a scalar value which becomes the default value for the entire column\ndf", "_____no_output_____" ], [ "df['Feedback'] = ['Positive', None, 'Negative'] # when we need to input a few values, we must supply the None values ourselves\ndf", "_____no_output_____" ], [ "# if rows have a unique index, we can assign a new column identifier to the series\nadf = df.reset_index()\n\n# think of it as create a index called 'Date' and set the value at row 0 to 'Dec 1'\nadf['Date'] = pd.Series({0: 'December 1', 2: 'mid-May'}) \n\nadf", "_____no_output_____" ], [ "staff_df = pd.DataFrame([{'Name': 'Kelly', 'Role': 'Director of HR'},\n {'Name': 'Sally', 'Role': 'Course liasion'},\n {'Name': 'James', 'Role': 'Grader'}])\n\nstaff_df = staff_df.set_index('Name')\n\nstudent_df = pd.DataFrame([{'Name': 'James', 'School': 'Business'},\n {'Name': 'Mike', 'School': 'Law'},\n {'Name': 'Sally', 'School': 'Engineering'}])\n\nstudent_df = student_df.set_index('Name')\n\nprint(staff_df.head())\nprint()\nprint(student_df.head())", "Role\nName \nKelly Director of HR\nSally Course liasion\nJames Grader\n\n School\nName \nJames Business\nMike Law\nSally Engineering\n" ], [ "# call merge passing in 'left' df and 'right' df, using an outer join. We also tell merge that we want to use the left and right indices as the joining key(s). In this case, the indices are the same. 'outer' will sort keys lexicographically.\npd.merge(staff_df, student_df, how='outer', left_index=True, right_index=True)", "_____no_output_____" ], [ "# inner join gives us a dataframe with just those students who are also staff\npd.merge(staff_df, student_df, how='inner', left_index=True, right_index=True)", "_____no_output_____" ], [ "# left gives us a list of all staff regardless of whether they are students are not. if they are students, also gets their student details \npd.merge(staff_df, student_df, how='left', left_index=True, right_index=True)", "_____no_output_____" ], [ "# right gives us a list of all students\npd.merge(staff_df, student_df, how='right', left_index=True, right_index=True)", "_____no_output_____" ], [ "# pd.merge can also use column names to join on\n\nstaff_df = staff_df.reset_index()\nstudent_df = student_df.reset_index()\n\npd.merge(staff_df, student_df, how='left', left_on='Name', right_on='Name')", "_____no_output_____" ], [ "# in cases where there is conflicting data, pandas preserves the data and appends either _x or _y depending on which dataframe it originated from\n# _x - left dataframe\n# _y - right dataframe\n\nstaff_df = pd.DataFrame([{'Name': 'Kelly', 'Role': 'Director of HR', 'Location': 'State Street'},\n {'Name': 'Sally', 'Role': 'Course liasion', 'Location': 'Washington Avenue'},\n {'Name': 'James', 'Role': 'Grader', 'Location': 'Washington Avenue'}])\n\nstudent_df = pd.DataFrame([{'Name': 'James', 'School': 'Business', 'Location': '1024 Billiard Avenue'},\n {'Name': 'Mike', 'School': 'Law', 'Location': 'Fraternity House #22'},\n {'Name': 'Sally', 'School': 'Engineering', 'Location': '512 Wilson Crescent'}])\n\npd.merge(staff_df, student_df, how='left', left_on='Name', right_on='Name')", "_____no_output_____" ], [ "# merge can be passed the suffixes parameter which takes a tuple to control how _x and _y are represented\n\npd.merge(staff_df, student_df, how='left', left_on='Name', right_on='Name', suffixes=('_Staff', '_Student'))", "_____no_output_____" ], [ "staff_df = pd.DataFrame([{'First Name': 'Kelly', 'Last Name': 'Desjardins', 'Role': 'Director of HR'},\n {'First Name': 'Sally', 'Last Name': 'Brooks', 'Role': 'Course liasion'},\n {'First Name': 'James', 'Last Name': 'Wilde', 'Role': 'Grader'}])\nstudent_df = pd.DataFrame([{'First Name': 'James', 'Last Name': 'Hammond', 'School': 'Business'},\n {'First Name': 'Mike', 'Last Name': 'Smith', 'School': 'Law'},\n {'First Name': 'Sally', 'Last Name': 'Brooks', 'School': 'Engineering'}])\n\nprint(staff_df)\nprint()\nprint(student_df)", "First Name Last Name Role\n0 Kelly Desjardins Director of HR\n1 Sally Brooks Course liasion\n2 James Wilde Grader\n\n First Name Last Name School\n0 James Hammond Business\n1 Mike Smith Law\n2 Sally Brooks Engineering\n" ], [ "# dataframes can also be merged on multiple column names and multi-indexes. The columns are passed as a list to the left_on/right_on parameteres. In this case, 'James Hammond' and 'James Wilde' will not satisfy the inner match condition.\n\npd.merge(staff_df, student_df, how='inner', left_on=['First Name', 'Last Name'], right_on=['First Name', 'Last Name'])", "_____no_output_____" ] ], [ [ "# Idiomatic Pandas: Making Code Pandorable", "_____no_output_____" ] ], [ [ "import pandas as pd, numpy as np, os\n\nos.chdir('/Users/riro/Documents/GitHub/umich_ds/intro to data science/files')\n\ndf = pd.read_csv('census.csv')\ndf", "_____no_output_____" ], [ "# Pandorable code using method chaining\n(df.where(df['SUMLEV'] == 50).dropna().set_index(['STNAME', 'CTYNAME']).rename(columns={'ESTIMATESBASE2010' : 'Estimates Base 2010'}))", "_____no_output_____" ], [ "# Unpandorable code\ndf = df[df['SUMLEV'] == 50]\ndf.set_index(['STNAME', 'CTYNAME'], inplace = True)\ndf.rename(columns={'ESTIMATESBASE2010' : 'Estimates Base 2010'})", "_____no_output_____" ], [ "# Using apply(): applies a function along an axis of the DataFrame. axis: '0' apply function to each column, '1' for each row\n\n# min_max takes in a particular row of data, finds a min and max value and returns a new row of data\ndef min_max(row):\n\n # create a small slice by projecting the population columns\n data = row[['POPESTIMATE2010',\n 'POPESTIMATE2011',\n 'POPESTIMATE2012',\n 'POPESTIMATE2013',\n 'POPESTIMATE2014',\n 'POPESTIMATE2015']]\n # create a new series with label values\n return pd.Series({'min': np.min(data), 'max': np.max(data)})\n\n# call apply on the dataframe\ndf.apply(min_max, axis=1)", "_____no_output_____" ], [ "def min_max(row):\n data = row[['POPESTIMATE2010',\n 'POPESTIMATE2011',\n 'POPESTIMATE2012',\n 'POPESTIMATE2013',\n 'POPESTIMATE2014',\n 'POPESTIMATE2015']]\n row['max'] = np.max(data)\n row['min'] = np.min(data)\n return row\n\ndf.apply(min_max, axis = 1)\n", "_____no_output_____" ], [ "# Using lambda functions\nrows = ['POPESTIMATE2010',\n 'POPESTIMATE2011',\n 'POPESTIMATE2012',\n 'POPESTIMATE2013',\n 'POPESTIMATE2014',\n 'POPESTIMATE2015']\ndf.apply(lambda x: np.max(x[rows]), axis=1)", "_____no_output_____" ], [ "help(df.apply)", "Help on method apply in module pandas.core.frame:\n\napply(func, axis=0, raw=False, result_type=None, args=(), **kwds) method of pandas.core.frame.DataFrame instance\n Apply a function along an axis of the DataFrame.\n \n Objects passed to the function are Series objects whose index is\n either the DataFrame's index (``axis=0``) or the DataFrame's columns\n (``axis=1``). By default (``result_type=None``), the final return type\n is inferred from the return type of the applied function. Otherwise,\n it depends on the `result_type` argument.\n \n Parameters\n ----------\n func : function\n Function to apply to each column or row.\n axis : {0 or 'index', 1 or 'columns'}, default 0\n Axis along which the function is applied:\n \n * 0 or 'index': apply function to each column.\n * 1 or 'columns': apply function to each row.\n \n raw : bool, default False\n Determines if row or column is passed as a Series or ndarray object:\n \n * ``False`` : passes each row or column as a Series to the\n function.\n * ``True`` : the passed function will receive ndarray objects\n instead.\n If you are just applying a NumPy reduction function this will\n achieve much better performance.\n \n result_type : {'expand', 'reduce', 'broadcast', None}, default None\n These only act when ``axis=1`` (columns):\n \n * 'expand' : list-like results will be turned into columns.\n * 'reduce' : returns a Series if possible rather than expanding\n list-like results. This is the opposite of 'expand'.\n * 'broadcast' : results will be broadcast to the original shape\n of the DataFrame, the original index and columns will be\n retained.\n \n The default behaviour (None) depends on the return value of the\n applied function: list-like results will be returned as a Series\n of those. However if the apply function returns a Series these\n are expanded to columns.\n \n .. versionadded:: 0.23.0\n \n args : tuple\n Positional arguments to pass to `func` in addition to the\n array/series.\n **kwds\n Additional keyword arguments to pass as keywords arguments to\n `func`.\n \n Returns\n -------\n Series or DataFrame\n Result of applying ``func`` along the given axis of the\n DataFrame.\n \n See Also\n --------\n DataFrame.applymap: For elementwise operations.\n DataFrame.aggregate: Only perform aggregating type operations.\n DataFrame.transform: Only perform transforming type operations.\n \n Examples\n --------\n \n >>> df = pd.DataFrame([[4, 9]] * 3, columns=['A', 'B'])\n >>> df\n A B\n 0 4 9\n 1 4 9\n 2 4 9\n \n Using a numpy universal function (in this case the same as\n ``np.sqrt(df)``):\n \n >>> df.apply(np.sqrt)\n A B\n 0 2.0 3.0\n 1 2.0 3.0\n 2 2.0 3.0\n \n Using a reducing function on either axis\n \n >>> df.apply(np.sum, axis=0)\n A 12\n B 27\n dtype: int64\n \n >>> df.apply(np.sum, axis=1)\n 0 13\n 1 13\n 2 13\n dtype: int64\n \n Returning a list-like will result in a Series\n \n >>> df.apply(lambda x: [1, 2], axis=1)\n 0 [1, 2]\n 1 [1, 2]\n 2 [1, 2]\n dtype: object\n \n Passing result_type='expand' will expand list-like results\n to columns of a Dataframe\n \n >>> df.apply(lambda x: [1, 2], axis=1, result_type='expand')\n 0 1\n 0 1 2\n 1 1 2\n 2 1 2\n \n Returning a Series inside the function is similar to passing\n ``result_type='expand'``. The resulting column names\n will be the Series index.\n \n >>> df.apply(lambda x: pd.Series([1, 2], index=['foo', 'bar']), axis=1)\n foo bar\n 0 1 2\n 1 1 2\n 2 1 2\n \n Passing ``result_type='broadcast'`` will ensure the same shape\n result, whether list-like or scalar is returned by the function,\n and broadcast it along the axis. The resulting column names will\n be the originals.\n \n >>> df.apply(lambda x: [1, 2], axis=1, result_type='broadcast')\n A B\n 0 1 2\n 1 1 2\n 2 1 2\n\n" ] ], [ [ "# Group by", "_____no_output_____" ] ], [ [ "# Looking at our backpacking equipment DataFrame, suppose we are interested in finding our total weight for each category. Use groupby to group the dataframe, and apply a function to calculate the total weight (Weight x Quantity) by category.\n\nimport pandas as pd, numpy as np \n\ndf = pd.DataFrame({'Item': ['Pack', 'Tent',\n 'Sleeping Pad', 'Sleeping Bag', 'Water Bottles'],\n 'Category': ['Pack', 'Shelter', 'Sleep', 'Sleep', 'Kitchen'], \n 'Quantity': [1, 1, 1, 1, 2],\n 'Weight (oz.)' : [33., 80., 27., 20., 35.]})\n\ndf = df.set_index('Item')\ndf", "_____no_output_____" ], [ "df.groupby(['Category']).apply(lambda x: sum(x['Quantity'] * x['Weight (oz.)']))", "_____no_output_____" ], [ "import pandas as pd\nimport numpy as np\ndf = pd.read_csv('census.csv')\ndf = df[df['SUMLEV']==50]", "_____no_output_____" ], [ "%%timeit -n 5\n\nfor state in df['STNAME'].unique():\n avg = np.average(df.where(df['STNAME'] == state).dropna()['CENSUS2010POP'])\n print('Counties in state ' + state + ' have an average population of ' + str(avg) )", "44155\nCounties in state Oregon have an average population of 106418.72222222222\nCounties in state Pennsylvania have an average population of 189587.74626865672\nCounties in state Rhode Island have an average population of 210513.4\nCounties in state South Carolina have an average population of 100551.39130434782\nCounties in state South Dakota have an average population of 12336.060606060606\nCounties in state Tennessee have an average population of 66801.1052631579\nCounties in state Texas have an average population of 98998.27165354331\nCounties in state Utah have an average population of 95306.37931034483\nCounties in state Vermont have an average population of 44695.78571428572\nCounties in state Virginia have an average population of 60111.29323308271\nCounties in state Washington have an average population of 172424.10256410256\nCounties in state West Virginia have an average population of 33690.8\nCounties in state Wisconsin have an average population of 78985.91666666667\nCounties in state Wyoming have an average population of 24505.478260869564\nCounties in state Alabama have an average population of 71339.34328358209\nCounties in state Alaska have an average population of 24490.724137931036\nCounties in state Arizona have an average population of 426134.4666666667\nCounties in state Arkansas have an average population of 38878.90666666667\nCounties in state California have an average population of 642309.5862068966\nCounties in state Colorado have an average population of 78581.1875\nCounties in state Connecticut have an average population of 446762.125\nCounties in state Delaware have an average population of 299311.3333333333\nCounties in state District of Columbia have an average population of 601723.0\nCounties in state Florida have an average population of 280616.5671641791\nCounties in state Georgia have an average population of 60928.63522012578\nCounties in state Hawaii have an average population of 272060.2\nCounties in state Idaho have an average population of 35626.86363636364\nCounties in state Illinois have an average population of 125790.50980392157\nCounties in state Indiana have an average population of 70476.10869565218\nCounties in state Iowa have an average population of 30771.262626262625\nCounties in state Kansas have an average population of 27172.55238095238\nCounties in state Kentucky have an average population of 36161.39166666667\nCounties in state Louisiana have an average population of 70833.9375\nCounties in state Maine have an average population of 83022.5625\nCounties in state Maryland have an average population of 240564.66666666666\nCounties in state Massachusetts have an average population of 467687.78571428574\nCounties in state Michigan have an average population of 119080.0\nCounties in state Minnesota have an average population of 60964.65517241379\nCounties in state Mississippi have an average population of 36186.54878048781\nCounties in state Missouri have an average population of 52077.62608695652\nCounties in state Montana have an average population of 17668.125\nCounties in state Nebraska have an average population of 19638.075268817203\nCounties in state Nevada have an average population of 158855.9411764706\nCounties in state New Hampshire have an average population of 131647.0\nCounties in state New Jersey have an average population of 418661.61904761905\nCounties in state New Mexico have an average population of 62399.36363636364\nCounties in state New York have an average population of 312550.03225806454\nCounties in state North Carolina have an average population of 95354.83\nCounties in state North Dakota have an average population of 12690.396226415094\nCounties in state Ohio have an average population of 131096.63636363635\nCounties in state Oklahoma have an average population of 48718.844155844155\nCounties in state Oregon have an average population of 106418.72222222222\nCounties in state Pennsylvania have an average population of 189587.74626865672\nCounties in state Rhode Island have an average population of 210513.4\nCounties in state South Carolina have an average population of 100551.39130434782\nCounties in state South Dakota have an average population of 12336.060606060606\nCounties in state Tennessee have an average population of 66801.1052631579\nCounties in state Texas have an average population of 98998.27165354331\nCounties in state Utah have an average population of 95306.37931034483\nCounties in state Vermont have an average population of 44695.78571428572\nCounties in state Virginia have an average population of 60111.29323308271\nCounties in state Washington have an average population of 172424.10256410256\nCounties in state West Virginia have an average population of 33690.8\nCounties in state Wisconsin have an average population of 78985.91666666667\nCounties in state Wyoming have an average population of 24505.478260869564\nCounties in state Alabama have an average population of 71339.34328358209\nCounties in state Alaska have an average population of 24490.724137931036\nCounties in state Arizona have an average population of 426134.4666666667\nCounties in state Arkansas have an average population of 38878.90666666667\nCounties in state California have an average population of 642309.5862068966\nCounties in state Colorado have an average population of 78581.1875\nCounties in state Connecticut have an average population of 446762.125\nCounties in state Delaware have an average population of 299311.3333333333\nCounties in state District of Columbia have an average population of 601723.0\nCounties in state Florida have an average population of 280616.5671641791\nCounties in state Georgia have an average population of 60928.63522012578\nCounties in state Hawaii have an average population of 272060.2\nCounties in state Idaho have an average population of 35626.86363636364\nCounties in state Illinois have an average population of 125790.50980392157\nCounties in state Indiana have an average population of 70476.10869565218\nCounties in state Iowa have an average population of 30771.262626262625\nCounties in state Kansas have an average population of 27172.55238095238\nCounties in state Kentucky have an average population of 36161.39166666667\nCounties in state Louisiana have an average population of 70833.9375\nCounties in state Maine have an average population of 83022.5625\nCounties in state Maryland have an average population of 240564.66666666666\nCounties in state Massachusetts have an average population of 467687.78571428574\nCounties in state Michigan have an average population of 119080.0\nCounties in state Minnesota have an average population of 60964.65517241379\nCounties in state Mississippi have an average population of 36186.54878048781\nCounties in state Missouri have an average population of 52077.62608695652\nCounties in state Montana have an average population of 17668.125\nCounties in state Nebraska have an average population of 19638.075268817203\nCounties in state Nevada have an average population of 158855.9411764706\nCounties in state New Hampshire have an average population of 131647.0\nCounties in state New Jersey have an average population of 418661.61904761905\nCounties in state New Mexico have an average population of 62399.36363636364\nCounties in state New York have an average population of 312550.03225806454\nCounties in state North Carolina have an average population of 95354.83\nCounties in state North Dakota have an average population of 12690.396226415094\nCounties in state Ohio have an average population of 131096.63636363635\nCounties in state Oklahoma have an average population of 48718.844155844155\nCounties in state Oregon have an average population of 106418.72222222222\nCounties in state Pennsylvania have an average population of 189587.74626865672\nCounties in state Rhode Island have an average population of 210513.4\nCounties in state South Carolina have an average population of 100551.39130434782\nCounties in state South Dakota have an average population of 12336.060606060606\nCounties in state Tennessee have an average population of 66801.1052631579\nCounties in state Texas have an average population of 98998.27165354331\nCounties in state Utah have an average population of 95306.37931034483\nCounties in state Vermont have an average population of 44695.78571428572\nCounties in state Virginia have an average population of 60111.29323308271\nCounties in state Washington have an average population of 172424.10256410256\nCounties in state West Virginia have an average population of 33690.8\nCounties in state Wisconsin have an average population of 78985.91666666667\nCounties in state Wyoming have an average population of 24505.478260869564\nCounties in state Alabama have an average population of 71339.34328358209\nCounties in state Alaska have an average population of 24490.724137931036\nCounties in state Arizona have an average population of 426134.4666666667\nCounties in state Arkansas have an average population of 38878.90666666667\nCounties in state California have an average population of 642309.5862068966\nCounties in state Colorado have an average population of 78581.1875\nCounties in state Connecticut have an average population of 446762.125\nCounties in state Delaware have an average population of 299311.3333333333\nCounties in state District of Columbia have an average population of 601723.0\nCounties in state Florida have an average population of 280616.5671641791\nCounties in state Georgia have an average population of 60928.63522012578\nCounties in state Hawaii have an average population of 272060.2\nCounties in state Idaho have an average population of 35626.86363636364\nCounties in state Illinois have an average population of 125790.50980392157\nCounties in state Indiana have an average population of 70476.10869565218\nCounties in state Iowa have an average population of 30771.262626262625\nCounties in state Kansas have an average population of 27172.55238095238\nCounties in state Kentucky have an average population of 36161.39166666667\nCounties in state Louisiana have an average population of 70833.9375\nCounties in state Maine have an average population of 83022.5625\nCounties in state Maryland have an average population of 240564.66666666666\nCounties in state Massachusetts have an average population of 467687.78571428574\nCounties in state Michigan have an average population of 119080.0\nCounties in state Minnesota have an average population of 60964.65517241379\nCounties in state Mississippi have an average population of 36186.54878048781\nCounties in state Missouri have an average population of 52077.62608695652\nCounties in state Montana have an average population of 17668.125\nCounties in state Nebraska have an average population of 19638.075268817203\nCounties in state Nevada have an average population of 158855.9411764706\nCounties in state New Hampshire have an average population of 131647.0\nCounties in state New Jersey have an average population of 418661.61904761905\nCounties in state New Mexico have an average population of 62399.36363636364\nCounties in state New York have an average population of 312550.03225806454\nCounties in state North Carolina have an average population of 95354.83\nCounties in state North Dakota have an average population of 12690.396226415094\nCounties in state Ohio have an average population of 131096.63636363635\nCounties in state Oklahoma have an average population of 48718.844155844155\nCounties in state Oregon have an average population of 106418.72222222222\nCounties in state Pennsylvania have an average population of 189587.74626865672\nCounties in state Rhode Island have an average population of 210513.4\nCounties in state South Carolina have an average population of 100551.39130434782\nCounties in state South Dakota have an average population of 12336.060606060606\nCounties in state Tennessee have an average population of 66801.1052631579\nCounties in state Texas have an average population of 98998.27165354331\nCounties in state Utah have an average population of 95306.37931034483\nCounties in state Vermont have an average population of 44695.78571428572\nCounties in state Virginia have an average population of 60111.29323308271\nCounties in state Washington have an average population of 172424.10256410256\nCounties in state West Virginia have an average population of 33690.8\nCounties in state Wisconsin have an average population of 78985.91666666667\nCounties in state Wyoming have an average population of 24505.478260869564\nCounties in state Alabama have an average population of 71339.34328358209\nCounties in state Alaska have an average population of 24490.724137931036\nCounties in state Arizona have an average population of 426134.4666666667\nCounties in state Arkansas have an average population of 38878.90666666667\nCounties in state California have an average population of 642309.5862068966\nCounties in state Colorado have an average population of 78581.1875\nCounties in state Connecticut have an average population of 446762.125\nCounties in state Delaware have an average population of 299311.3333333333\nCounties in state District of Columbia have an average population of 601723.0\nCounties in state Florida have an average population of 280616.5671641791\nCounties in state Georgia have an average population of 60928.63522012578\nCounties in state Hawaii have an average population of 272060.2\nCounties in state Idaho have an average population of 35626.86363636364\nCounties in state Illinois have an average population of 125790.50980392157\nCounties in state Indiana have an average population of 70476.10869565218\nCounties in state Iowa have an average population of 30771.262626262625\nCounties in state Kansas have an average population of 27172.55238095238\nCounties in state Kentucky have an average population of 36161.39166666667\nCounties in state Louisiana have an average population of 70833.9375\nCounties in state Maine have an average population of 83022.5625\nCounties in state Maryland have an average population of 240564.66666666666\nCounties in state Massachusetts have an average population of 467687.78571428574\nCounties in state Michigan have an average population of 119080.0\nCounties in state Minnesota have an average population of 60964.65517241379\nCounties in state Mississippi have an average population of 36186.54878048781\nCounties in state Missouri have an average population of 52077.62608695652\nCounties in state Montana have an average population of 17668.125\nCounties in state Nebraska have an average population of 19638.075268817203\nCounties in state Nevada have an average population of 158855.9411764706\nCounties in state New Hampshire have an average population of 131647.0\nCounties in state New Jersey have an average population of 418661.61904761905\nCounties in state New Mexico have an average population of 62399.36363636364\nCounties in state New York have an average population of 312550.03225806454\nCounties in state North Carolina have an average population of 95354.83\nCounties in state North Dakota have an average population of 12690.396226415094\nCounties in state Ohio have an average population of 131096.63636363635\nCounties in state Oklahoma have an average population of 48718.844155844155\nCounties in state Oregon have an average population of 106418.72222222222\nCounties in state Pennsylvania have an average population of 189587.74626865672\nCounties in state Rhode Island have an average population of 210513.4\nCounties in state South Carolina have an average population of 100551.39130434782\nCounties in state South Dakota have an average population of 12336.060606060606\nCounties in state Tennessee have an average population of 66801.1052631579\nCounties in state Texas have an average population of 98998.27165354331\nCounties in state Utah have an average population of 95306.37931034483\nCounties in state Vermont have an average population of 44695.78571428572\nCounties in state Virginia have an average population of 60111.29323308271\nCounties in state Washington have an average population of 172424.10256410256\nCounties in state West Virginia have an average population of 33690.8\nCounties in state Wisconsin have an average population of 78985.91666666667\nCounties in state Wyoming have an average population of 24505.478260869564\nCounties in state Alabama have an average population of 71339.34328358209\nCounties in state Alaska have an average population of 24490.724137931036\nCounties in state Arizona have an average population of 426134.4666666667\nCounties in state Arkansas have an average population of 38878.90666666667\nCounties in state California have an average population of 642309.5862068966\nCounties in state Colorado have an average population of 78581.1875\nCounties in state Connecticut have an average population of 446762.125\nCounties in state Delaware have an average population of 299311.3333333333\nCounties in state District of Columbia have an average population of 601723.0\nCounties in state Florida have an average population of 280616.5671641791\nCounties in state Georgia have an average population of 60928.63522012578\nCounties in state Hawaii have an average population of 272060.2\nCounties in state Idaho have an average population of 35626.86363636364\nCounties in state Illinois have an average population of 125790.50980392157\nCounties in state Indiana have an average population of 70476.10869565218\nCounties in state Iowa have an average population of 30771.262626262625\nCounties in state Kansas have an average population of 27172.55238095238\nCounties in state Kentucky have an average population of 36161.39166666667\nCounties in state Louisiana have an average population of 70833.9375\nCounties in state Maine have an average population of 83022.5625\nCounties in state Maryland have an average population of 240564.66666666666\nCounties in state Massachusetts have an average population of 467687.78571428574\nCounties in state Michigan have an average population of 119080.0\nCounties in state Minnesota have an average population of 60964.65517241379\nCounties in state Mississippi have an average population of 36186.54878048781\nCounties in state Missouri have an average population of 52077.62608695652\nCounties in state Montana have an average population of 17668.125\nCounties in state Nebraska have an average population of 19638.075268817203\nCounties in state Nevada have an average population of 158855.9411764706\nCounties in state New Hampshire have an average population of 131647.0\nCounties in state New Jersey have an average population of 418661.61904761905\nCounties in state New Mexico have an average population of 62399.36363636364\nCounties in state New York have an average population of 312550.03225806454\nCounties in state North Carolina have an average population of 95354.83\nCounties in state North Dakota have an average population of 12690.396226415094\nCounties in state Ohio have an average population of 131096.63636363635\nCounties in state Oklahoma have an average population of 48718.844155844155\nCounties in state Oregon have an average population of 106418.72222222222\nCounties in state Pennsylvania have an average population of 189587.74626865672\nCounties in state Rhode Island have an average population of 210513.4\nCounties in state South Carolina have an average population of 100551.39130434782\nCounties in state South Dakota have an average population of 12336.060606060606\nCounties in state Tennessee have an average population of 66801.1052631579\nCounties in state Texas have an average population of 98998.27165354331\nCounties in state Utah have an average population of 95306.37931034483\nCounties in state Vermont have an average population of 44695.78571428572\nCounties in state Virginia have an average population of 60111.29323308271\nCounties in state Washington have an average population of 172424.10256410256\nCounties in state West Virginia have an average population of 33690.8\nCounties in state Wisconsin have an average population of 78985.91666666667\nCounties in state Wyoming have an average population of 24505.478260869564\n839 ms ± 41 ms per loop (mean ± std. dev. of 7 runs, 5 loops each)\n" ], [ "%%timeit -n 5\n\n# Using groupby is much quicker than iterating over the dataframe. groupby takes a col. name/names as an argument and splits the database based on those names\n\n\n\nfor group, frame in df.groupby('STNAME'):\n avg = np.average(frame['CENSUS2010POP'])\n print('Counties in state ' + group + ' have an average population of ' + str(avg))", "55\nCounties in state Oregon have an average population of 106418.72222222222\nCounties in state Pennsylvania have an average population of 189587.74626865672\nCounties in state Rhode Island have an average population of 210513.4\nCounties in state South Carolina have an average population of 100551.39130434782\nCounties in state South Dakota have an average population of 12336.060606060606\nCounties in state Tennessee have an average population of 66801.1052631579\nCounties in state Texas have an average population of 98998.27165354331\nCounties in state Utah have an average population of 95306.37931034483\nCounties in state Vermont have an average population of 44695.78571428572\nCounties in state Virginia have an average population of 60111.29323308271\nCounties in state Washington have an average population of 172424.10256410256\nCounties in state West Virginia have an average population of 33690.8\nCounties in state Wisconsin have an average population of 78985.91666666667\nCounties in state Wyoming have an average population of 24505.478260869564\nCounties in state Alabama have an average population of 71339.34328358209\nCounties in state Alaska have an average population of 24490.724137931036\nCounties in state Arizona have an average population of 426134.4666666667\nCounties in state Arkansas have an average population of 38878.90666666667\nCounties in state California have an average population of 642309.5862068966\nCounties in state Colorado have an average population of 78581.1875\nCounties in state Connecticut have an average population of 446762.125\nCounties in state Delaware have an average population of 299311.3333333333\nCounties in state District of Columbia have an average population of 601723.0\nCounties in state Florida have an average population of 280616.5671641791\nCounties in state Georgia have an average population of 60928.63522012578\nCounties in state Hawaii have an average population of 272060.2\nCounties in state Idaho have an average population of 35626.86363636364\nCounties in state Illinois have an average population of 125790.50980392157\nCounties in state Indiana have an average population of 70476.10869565218\nCounties in state Iowa have an average population of 30771.262626262625\nCounties in state Kansas have an average population of 27172.55238095238\nCounties in state Kentucky have an average population of 36161.39166666667\nCounties in state Louisiana have an average population of 70833.9375\nCounties in state Maine have an average population of 83022.5625\nCounties in state Maryland have an average population of 240564.66666666666\nCounties in state Massachusetts have an average population of 467687.78571428574\nCounties in state Michigan have an average population of 119080.0\nCounties in state Minnesota have an average population of 60964.65517241379\nCounties in state Mississippi have an average population of 36186.54878048781\nCounties in state Missouri have an average population of 52077.62608695652\nCounties in state Montana have an average population of 17668.125\nCounties in state Nebraska have an average population of 19638.075268817203\nCounties in state Nevada have an average population of 158855.9411764706\nCounties in state New Hampshire have an average population of 131647.0\nCounties in state New Jersey have an average population of 418661.61904761905\nCounties in state New Mexico have an average population of 62399.36363636364\nCounties in state New York have an average population of 312550.03225806454\nCounties in state North Carolina have an average population of 95354.83\nCounties in state North Dakota have an average population of 12690.396226415094\nCounties in state Ohio have an average population of 131096.63636363635\nCounties in state Oklahoma have an average population of 48718.844155844155\nCounties in state Oregon have an average population of 106418.72222222222\nCounties in state Pennsylvania have an average population of 189587.74626865672\nCounties in state Rhode Island have an average population of 210513.4\nCounties in state South Carolina have an average population of 100551.39130434782\nCounties in state South Dakota have an average population of 12336.060606060606\nCounties in state Tennessee have an average population of 66801.1052631579\nCounties in state Texas have an average population of 98998.27165354331\nCounties in state Utah have an average population of 95306.37931034483\nCounties in state Vermont have an average population of 44695.78571428572\nCounties in state Virginia have an average population of 60111.29323308271\nCounties in state Washington have an average population of 172424.10256410256\nCounties in state West Virginia have an average population of 33690.8\nCounties in state Wisconsin have an average population of 78985.91666666667\nCounties in state Wyoming have an average population of 24505.478260869564\nCounties in state Alabama have an average population of 71339.34328358209\nCounties in state Alaska have an average population of 24490.724137931036\nCounties in state Arizona have an average population of 426134.4666666667\nCounties in state Arkansas have an average population of 38878.90666666667\nCounties in state California have an average population of 642309.5862068966\nCounties in state Colorado have an average population of 78581.1875\nCounties in state Connecticut have an average population of 446762.125\nCounties in state Delaware have an average population of 299311.3333333333\nCounties in state District of Columbia have an average population of 601723.0\nCounties in state Florida have an average population of 280616.5671641791\nCounties in state Georgia have an average population of 60928.63522012578\nCounties in state Hawaii have an average population of 272060.2\nCounties in state Idaho have an average population of 35626.86363636364\nCounties in state Illinois have an average population of 125790.50980392157\nCounties in state Indiana have an average population of 70476.10869565218\nCounties in state Iowa have an average population of 30771.262626262625\nCounties in state Kansas have an average population of 27172.55238095238\nCounties in state Kentucky have an average population of 36161.39166666667\nCounties in state Louisiana have an average population of 70833.9375\nCounties in state Maine have an average population of 83022.5625\nCounties in state Maryland have an average population of 240564.66666666666\nCounties in state Massachusetts have an average population of 467687.78571428574\nCounties in state Michigan have an average population of 119080.0\nCounties in state Minnesota have an average population of 60964.65517241379\nCounties in state Mississippi have an average population of 36186.54878048781\nCounties in state Missouri have an average population of 52077.62608695652\nCounties in state Montana have an average population of 17668.125\nCounties in state Nebraska have an average population of 19638.075268817203\nCounties in state Nevada have an average population of 158855.9411764706\nCounties in state New Hampshire have an average population of 131647.0\nCounties in state New Jersey have an average population of 418661.61904761905\nCounties in state New Mexico have an average population of 62399.36363636364\nCounties in state New York have an average population of 312550.03225806454\nCounties in state North Carolina have an average population of 95354.83\nCounties in state North Dakota have an average population of 12690.396226415094\nCounties in state Ohio have an average population of 131096.63636363635\nCounties in state Oklahoma have an average population of 48718.844155844155\nCounties in state Oregon have an average population of 106418.72222222222\nCounties in state Pennsylvania have an average population of 189587.74626865672\nCounties in state Rhode Island have an average population of 210513.4\nCounties in state South Carolina have an average population of 100551.39130434782\nCounties in state South Dakota have an average population of 12336.060606060606\nCounties in state Tennessee have an average population of 66801.1052631579\nCounties in state Texas have an average population of 98998.27165354331\nCounties in state Utah have an average population of 95306.37931034483\nCounties in state Vermont have an average population of 44695.78571428572\nCounties in state Virginia have an average population of 60111.29323308271\nCounties in state Washington have an average population of 172424.10256410256\nCounties in state West Virginia have an average population of 33690.8\nCounties in state Wisconsin have an average population of 78985.91666666667\nCounties in state Wyoming have an average population of 24505.478260869564\nCounties in state Alabama have an average population of 71339.34328358209\nCounties in state Alaska have an average population of 24490.724137931036\nCounties in state Arizona have an average population of 426134.4666666667\nCounties in state Arkansas have an average population of 38878.90666666667\nCounties in state California have an average population of 642309.5862068966\nCounties in state Colorado have an average population of 78581.1875\nCounties in state Connecticut have an average population of 446762.125\nCounties in state Delaware have an average population of 299311.3333333333\nCounties in state District of Columbia have an average population of 601723.0\nCounties in state Florida have an average population of 280616.5671641791\nCounties in state Georgia have an average population of 60928.63522012578\nCounties in state Hawaii have an average population of 272060.2\nCounties in state Idaho have an average population of 35626.86363636364\nCounties in state Illinois have an average population of 125790.50980392157\nCounties in state Indiana have an average population of 70476.10869565218\nCounties in state Iowa have an average population of 30771.262626262625\nCounties in state Kansas have an average population of 27172.55238095238\nCounties in state Kentucky have an average population of 36161.39166666667\nCounties in state Louisiana have an average population of 70833.9375\nCounties in state Maine have an average population of 83022.5625\nCounties in state Maryland have an average population of 240564.66666666666\nCounties in state Massachusetts have an average population of 467687.78571428574\nCounties in state Michigan have an average population of 119080.0\nCounties in state Minnesota have an average population of 60964.65517241379\nCounties in state Mississippi have an average population of 36186.54878048781\nCounties in state Missouri have an average population of 52077.62608695652\nCounties in state Montana have an average population of 17668.125\nCounties in state Nebraska have an average population of 19638.075268817203\nCounties in state Nevada have an average population of 158855.9411764706\nCounties in state New Hampshire have an average population of 131647.0\nCounties in state New Jersey have an average population of 418661.61904761905\nCounties in state New Mexico have an average population of 62399.36363636364\nCounties in state New York have an average population of 312550.03225806454\nCounties in state North Carolina have an average population of 95354.83\nCounties in state North Dakota have an average population of 12690.396226415094\nCounties in state Ohio have an average population of 131096.63636363635\nCounties in state Oklahoma have an average population of 48718.844155844155\nCounties in state Oregon have an average population of 106418.72222222222\nCounties in state Pennsylvania have an average population of 189587.74626865672\nCounties in state Rhode Island have an average population of 210513.4\nCounties in state South Carolina have an average population of 100551.39130434782\nCounties in state South Dakota have an average population of 12336.060606060606\nCounties in state Tennessee have an average population of 66801.1052631579\nCounties in state Texas have an average population of 98998.27165354331\nCounties in state Utah have an average population of 95306.37931034483\nCounties in state Vermont have an average population of 44695.78571428572\nCounties in state Virginia have an average population of 60111.29323308271\nCounties in state Washington have an average population of 172424.10256410256\nCounties in state West Virginia have an average population of 33690.8\nCounties in state Wisconsin have an average population of 78985.91666666667\nCounties in state Wyoming have an average population of 24505.478260869564\nCounties in state Alabama have an average population of 71339.34328358209\nCounties in state Alaska have an average population of 24490.724137931036\nCounties in state Arizona have an average population of 426134.4666666667\nCounties in state Arkansas have an average population of 38878.90666666667\nCounties in state California have an average population of 642309.5862068966\nCounties in state Colorado have an average population of 78581.1875\nCounties in state Connecticut have an average population of 446762.125\nCounties in state Delaware have an average population of 299311.3333333333\nCounties in state District of Columbia have an average population of 601723.0\nCounties in state Florida have an average population of 280616.5671641791\nCounties in state Georgia have an average population of 60928.63522012578\nCounties in state Hawaii have an average population of 272060.2\nCounties in state Idaho have an average population of 35626.86363636364\nCounties in state Illinois have an average population of 125790.50980392157\nCounties in state Indiana have an average population of 70476.10869565218\nCounties in state Iowa have an average population of 30771.262626262625\nCounties in state Kansas have an average population of 27172.55238095238\nCounties in state Kentucky have an average population of 36161.39166666667\nCounties in state Louisiana have an average population of 70833.9375\nCounties in state Maine have an average population of 83022.5625\nCounties in state Maryland have an average population of 240564.66666666666\nCounties in state Massachusetts have an average population of 467687.78571428574\nCounties in state Michigan have an average population of 119080.0\nCounties in state Minnesota have an average population of 60964.65517241379\nCounties in state Mississippi have an average population of 36186.54878048781\nCounties in state Missouri have an average population of 52077.62608695652\nCounties in state Montana have an average population of 17668.125\nCounties in state Nebraska have an average population of 19638.075268817203\nCounties in state Nevada have an average population of 158855.9411764706\nCounties in state New Hampshire have an average population of 131647.0\nCounties in state New Jersey have an average population of 418661.61904761905\nCounties in state New Mexico have an average population of 62399.36363636364\nCounties in state New York have an average population of 312550.03225806454\nCounties in state North Carolina have an average population of 95354.83\nCounties in state North Dakota have an average population of 12690.396226415094\nCounties in state Ohio have an average population of 131096.63636363635\nCounties in state Oklahoma have an average population of 48718.844155844155\nCounties in state Oregon have an average population of 106418.72222222222\nCounties in state Pennsylvania have an average population of 189587.74626865672\nCounties in state Rhode Island have an average population of 210513.4\nCounties in state South Carolina have an average population of 100551.39130434782\nCounties in state South Dakota have an average population of 12336.060606060606\nCounties in state Tennessee have an average population of 66801.1052631579\nCounties in state Texas have an average population of 98998.27165354331\nCounties in state Utah have an average population of 95306.37931034483\nCounties in state Vermont have an average population of 44695.78571428572\nCounties in state Virginia have an average population of 60111.29323308271\nCounties in state Washington have an average population of 172424.10256410256\nCounties in state West Virginia have an average population of 33690.8\nCounties in state Wisconsin have an average population of 78985.91666666667\nCounties in state Wyoming have an average population of 24505.478260869564\nCounties in state Alabama have an average population of 71339.34328358209\nCounties in state Alaska have an average population of 24490.724137931036\nCounties in state Arizona have an average population of 426134.4666666667\nCounties in state Arkansas have an average population of 38878.90666666667\nCounties in state California have an average population of 642309.5862068966\nCounties in state Colorado have an average population of 78581.1875\nCounties in state Connecticut have an average population of 446762.125\nCounties in state Delaware have an average population of 299311.3333333333\nCounties in state District of Columbia have an average population of 601723.0\nCounties in state Florida have an average population of 280616.5671641791\nCounties in state Georgia have an average population of 60928.63522012578\nCounties in state Hawaii have an average population of 272060.2\nCounties in state Idaho have an average population of 35626.86363636364\nCounties in state Illinois have an average population of 125790.50980392157\nCounties in state Indiana have an average population of 70476.10869565218\nCounties in state Iowa have an average population of 30771.262626262625\nCounties in state Kansas have an average population of 27172.55238095238\nCounties in state Kentucky have an average population of 36161.39166666667\nCounties in state Louisiana have an average population of 70833.9375\nCounties in state Maine have an average population of 83022.5625\nCounties in state Maryland have an average population of 240564.66666666666\nCounties in state Massachusetts have an average population of 467687.78571428574\nCounties in state Michigan have an average population of 119080.0\nCounties in state Minnesota have an average population of 60964.65517241379\nCounties in state Mississippi have an average population of 36186.54878048781\nCounties in state Missouri have an average population of 52077.62608695652\nCounties in state Montana have an average population of 17668.125\nCounties in state Nebraska have an average population of 19638.075268817203\nCounties in state Nevada have an average population of 158855.9411764706\nCounties in state New Hampshire have an average population of 131647.0\nCounties in state New Jersey have an average population of 418661.61904761905\nCounties in state New Mexico have an average population of 62399.36363636364\nCounties in state New York have an average population of 312550.03225806454\nCounties in state North Carolina have an average population of 95354.83\nCounties in state North Dakota have an average population of 12690.396226415094\nCounties in state Ohio have an average population of 131096.63636363635\nCounties in state Oklahoma have an average population of 48718.844155844155\nCounties in state Oregon have an average population of 106418.72222222222\nCounties in state Pennsylvania have an average population of 189587.74626865672\nCounties in state Rhode Island have an average population of 210513.4\nCounties in state South Carolina have an average population of 100551.39130434782\nCounties in state South Dakota have an average population of 12336.060606060606\nCounties in state Tennessee have an average population of 66801.1052631579\nCounties in state Texas have an average population of 98998.27165354331\nCounties in state Utah have an average population of 95306.37931034483\nCounties in state Vermont have an average population of 44695.78571428572\nCounties in state Virginia have an average population of 60111.29323308271\nCounties in state Washington have an average population of 172424.10256410256\nCounties in state West Virginia have an average population of 33690.8\nCounties in state Wisconsin have an average population of 78985.91666666667\nCounties in state Wyoming have an average population of 24505.478260869564\n34.7 ms ± 8.65 ms per loop (mean ± std. dev. of 7 runs, 5 loops each)\n" ], [ "df.head()", "_____no_output_____" ], [ "# groupby can also be passed functions that are used for segmenting data\n\ndf = df.set_index('STNAME')\n\ndef fun(item):\n if item[0] < 'M':\n return 0\n if item[0] > 'Q':\n return 1\n else:\n return 2\n\n# lightweight hashing which distributes tasks across multiple workers, eg. cores in a processor, nodes in a supercomputer or disks in a database \n\nfor group, frame in df.groupby(fun):\n print('There are ' + str(len(frame)) + ' records in group ' + str(group) + ' for processing.')\n ", "There are 1177 records in group 0 for processing.\nThere are 831 records in group 1 for processing.\nThere are 1134 records in group 2 for processing.\n" ], [ "df = pd.read_csv('census.csv')\ndf = df[df['SUMLEV'] == 50]", "_____no_output_____" ], [ "# groupby's agg method applies a fucntion to the column/columns in a the group and returns the result. \n\ndf.groupby('STNAME').agg({'CENSUS2010POP': np.average})", "_____no_output_____" ], [ "print(type(df.groupby(level=0)['POPESTIMATE2010', 'POPESTIMATE2011']))\nprint(type(df.groupby(level=0)['POPESTIMATE2010']))", "<class 'pandas.core.groupby.generic.DataFrameGroupBy'>\n<class 'pandas.core.groupby.generic.SeriesGroupBy'>\n" ] ], [ [ "groupby.agg() with a dictionary when renaming has been deprecated (https://pandas.pydata.org/pandas-docs/stable/whatsnew/v0.20.0.html#whatsnew-0200-api-breaking-deprecate-group-agg-dict)", "_____no_output_____" ] ], [ [ "(df.set_index('STNAME').groupby(level=0)['POPESTIMATE2010','POPESTIMATE2011']\n .agg(np.sum, np.average).rename(columns={'POPESTIMATE2010': 'avg'}))", "_____no_output_____" ], [ "df.groupby('STNAME').agg({'CENSUS2010POP' : [np.average , np.sum]})", "_____no_output_____" ], [ "(df.set_index('STNAME').groupby(level=0)['POPESTIMATE2010','POPESTIMATE2011']\n .agg({'POPESTIMATE2010': np.average, 'POPESTIMATE2011': np.sum}))", "_____no_output_____" ] ], [ [ "# Data Scales\n\n## (a,b)(c,d): Scales\n\nAs a data scientist, there are four scales worth knowing:\n\n* Ratio scale\n * units are equally spaced\n * mathematical operations of +-/* are all valid\n * eg. height and weight\n\n* Interval scale\n * units are equally spaced, but there is no true zero\n * eg. temperature, since there is never an abscence of temperature, where 0°C is a valid value\n\n* Ordinal scale\n * the order of the units is important, but not evenly spaced\n * letter grades such as A, A+ are a good example\n\n* Nominal scale\n * categories of data, but the categories have no order with respect to one another\n * categories where there are only two possible values are binary\n * eg. teams of a sport\n\nNominal data in pandas is called categorical data. pandas has a built-in type for categorical data and we can set a column of our data to categorical by using the astype method. astype changes the underlying type of the data, in this case to category data.\n", "_____no_output_____" ] ], [ [ "df = pd.DataFrame(['A+', 'A', 'A-', 'B+', 'B', 'B-', 'C+', 'C', 'C-', 'D+', 'D'],\n index=['excellent', 'excellent', 'excellent', 'good', 'good', 'good', 'ok', 'ok', 'ok', 'poor', 'poor'])\n\ndf.rename(columns={0:'Grades'}, inplace=True)\ndf", "_____no_output_____" ], [ "# set the data type as category using astype\ndf['Grades'].astype('category').head()", "_____no_output_____" ], [ "# astype categories parameter deprecated, use pd.api.types.CategoricalDtype instead\n\ngrades = pd.DataFrame(['A+', 'A', 'A-', 'B+', 'B', 'B-', 'C+', 'C', 'C-', 'D+', 'D'],\n index=['excellent', 'excellent', 'excellent', 'good', 'good', 'good', 'ok', 'ok', 'ok', 'poor', 'poor']).astype(pd.api.types.CategoricalDtype(categories=['D', 'D+', 'C-', 'C', 'C+', 'B-', 'B', 'B+', 'A-', 'A', 'A+'], ordered = True))\n\ngrades.rename(columns={0: 'Grades'}, inplace = True)", "_____no_output_____" ], [ "grades.head()", "_____no_output_____" ], [ "grades > 'C'", "_____no_output_____" ], [ "s = pd.Series(['Low', 'Low', 'High', 'Medium', 'Low', 'High', 'Low'])\ns.astype(pd.api.types.CategoricalDtype(categories = ['Low', 'Medium', 'High'], ordered = True))", "_____no_output_____" ], [ "# get_dummies\n\ns = pd.Series(list('abca'))\ns", "_____no_output_____" ], [ "pd.get_dummies(s)", "_____no_output_____" ], [ "help(pd.get_dummies)", "Help on function get_dummies in module pandas.core.reshape.reshape:\n\nget_dummies(data, prefix=None, prefix_sep='_', dummy_na=False, columns=None, sparse=False, drop_first=False, dtype=None) -> 'DataFrame'\n Convert categorical variable into dummy/indicator variables.\n \n Parameters\n ----------\n data : array-like, Series, or DataFrame\n Data of which to get dummy indicators.\n prefix : str, list of str, or dict of str, default None\n String to append DataFrame column names.\n Pass a list with length equal to the number of columns\n when calling get_dummies on a DataFrame. Alternatively, `prefix`\n can be a dictionary mapping column names to prefixes.\n prefix_sep : str, default '_'\n If appending prefix, separator/delimiter to use. Or pass a\n list or dictionary as with `prefix`.\n dummy_na : bool, default False\n Add a column to indicate NaNs, if False NaNs are ignored.\n columns : list-like, default None\n Column names in the DataFrame to be encoded.\n If `columns` is None then all the columns with\n `object` or `category` dtype will be converted.\n sparse : bool, default False\n Whether the dummy-encoded columns should be backed by\n a :class:`SparseArray` (True) or a regular NumPy array (False).\n drop_first : bool, default False\n Whether to get k-1 dummies out of k categorical levels by removing the\n first level.\n dtype : dtype, default np.uint8\n Data type for new columns. Only a single dtype is allowed.\n \n .. versionadded:: 0.23.0\n \n Returns\n -------\n DataFrame\n Dummy-coded data.\n \n See Also\n --------\n Series.str.get_dummies : Convert Series to dummy codes.\n \n Examples\n --------\n >>> s = pd.Series(list('abca'))\n \n >>> pd.get_dummies(s)\n a b c\n 0 1 0 0\n 1 0 1 0\n 2 0 0 1\n 3 1 0 0\n \n >>> s1 = ['a', 'b', np.nan]\n \n >>> pd.get_dummies(s1)\n a b\n 0 1 0\n 1 0 1\n 2 0 0\n \n >>> pd.get_dummies(s1, dummy_na=True)\n a b NaN\n 0 1 0 0\n 1 0 1 0\n 2 0 0 1\n \n >>> df = pd.DataFrame({'A': ['a', 'b', 'a'], 'B': ['b', 'a', 'c'],\n ... 'C': [1, 2, 3]})\n \n >>> pd.get_dummies(df, prefix=['col1', 'col2'])\n C col1_a col1_b col2_a col2_b col2_c\n 0 1 1 0 0 1 0\n 1 2 0 1 1 0 0\n 2 3 1 0 0 0 1\n \n >>> pd.get_dummies(pd.Series(list('abcaa')))\n a b c\n 0 1 0 0\n 1 0 1 0\n 2 0 0 1\n 3 1 0 0\n 4 1 0 0\n \n >>> pd.get_dummies(pd.Series(list('abcaa')), drop_first=True)\n b c\n 0 0 0\n 1 1 0\n 2 0 1\n 3 0 0\n 4 0 0\n \n >>> pd.get_dummies(pd.Series(list('abc')), dtype=float)\n a b c\n 0 1.0 0.0 0.0\n 1 0.0 1.0 0.0\n 2 0.0 0.0 1.0\n\n" ] ], [ [ "# Cut", "_____no_output_____" ] ], [ [ "df = pd.read_csv('census.csv')\ndf = df[df['SUMLEV']==50]\n\ndf", "_____no_output_____" ], [ "df = df.set_index('STNAME').groupby(level=0)['CENSUS2010POP'].agg([np.average])", "_____no_output_____" ], [ "# Using cut\n\ndf.rename(columns={'average' : 'avg_pop'}, inplace = True)\npd.cut(df['avg_pop'], 10)", "_____no_output_____" ], [ "s = pd.Series([168, 180, 174, 190, 170, 185, 179, 181, 175, 169, 182, 177, 180, 171])\n\npd.cut(s, 3, labels = ['Small', 'Medium', 'Large'])", "_____no_output_____" ] ], [ [ "# Pivot Tables", "_____no_output_____" ] ], [ [ "df = pd.read_csv('cars.csv')", "_____no_output_____" ], [ "df.head()", "_____no_output_____" ], [ "df.pivot_table(values='(kW)', index='YEAR', columns='Make', aggfunc=np.mean)", "_____no_output_____" ], [ "# aggfunc can also be passed a list of functions, and pandas will provide the result using hierarchical column names\n# we also pass margins = True, so for each category now there is an all category, which shows the overall mean and min values for each year for a given year and vendor\n# \ndf.pivot_table(values='(kW)', index='YEAR', columns='Make', aggfunc=[np.mean, np.min], margins = True)", "_____no_output_____" ], [ "import pandas as pd, os\ndf = pd.read_csv('mpg.csv')", "_____no_output_____" ], [ "df.head()", "_____no_output_____" ], [ "df", "_____no_output_____" ], [ "# see the cty and hwy mpg values for each car made by each manufacturer in every year in the dataset\ndf.pivot_table(values = ['cty', 'hwy'], index = ['manufacturer', 'class', 'model'], columns = 'year')", "_____no_output_____" ] ], [ [ "# Date Functionality in pandas", "_____no_output_____" ] ], [ [ "import pandas as pd\nimport numpy as np", "_____no_output_____" ] ], [ [ "### Timestamp", "_____no_output_____" ] ], [ [ "pd.Timestamp('07/20/2020 10:05 AM')", "_____no_output_____" ] ], [ [ "### Period", "_____no_output_____" ] ], [ [ "pd.Period('1/2020')", "_____no_output_____" ], [ "pd.Period('3/1/2020')", "_____no_output_____" ] ], [ [ "### DatetimeIndex", "_____no_output_____" ] ], [ [ "t1 = pd.Series(list('abc'), [pd.Timestamp('2020-07-18'), pd.Timestamp('2020-07-19'), pd.Timestamp('2020-07-20') ])\nt1", "_____no_output_____" ], [ "type(t1.index)", "_____no_output_____" ] ], [ [ "### Period Index", "_____no_output_____" ] ], [ [ "t2 = pd.Series(list('def'), [pd.Period('2020-04'), pd.Period('2020-05'), pd.Period('2020-06') ])\nt2", "_____no_output_____" ], [ "type(t2.index)", "_____no_output_____" ] ], [ [ "### Converting to Datetime", "_____no_output_____" ] ], [ [ "d1 = ['2 June 2013', 'Aug 29, 2014', '2015-06-26', '7/12/16']\nt3 = pd.DataFrame(np.random.randint(10, 100, (4, 2)), index = d1, columns=list('ab'))\nt3", "_____no_output_____" ], [ "t3.index = pd.to_datetime(t3.index)\nt3", "_____no_output_____" ], [ "pd.to_datetime('4.7.12', dayfirst=True)", "_____no_output_____" ] ], [ [ "### Timedeltas", "_____no_output_____" ] ], [ [ "pd.Timestamp('9/3/2016') - pd.Timestamp('9/1/2016')", "_____no_output_____" ], [ "pd.Timestamp('9/2/2016 8:10AM') + pd.Timedelta('12D 3H')", "_____no_output_____" ] ], [ [ "### Working with dates in a dataframe", "_____no_output_____" ] ], [ [ "dates = pd.date_range('10-01-2016', periods = 9, freq='2W-SUN')\ndates", "_____no_output_____" ], [ "df = pd.DataFrame({'Count 1': 100 + np.random.randint(-5, 10, 9).cumsum(),\n 'Count 2': 120 + np.random.randint(-5, 10, 9)}, index=dates)\ndf", "_____no_output_____" ], [ "# cumsum(): returns the cumulative sum of elements along a given axis\n\na = np.random.randint(1, 10, 9)\nprint(a)\nb = a.cumsum()\nprint(b)", "[1 2 3 3 1 6 5 4 1]\n[ 1 3 6 9 10 16 21 25 26]\n" ], [ "# weekday_name deprecated, use day_name() instead\ndf.index.day_name()", "_____no_output_____" ], [ "# diff(): calculates the difference of a DataFrame elements compared with another element in the DatFrame. defualts to the element of the previous row in the same column \ndf.diff()", "_____no_output_____" ], [ "# resample(): method for frequency conversion and resampling of a time series. Object must have a datetime index.\n\n# DateOffset objects - https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#dateoffset-objects\n\ndf.resample('M').mean()", "_____no_output_____" ], [ "# Using partial string indexing\n\ndf['2017']", "_____no_output_____" ], [ "df['2016-12']", "_____no_output_____" ], [ "df['2016-12':]", "_____no_output_____" ], [ "df.asfreq('W', method='ffill')", "_____no_output_____" ], [ "import matplotlib.pyplot as plt\n%matplotlib inline\n\ndf.plot()", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code", "code", "code" ] ]

4ad9637727e946928f8ec325ecd5caf847225dd7

439,719

ipynb

Jupyter Notebook

_source/raw/movielens_1m_wordcloud.ipynb

sparsh-ai/reco-tutorials

7be837ca7105424aaf43148b334dc9d2e0e66368

[ "Apache-2.0" ]

3

2021-08-29T13:18:25.000Z

2022-03-08T19:48:32.000Z

code/movielens_1m_wordcloud.ipynb

sparsh-ai/recsys-colab

c0aa0dceca5a4d8ecd42b61c4e906035fe1614f3

[ "MIT" ]

null

code/movielens_1m_wordcloud.ipynb

sparsh-ai/recsys-colab

c0aa0dceca5a4d8ecd42b61c4e906035fe1614f3

[ "MIT" ]

4

2021-06-16T03:07:10.000Z

2022-03-26T04:22:04.000Z

433.647929

403,308

0.921882

[ [ [ "import pandas as pd\r\nimport numpy as np\r\nimport matplotlib.pyplot as plt", "_____no_output_____" ], [ "base_path = './data/ML-1M/'\r\n\r\nratings = pd.read_csv(base_path+'ratings.csv', sep='\\t', \r\n encoding='latin-1', \r\n usecols=['user_id', 'movie_id', 'rating'])\r\n\r\nusers = pd.read_csv(base_path+'users.csv', sep='\\t', \r\n encoding='latin-1', \r\n usecols=['user_id', 'gender', 'zipcode', \r\n 'age_desc', 'occ_desc'])\r\n\r\nmovies = pd.read_csv(base_path+'movies.csv', sep='\\t', \r\n encoding='latin-1', \r\n usecols=['movie_id', 'title', 'genres'])", "_____no_output_____" ], [ "ratings.head()", "_____no_output_____" ], [ "users.head()", "_____no_output_____" ], [ "movies.head()", "_____no_output_____" ] ], [ [ "Plot the wordcloud", "_____no_output_____" ] ], [ [ "%matplotlib inline\r\nimport wordcloud\r\nfrom wordcloud import WordCloud, STOPWORDS\r\n\r\n# Create a wordcloud of the movie titles\r\nmovies['title'] = movies['title'].fillna(\"\").astype('str')\r\ntitle_corpus = ' '.join(movies['title'])\r\ntitle_wordcloud = WordCloud(stopwords=STOPWORDS, background_color='black', height=2000, width=4000).generate(title_corpus)\r\n\r\n# Plot the wordcloud\r\nplt.figure(figsize=(16,8))\r\nplt.imshow(title_wordcloud)\r\nplt.axis('off')\r\nplt.show()", "_____no_output_____" ] ], [ [ "Genre-based recommendations", "_____no_output_____" ] ], [ [ "# Import libraries\r\nfrom sklearn.feature_extraction.text import TfidfVectorizer\r\nfrom sklearn.metrics.pairwise import linear_kernel\r\n\r\n# Break up the big genre string into a string array\r\nmovies['genres'] = movies['genres'].str.split('|')\r\n\r\n# Convert genres to string value\r\nmovies['genres'] = movies['genres'].fillna(\"\").astype('str')\r\n\r\n# Movie feature vector\r\ntf = TfidfVectorizer(analyzer='word', ngram_range=(1, 2), min_df=0, \r\n stop_words='english')\r\ntfidf_matrix = tf.fit_transform(movies['genres'])\r\n\r\n# Movie similarity matrix\r\ncosine_sim = linear_kernel(tfidf_matrix, tfidf_matrix)\r\n\r\n# 1-d array of movie titles\r\ntitles = movies['title']\r\nindices = pd.Series(movies.index, index=movies['title'])\r\n\r\n# Function to return top-k most similar movies\r\ndef genre_recommendations(title, topk=20):\r\n idx = indices[title]\r\n sim_scores = list(enumerate(cosine_sim[idx]))\r\n sim_scores = sorted(sim_scores, key=lambda x: x[1], reverse=True)\r\n sim_scores = sim_scores[1:topk+1]\r\n movie_indices = [i[0] for i in sim_scores]\r\n return titles.iloc[movie_indices].reset_index(drop=True)", "_____no_output_____" ], [ "# Checkout the results\r\n# genre_recommendations('Good Will Hunting (1997)')\r\ngenre_recommendations('Toy Story (1995)')\r\n# genre_recommendations('Saving Private Ryan (1998)')", "_____no_output_____" ] ], [ [ "Simple collaborative filtering", "_____no_output_____" ] ], [ [ "from math import sqrt\r\nfrom sklearn.metrics import mean_squared_error\r\nfrom sklearn.model_selection import train_test_split\r\nfrom sklearn.metrics.pairwise import pairwise_distances\r\n\r\n# Fill NaN values in user_id and movie_id column with 0\r\nratings['user_id'] = ratings['user_id'].fillna(0)\r\nratings['movie_id'] = ratings['movie_id'].fillna(0)\r\n\r\n# Replace NaN values in rating column with average of all values\r\nratings['rating'] = ratings['rating'].fillna(ratings['rating'].mean())\r\n\r\n# Randomly sample 1% for faster processing\r\nsmall_data = ratings.sample(frac=0.01)\r\n\r\n# Split into train and test\r\ntrain_data, test_data = train_test_split(small_data, test_size=0.2)\r\n\r\n# Create two user-item matrices, one for training and another for testing\r\ntrain_data_matrix = train_data.pivot(index='user_id', columns='movie_id', values='rating').fillna(0)\r\ntest_data_matrix = test_data.pivot(index='user_id', columns='movie_id', values='rating').fillna(0)\r\n\r\n# Create user similarity using Pearson correlation\r\nuser_correlation = 1 - pairwise_distances(train_data_matrix, metric='correlation')\r\nuser_correlation[np.isnan(user_correlation)] = 0\r\n\r\n# Create item similarity using Pearson correlation\r\nitem_correlation = 1 - pairwise_distances(train_data_matrix.T, metric='correlation')\r\nitem_correlation[np.isnan(item_correlation)] = 0\r\n\r\n# Function to predict ratings\r\ndef predict(ratings, similarity, type='user'):\r\n if type == 'user':\r\n mean_user_rating = ratings.mean(axis=1)\r\n # Use np.newaxis so that mean_user_rating has same format as ratings\r\n ratings_diff = (ratings - mean_user_rating.values[:, np.newaxis])\r\n pred = mean_user_rating.values[:, np.newaxis] + similarity.dot(ratings_diff) / np.array([np.abs(similarity).sum(axis=1)]).T\r\n elif type == 'item':\r\n pred = ratings.dot(similarity) / np.array([np.abs(similarity).sum(axis=1)])\r\n return pred\r\n\r\n# Function to calculate RMSE\r\ndef rmse(pred, actual):\r\n # Ignore nonzero terms.\r\n pred = pd.DataFrame(pred).values\r\n actual = actual.values\r\n pred = pred[actual.nonzero()].flatten()\r\n actual = actual[actual.nonzero()].flatten()\r\n return sqrt(mean_squared_error(pred, actual))\r\n\r\n# Predict ratings on the training data with both similarity score\r\nuser_prediction = predict(train_data_matrix, user_correlation, type='user')\r\nitem_prediction = predict(train_data_matrix, item_correlation, type='item')\r\n\r\n# RMSE on the train data\r\nprint('User-based CF RMSE Train: ' + str(rmse(user_prediction, train_data_matrix)))\r\nprint('Item-based CF RMSE Train: ' + str(rmse(item_prediction, train_data_matrix)))\r\n\r\n# RMSE on the test data\r\nprint('User-based CF RMSE Test: ' + str(rmse(user_prediction, test_data_matrix)))\r\nprint('Item-based CF RMSE Test: ' + str(rmse(item_prediction, test_data_matrix)))", "User-based CF RMSE Train: 2.6156258817338545\nItem-based CF RMSE Train: 2.8395461928591756\nUser-based CF RMSE Test: 3.740786194256653\nItem-based CF RMSE Test: 3.7432481961974267\n" ] ], [ [ "SVD matrix factorization based collaborative filtering", "_____no_output_____" ] ], [ [ "!pip install surprise", "Collecting surprise\n Downloading https://files.pythonhosted.org/packages/61/de/e5cba8682201fcf9c3719a6fdda95693468ed061945493dea2dd37c5618b/surprise-0.1-py2.py3-none-any.whl\nCollecting scikit-surprise\n\u001b[?25l Downloading https://files.pythonhosted.org/packages/97/37/5d334adaf5ddd65da99fc65f6507e0e4599d092ba048f4302fe8775619e8/scikit-surprise-1.1.1.tar.gz (11.8MB)\n\u001b[K |████████████████████████████████| 11.8MB 8.0MB/s \n\u001b[?25hRequirement already satisfied: joblib>=0.11 in /usr/local/lib/python3.6/dist-packages (from scikit-surprise->surprise) (1.0.0)\nRequirement already satisfied: numpy>=1.11.2 in /usr/local/lib/python3.6/dist-packages (from scikit-surprise->surprise) (1.19.4)\nRequirement already satisfied: scipy>=1.0.0 in /usr/local/lib/python3.6/dist-packages (from scikit-surprise->surprise) (1.4.1)\nRequirement already satisfied: six>=1.10.0 in /usr/local/lib/python3.6/dist-packages (from scikit-surprise->surprise) (1.15.0)\nBuilding wheels for collected packages: scikit-surprise\n Building wheel for scikit-surprise (setup.py) ... \u001b[?25l\u001b[?25hdone\n Created wheel for scikit-surprise: filename=scikit_surprise-1.1.1-cp36-cp36m-linux_x86_64.whl size=1618266 sha256=8ea4c56743b89b02a764481272bfcb2f25f76e68574c6d695397e9579c503de3\n Stored in directory: /root/.cache/pip/wheels/78/9c/3d/41b419c9d2aff5b6e2b4c0fc8d25c538202834058f9ed110d0\nSuccessfully built scikit-surprise\nInstalling collected packages: scikit-surprise, surprise\nSuccessfully installed scikit-surprise-1.1.1 surprise-0.1\n" ], [ "from scipy.sparse.linalg import svds\r\n\r\n# Create the interaction matrix\r\ninteractions = ratings.pivot(index='user_id', columns='movie_id', values='rating').fillna(0)\r\nprint(pd.DataFrame(interactions.values).head())\r\n\r\n# De-normalize the data (normalize by each users mean)\r\nuser_ratings_mean = np.mean(interactions.values, axis=1)\r\ninteractions_normalized = interactions.values - user_ratings_mean.reshape(-1, 1)\r\nprint(pd.DataFrame(interactions_normalized).head())\r\n\r\n# Calculating SVD\r\nU, sigma, Vt = svds(interactions_normalized, k=50)\r\nsigma = np.diag(sigma)\r\n\r\n# Make predictions from the decomposed matrix by matrix multiply U, Σ, and VT \r\n# back to get the rank k=50 approximation of A.\r\nall_user_predicted_ratings = np.dot(np.dot(U, sigma), Vt) + user_ratings_mean.reshape(-1, 1)\r\npreds = pd.DataFrame(all_user_predicted_ratings, columns=interactions.columns)\r\nprint(preds.head().values)\r\n\r\n# Get the movie with the highest predicted rating\r\ndef recommend_movies(predictions, userID, movies, original_ratings, num_recommendations):\r\n \r\n # Get and sort the user's predictions\r\n user_row_number = userID - 1 # User ID starts at 1, not 0\r\n sorted_user_predictions = preds.iloc[user_row_number].sort_values(ascending=False) # User ID starts at 1\r\n \r\n # Get the user's data and merge in the movie information.\r\n user_data = original_ratings[original_ratings.user_id == (userID)]\r\n user_full = (user_data.merge(movies, how = 'left', left_on = 'movie_id', right_on = 'movie_id').\r\n sort_values(['rating'], ascending=False)\r\n )\r\n\r\n print('User {0} has already rated {1} movies.'.format(userID, user_full.shape[0]))\r\n print('Recommending highest {0} predicted ratings movies not already rated.'.format(num_recommendations))\r\n \r\n # Recommend the highest predicted rating movies that the user hasn't seen yet.\r\n recommendations = (movies[~movies['movie_id'].isin(user_full['movie_id'])].\r\n merge(pd.DataFrame(sorted_user_predictions).reset_index(), how = 'left',\r\n left_on = 'movie_id',\r\n right_on = 'movie_id').\r\n rename(columns = {user_row_number: 'Predictions'}).\r\n sort_values('Predictions', ascending = False).\r\n iloc[:num_recommendations, :-1]\r\n )\r\n\r\n return user_full, recommendations\r\n\r\n# Let's try to recommend 20 movies for user with ID 1310\r\nalready_rated, predictions = recommend_movies(preds, 1310, movies, ratings, 20)\r\n\r\n# Top 20 movies that User 1310 has rated \r\nprint(already_rated.head(20))\r\n\r\n# Top 20 movies that User 1310 hopefully will enjoy\r\nprint(predictions)", " 0 1 2 3 4 5 ... 3700 3701 3702 3703 3704 3705\n0 5.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0\n1 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0\n2 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0\n3 0.0 0.0 0.0 0.0 0.0 0.0 ... 0.0 0.0 0.0 0.0 0.0 0.0\n4 0.0 0.0 0.0 0.0 0.0 2.0 ... 0.0 0.0 0.0 0.0 0.0 0.0\n\n[5 rows x 3706 columns]\n 0 1 2 ... 3703 3704 3705\n0 4.940097 -0.059903 -0.059903 ... -0.059903 -0.059903 -0.059903\n1 -0.129250 -0.129250 -0.129250 ... -0.129250 -0.129250 -0.129250\n2 -0.053697 -0.053697 -0.053697 ... -0.053697 -0.053697 -0.053697\n3 -0.023745 -0.023745 -0.023745 ... -0.023745 -0.023745 -0.023745\n4 -0.168106 -0.168106 -0.168106 ... -0.168106 -0.168106 -0.168106\n\n[5 rows x 3706 columns]\n[[ 4.28886061 0.14305516 -0.1950795 ... 0.03191195 0.05044975\n 0.08891033]\n [ 0.74471587 0.16965927 0.33541808 ... -0.10110207 -0.0540982\n -0.14018846]\n [ 1.81882382 0.45613623 0.09097801 ... 0.01234452 0.01514752\n -0.10995596]\n [ 0.40805697 -0.07296018 0.03964241 ... -0.02605009 0.01484101\n -0.03422403]\n [ 1.57427245 0.02123904 -0.05129994 ... 0.03383015 0.12570625\n 0.19924413]]\nUser 1310 has already rated 24 movies.\nRecommending highest 20 predicted ratings movies not already rated.\n user_id ... genres\n5 1310 ... ['[\\'[\"[\\\\\\'Comedy\\\\\\', \\\\\\'Drama\\\\\\', \\\\\\'Rom...\n6 1310 ... ['[\\'[\"[\\\\\\'Drama\\\\\\', \\\\\\'Romance\\\\\\']\"]\\']']\n7 1310 ... ['[\\'[\"[\\\\\\'Drama\\\\\\', \\\\\\'Film-Noir\\\\\\']\"]\\']']\n15 1310 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n1 1310 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n12 1310 ... ['[\\'[\"[\\\\\\'Thriller\\\\\\']\"]\\']']\n11 1310 ... ['[\\'[\"[\\\\\\'Thriller\\\\\\']\"]\\']']\n20 1310 ... ['[\\'[\"[\\\\\\'Action\\\\\\', \\\\\\'Comedy\\\\\\', \\\\\\'Cr...\n18 1310 ... ['[\\'[\"[\\\\\\'Comedy\\\\\\', \\\\\\'Drama\\\\\\']\"]\\']']\n17 1310 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n13 1310 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n23 1310 ... ['[\\'[\\\\\\'[\"Children\\\\\\\\\\\\\\'s\", \\\\\\\\\\\\\\'Drama\\...\n10 1310 ... ['[\\'[\"[\\\\\\'Action\\\\\\', \\\\\\'Adventure\\\\\\', \\\\\\...\n9 1310 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n8 1310 ... ['[\\'[\"[\\\\\\'Comedy\\\\\\', \\\\\\'Drama\\\\\\', \\\\\\'Rom...\n4 1310 ... ['[\\'[\"[\\\\\\'Comedy\\\\\\', \\\\\\'Drama\\\\\\', \\\\\\'Rom...\n3 1310 ... ['[\\'[\"[\\\\\\'Drama\\\\\\', \\\\\\'War\\\\\\']\"]\\']']\n16 1310 ... ['[\\'[\"[\\\\\\'Comedy\\\\\\']\"]\\']']\n19 1310 ... ['[\\'[\"[\\\\\\'Drama\\\\\\', \\\\\\'War\\\\\\']\"]\\']']\n0 1310 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n\n[20 rows x 5 columns]\n movie_id ... genres\n1618 1674 ... ['[\\'[\"[\\\\\\'Drama\\\\\\', \\\\\\'Romance\\\\\\', \\\\\\'Th...\n1880 1961 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n1187 1210 ... ['[\\'[\"[\\\\\\'Action\\\\\\', \\\\\\'Adventure\\\\\\', \\\\\\...\n1216 1242 ... ['[\\'[\"[\\\\\\'Action\\\\\\', \\\\\\'Drama\\\\\\', \\\\\\'War...\n1202 1225 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n1273 1302 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n1220 1246 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n1881 1962 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n1877 1957 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n1938 2020 ... ['[\\'[\"[\\\\\\'Drama\\\\\\', \\\\\\'Romance\\\\\\']\"]\\']']\n1233 1259 ... ['[\\'[\"[\\\\\\'Adventure\\\\\\', \\\\\\'Comedy\\\\\\', \\\\\\...\n3011 3098 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n2112 2194 ... ['[\\'[\"[\\\\\\'Action\\\\\\', \\\\\\'Crime\\\\\\', \\\\\\'Dra...\n1876 1956 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n1268 1296 ... ['[\\'[\"[\\\\\\'Drama\\\\\\', \\\\\\'Romance\\\\\\']\"]\\']']\n2267 2352 ... ['[\\'[\"[\\\\\\'Comedy\\\\\\', \\\\\\'Drama\\\\\\']\"]\\']']\n1278 1307 ... ['[\\'[\"[\\\\\\'Comedy\\\\\\', \\\\\\'Romance\\\\\\']\"]\\']']\n1165 1186 ... ['[\\'[\"[\\\\\\'Drama\\\\\\']\"]\\']']\n1199 1222 ... ['[\\'[\"[\\\\\\'Action\\\\\\', \\\\\\'Drama\\\\\\', \\\\\\'War...\n2833 2919 ... ['[\\'[\"[\\\\\\'Drama\\\\\\', \\\\\\'Romance\\\\\\']\"]\\']']\n\n[20 rows x 3 columns]\n" ], [ "from surprise import Reader, Dataset, SVD\r\nfrom surprise.model_selection import cross_validate\r\n\r\n# Load Reader library\r\nreader = Reader()\r\n\r\n# Load ratings dataset with Dataset library\r\ndata = Dataset.load_from_df(ratings[['user_id', 'movie_id', 'rating']], reader)\r\n\r\n# Use the SVD algorithm\r\nsvd = SVD()\r\n\r\n# Compute the RMSE of the SVD algorithm\r\ncross_validate(svd, data, cv=5, measures=['RMSE'], verbose=True)\r\n\r\n# Train on the dataset and arrive at predictions\r\ntrainset = data.build_full_trainset()\r\nsvd.fit(trainset)\r\n\r\n# Let's pick again user with ID 1310 and check the ratings he has given\r\nprint(ratings[ratings['user_id'] == 1310])\r\n\r\n# Now let's use SVD to predict the rating that 1310 will give to movie 1994\r\nprint(svd.predict(1310, 1994))", "Evaluating RMSE of algorithm SVD on 5 split(s).\n\n Fold 1 Fold 2 Fold 3 Fold 4 Fold 5 Mean Std \nRMSE (testset) 0.8726 0.8748 0.8735 0.8713 0.8757 0.8736 0.0016 \nFit time 53.92 52.58 52.30 52.06 52.18 52.61 0.68 \nTest time 3.01 2.84 2.85 2.47 2.79 2.79 0.18 \n user_id movie_id rating\n215928 1310 2988 3\n215929 1310 1293 5\n215930 1310 1295 2\n215931 1310 1299 4\n215932 1310 2243 4\n215933 1310 2248 5\n215934 1310 2620 5\n215935 1310 3683 5\n215936 1310 3685 4\n215937 1310 1185 4\n215938 1310 1196 4\n215939 1310 1343 4\n215940 1310 3101 4\n215941 1310 3111 4\n215942 1310 2313 2\n215943 1310 1704 5\n215944 1310 144 3\n215945 1310 3360 4\n215946 1310 3526 4\n215947 1310 1960 3\n215948 1310 2000 4\n215949 1310 1231 2\n215950 1310 1090 2\n215951 1310 1097 4\nuser: 1310 item: 1994 r_ui = None est = 3.60 {'was_impossible': False}\n" ] ] ]

[ "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code" ]

[ [ "code", "code", "code", "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code" ] ]

4ad9678ab21e5330a2893cf00591b88b45c4bf58

34,271

ipynb

Jupyter Notebook

Running_LTC_UT.ipynb

JefeLitman/Biv2LabNN

a89789fa10afb8a405e22c5d85a0836af275fbc8

[ "MIT" ]

null

Running_LTC_UT.ipynb

JefeLitman/Biv2LabNN

a89789fa10afb8a405e22c5d85a0836af275fbc8

[ "MIT" ]

null

Running_LTC_UT.ipynb

JefeLitman/Biv2LabNN

a89789fa10afb8a405e22c5d85a0836af275fbc8

[ "MIT" ]

null

34.722391

273

0.378746

[ [ [ "# Configuracion de grafica a usar", "_____no_output_____" ] ], [ [ "import os\nos.environ[\"CUDA_DEVICE_ORDER\"]=\"PCI_BUS_ID\";\n \n# lA ID de la GPU a usar, puede ser desde 0 hasta las N GPU's. Si es -1 significa que es en la CPU\nos.environ[\"CUDA_VISIBLE_DEVICES\"]=\"1\";", "_____no_output_____" ] ], [ [ "# Importacion de librerias", "_____no_output_____" ] ], [ [ "from __future__ import absolute_import, division, print_function, unicode_literals\nimport tensorflow as tf\nfrom tensorflow import keras\nfrom IPython.display import display, clear_output\nfrom ipywidgets import interact, IntSlider\nimport h5py\nimport numpy as np\n\n%matplotlib inline\nimport sys\nimport matplotlib.pyplot as plt\nfrom sklearn.preprocessing import MinMaxScaler\nfrom LTC import *\nsys.path.append('../')\nfrom Datasets_utils.DatasetsLoader import VideoDataGenerator", "_____no_output_____" ] ], [ [ "# Configuraciones para Tensorflow y Keras", "_____no_output_____" ] ], [ [ "print(\"Num GPUs Available: \", len(tf.config.experimental.list_physical_devices('GPU')))", "Num GPUs Available: 1\n" ], [ "gpus = tf.config.experimental.list_physical_devices('GPU')\ntf.config.experimental.set_memory_growth(gpus[0], True)", "_____no_output_____" ], [ "tf.debugging.set_log_device_placement(False)", "_____no_output_____" ], [ "#Comprobar que estoy ejecutandome en modo eagerly\ntf.executing_eagerly()", "_____no_output_____" ] ], [ [ "# Carga de Datos", "_____no_output_____" ] ], [ [ "root_path = \"/home/jefelitman/DataSets/ut/videos_set_1\"\nroot_path", "_____no_output_____" ], [ "batch_size = 12\noriginal_size = [171,128]\nsize = [112,112]\nframes = 16\ncanales = 3", "_____no_output_____" ], [ "def custom_steps_temporal(frames):\n mitad = len(frames)//2\n paso = len(frames)//16\n videos = []\n for i in range(paso):\n indices = range(i, len(frames),paso)\n videos.append([frames[j] for j in indices[:16]])\n return videos", "_____no_output_____" ], [ "def custom_temp_crop_unique(frames):\n mitad = len(frames)//2\n paso = len(frames)//16\n indices = sorted(list(range(mitad, -1,-paso)) + list(range(mitad+paso, len(frames),paso)))\n indices = indices[len(indices)//2 - 8 : len(indices)//2 + 8]\n return [[frames[i] for i in indices]]", "_____no_output_____" ], [ "def half_video_temporal(frames):\n mitad = len(frames)//2\n return [frames[mitad-8*2:mitad+8*2]]", "_____no_output_____" ], [ "def video_transf(video):\n escalador = MinMaxScaler()\n new_video = video.reshape((video.shape[0]*video.shape[1]*video.shape[2]*video.shape[3],1))\n new_video = escalador.fit_transform(new_video)\n return new_video.reshape((video.shape[0],video.shape[1],video.shape[2],video.shape[3]))", "_____no_output_____" ], [ "def flip_vertical(volume):\n return np.flip(volume, (0, 2))[::-1]", "_____no_output_____" ], [ "def corner_frame_crop(original_width, original_height):\n x = original_width-112\n y = original_height-112\n return [[x//2, original_width - x//2 -1, y//2, original_height-y//2],\n [x//2, original_width - x//2 -1, y//2, original_height-y//2]\n ]", "_____no_output_____" ], [ "dataset = VideoDataGenerator(directory_path = root_path, \n table_paths = None, \n batch_size = batch_size, \n original_frame_size = original_size, \n frame_size=size, \n video_frames = frames, \n temporal_crop = (\"custom\", custom_steps_temporal),\n video_transformation = [(\"augmented\",flip_vertical)],\n frame_crop = (\"custom\", corner_frame_crop), \n shuffle = True, \n conserve_original = False)", "_____no_output_____" ] ], [ [ "# Red Neuronal LTC", "_____no_output_____" ], [ "### Construccion del modelo", "_____no_output_____" ] ], [ [ "#Entrada de la red neuronal\nvideo_shape = tuple([frames]+size[::-1]+[canales])\ndropout = 0.5\nlr = 1e-3\nweigh_decay = 5e-3\n\nltc_save_path = '/home/jefelitman/Saved_Models/trained_ut/Encoder/Inception/inception_enhan/LTC-enhan-noMaxPoolT-SLTEnc_Seq_split1_{w}x{h}x{f}_softmax_sgd_'.format(\n w=size[0], h=size[1],f=frames)\nif canales == 3:\n ltc_save_path += 'RGB_'\nelse:\n ltc_save_path += 'B&N_'\n\nltc_save_path += 'lr={l}__dec-ori_2center-frame-crop_video-flip_temporal-dynamic_batchNorm_pretrained-c3d'.format(l = lr)\n\n#Creacion de la carpeta donde se salvara el modelo\nif not os.path.isdir(ltc_save_path):\n os.mkdir(ltc_save_path)\nmodel_saves_path = os.path.join(ltc_save_path,'model_saves')\nif not os.path.isdir(model_saves_path):\n os.mkdir(model_saves_path)\nltc_save_path", "_____no_output_____" ], [ "lr = 1e-3\nweigh_decay = 5e-3\n#Parametros para la compilacion del modelo\noptimizador = keras.optimizers.SGD(learning_rate=lr, momentum=0.9)\n#optimizador = keras.optimizers.Adam(learning_rate=lr)\nperdida = keras.losses.SparseCategoricalCrossentropy()\nprecision = keras.metrics.SparseCategoricalAccuracy()", "_____no_output_____" ], [ "entrada = keras.Input(shape=(300,224,224,3),batch_size=1,\n name=\"Input_video\")\n#Conv1\nx = keras.layers.Conv3D(filters=16, kernel_size=3, padding=\"same\", activation=\"relu\", \n kernel_regularizer=keras.regularizers.l2(weigh_decay),\n name='conv3d_1')(entrada)\nx = keras.layers.MaxPool3D(pool_size=(1,2,2),strides=(1,2,2), name='max_pooling3d_1')(x)\n\n#Conv2\nx = keras.layers.Conv3D(filters=32, kernel_size=3, padding=\"same\", activation=\"relu\", \n kernel_regularizer=keras.regularizers.l2(weigh_decay),\n name='conv3d_2')(x)\nx = keras.layers.MaxPool3D(pool_size=(1,2,2),strides=(1,2,2), name='max_pooling3d_2')(x)\n\n#Conv3\nx = keras.layers.Conv3D(filters=64, kernel_size=3, padding=\"same\", activation=\"relu\", \n kernel_regularizer=keras.regularizers.l2(weigh_decay),\n name='conv3d_3')(x)\nx = keras.layers.MaxPool3D(pool_size=(1,2,2),strides=(1,2,2),name='max_pooling3d_3')(x)\n\n#Conv4\nx = keras.layers.Conv3D(filters=128, kernel_size=3, padding=\"same\", activation=\"relu\", \n kernel_regularizer=keras.regularizers.l2(weigh_decay),\n name='conv3d_4')(x)\nx = keras.layers.MaxPool3D(pool_size=(1,2,2),strides=(1,2,2),name='max_pooling3d_4')(x)\n\nltc = keras.Model(entrada, x, name=\"LTC_original\")", "_____no_output_____" ], [ "ltc = get_LTC_encoder_slt_I(video_shape, len(dataset.to_class),dropout, weigh_decay, 256, 512, True)", "_____no_output_____" ], [ "#Compilacion del modelo\nltc.compile(optimizer = optimizador,\n loss = perdida,\n metrics = [precision])", "_____no_output_____" ], [ "#keras.utils.plot_model(ltc, 'LTC.png', show_shapes=True)", "_____no_output_____" ], [ "#ltc = keras.models.load_model('/home/jefelitman/Saved_Models/trained_ut/Inception/conv_channels/LTC-incept-channels_split1_112x112x16_softmax_sgd_RGB_lr=0.001_decreased-original_2center-frame-crop_video-flip_temporal-dynamic_batchNorm_pretrained-c3d/ltc_final_1.h5')", "_____no_output_____" ], [ "ltc.summary()", "Model: \"LTC_original\"\n_________________________________________________________________\nLayer (type) Output Shape Param # \n=================================================================\nInput_video (InputLayer) [(1, 300, 224, 224, 3)] 0 \n_________________________________________________________________\nconv3d_1 (Conv3D) (1, 300, 224, 224, 16) 1312 \n_________________________________________________________________\nmax_pooling3d_1 (MaxPooling3 (1, 300, 112, 112, 16) 0 \n_________________________________________________________________\nconv3d_2 (Conv3D) (1, 300, 112, 112, 32) 13856 \n_________________________________________________________________\nmax_pooling3d_2 (MaxPooling3 (1, 300, 56, 56, 32) 0 \n_________________________________________________________________\nconv3d_3 (Conv3D) (1, 300, 56, 56, 64) 55360 \n_________________________________________________________________\nmax_pooling3d_3 (MaxPooling3 (1, 300, 28, 28, 64) 0 \n_________________________________________________________________\nconv3d_4 (Conv3D) (1, 300, 28, 28, 128) 221312 \n_________________________________________________________________\nmax_pooling3d_4 (MaxPooling3 (1, 300, 14, 14, 128) 0 \n=================================================================\nTotal params: 291,840\nTrainable params: 291,840\nNon-trainable params: 0\n_________________________________________________________________\n" ], [ "ltc.predict(np.zeros((1,300,224,224,3)))", "_____no_output_____" ] ], [ [ "### Cargo los pesos pre entrenados", "_____no_output_____" ], [ "##### Pesos del C3D", "_____no_output_____" ] ], [ [ "c3d_weights = h5py.File('/home/jefelitman/Saved_Models/c3d-sports1M_weights.h5', 'r')\nprint(c3d_weights.keys())", "_____no_output_____" ], [ "c3d_weights['layer_0'].keys()", "_____no_output_____" ], [ "weights = []\nfor capa in ['layer_0','layer_2','layer_4','layer_5','layer_5']:\n weights.append([\n np.moveaxis(np.r_[c3d_weights[capa]['param_0']], (0,1),(4,3)), #Cambio los ejes porque c3d estan con canales primero\n np.r_[c3d_weights[capa]['param_1']]\n ])\nfor index, capa in enumerate(['conv3d_1','conv3d_2','conv3d_3','conv3d_4','conv3d_5']):\n ltc.get_layer(capa).set_weights(weights[index])", "_____no_output_____" ] ], [ [ "##### Pesos de la InceptionV3", "_____no_output_____" ] ], [ [ "inceptionv3 = keras.applications.InceptionV3(weights=\"imagenet\")", "_____no_output_____" ], [ "for layer, index in [('conv3d_1',1),('conv3d_2',4),('conv3d_3',7),('conv3d_4',11),('conv3d_5',14)]:\n old_weights, old_bias = ltc.get_layer(layer).get_weights()\n new_weight = np.zeros(old_weights.shape)\n new_bias = np.zeros(old_bias.shape)\n pesos = inceptionv3.layers[index].get_weights()[0]\n for entrada in range(old_weights.shape[3]):\n for salida in range(old_weights.shape[4]):\n new_weight[:,:,:,entrada,salida] = np.stack([pesos[:,:,entrada%pesos.shape[2],salida%pesos.shape[3]], \n pesos[:,:,entrada%pesos.shape[2],salida%pesos.shape[3]], \n pesos[:,:,entrada%pesos.shape[2],salida%pesos.shape[3]]\n ]\n )/3\n ltc.get_layer(layer).set_weights([new_weight, new_bias])", "_____no_output_____" ] ], [ [ "### Entrenamiento de la red con el generador", "_____no_output_____" ] ], [ [ "#Funcion customizadas para el entrenamiento del modelo \nclass custom_callback(keras.callbacks.Callback):\n def __init__(self):\n self.accuracies = []\n self.losses = []\n self.val_accuracies = []\n self.val_loss = []\n \n def on_batch_end(self, batch, logs):\n corte = dataset.train_batches//3 + 1\n if batch == corte or batch == corte*2:\n keras.backend.set_value(optimizador.lr, optimizador.lr.numpy()*0.1)\n for i in ['conv3d_1','conv3d_2','conv3d_3','conv3d_4','conv3d_5','dense_8','dense_9','dense_10']:\n weigh_decay = ltc.get_layer(i).kernel_regularizer.get_config()['l2'] * 0.1\n ltc.get_layer(i).kernel_regularizer = keras.regularizers.l2(weigh_decay)\n print(\"\\n\",\"Actual LR: \", str(optimizador.lr.numpy()))\n \n self.accuracies.append(logs['sparse_categorical_accuracy'])\n self.losses.append(logs['loss'])\n \n def on_epoch_begin(self, epoch, logs):\n keras.backend.set_value(optimizador.lr, 0.001)\n \n def on_epoch_end(self,batch, logs):\n self.val_accuracies.append(logs['val_sparse_categorical_accuracy'])\n self.val_loss.append(logs['val_loss'])\n\nfunciones = [\n keras.callbacks.ModelCheckpoint(\n filepath=os.path.join(model_saves_path,'ltc_epoch_{epoch}.h5'),\n save_best_only=True,\n monitor='val_sparse_categorical_accuracy',\n verbose=1),\n keras.callbacks.CSVLogger(os.path.join(ltc_save_path,'output.csv')),\n custom_callback()\n]", "_____no_output_____" ], [ "epoch = 1\nhistorial = ltc.fit(x = dataset.get_train_generator(canales),\n steps_per_epoch=dataset.train_batches,\n epochs=epoch,\n callbacks=funciones,\n validation_data= dataset.get_test_generator(canales),\n validation_steps=dataset.test_batches,\n max_queue_size=batch_size%24)", "_____no_output_____" ] ], [ [ "### Guardado del modelo ", "_____no_output_____" ] ], [ [ "#Salvado final definitivo del modelo una vez se detenga\nltc.save(os.path.join(ltc_save_path,\"ltc_final_{e}.h5\".format(e=epoch)))", "_____no_output_____" ] ], [ [ "### Graficas de los resultados de entrenamiento", "_____no_output_____" ] ], [ [ "fig = plt.figure()\nplt.plot(funciones[-1].losses,'k--')\nrango = [i*dataset.train_batches-1 for i in range(1,epoch+1)]\nplt.plot(rango, funciones[-1].val_loss,'bo')\nplt.title('Loss over steps')\nplt.legend(labels=[\"Loss\",\"Test Loss\"])\nplt.show()\nfig.savefig(os.path.join(ltc_save_path,'train_loss_steps_{e}.png'.format(e=dataset.train_batches)))", "_____no_output_____" ], [ "fig = plt.figure()\nplt.plot(funciones[-1].accuracies,'k--')\nrango = [i*dataset.train_batches-1 for i in range(1,epoch+1)]\nplt.plot(rango, funciones[-1].val_accuracies,'bo')\nplt.title('Accuracy over steps')\nplt.legend(labels=[\"Accuracy\",\"Test Accuracy\"])\nplt.show()\nfig.savefig(os.path.join(ltc_save_path,'train_accuracy_steps_{e}.png'.format(e=dataset.train_batches)))", "_____no_output_____" ] ], [ [ "### Evaluacion del entrenamiento", "_____no_output_____" ] ], [ [ "resultados = ltc.evaluate_generator(generator=dataset.get_test_generator(canales),\n steps=dataset.test_batches,\n max_queue_size=batch_size)\nprint(\"\"\"Los resultados de la evaluacion del modelo fueron: \nPerdida: {l}\nPrecision: {a}\"\"\".format(l=resultados[0],a=resultados[1]))", "_____no_output_____" ] ] ]

[ "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "raw", "markdown", "code", "markdown", "code", "markdown", "code", "markdown", "raw" ]

[ [ "markdown" ], [ "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code", "code", "code" ], [ "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code", "code", "code", "code", "code", "code", "code" ], [ "markdown", "markdown" ], [ "code", "code", "code" ], [ "markdown" ], [ "raw", "raw" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "code" ], [ "markdown" ], [ "code", "code" ], [ "markdown" ], [ "raw" ] ]

4ad975775c3298e96f9ece920cb34600ff5ed782

22,802

ipynb

Jupyter Notebook

Data Processing Code/EMBEDDING-TESTING.ipynb

maxby12/CiTius-Summer-Research

e3e3c39ad8d2045410e01d4f258e138616ef4c36

[ "MIT" ]

null

Data Processing Code/EMBEDDING-TESTING.ipynb

maxby12/CiTius-Summer-Research

e3e3c39ad8d2045410e01d4f258e138616ef4c36

[ "MIT" ]

null

Data Processing Code/EMBEDDING-TESTING.ipynb

maxby12/CiTius-Summer-Research

e3e3c39ad8d2045410e01d4f258e138616ef4c36

[ "MIT" ]

null

86.371212

9,712

0.821068

[ [ [ "import matplotlib.pyplot as plt\nimport numpy as np\nimport random", "_____no_output_____" ], [ "x,y = np.random.uniform(low=10, high=11, size=100), np.random.uniform(low=10, high=11, size=100)\ncoords = list(zip(x,y))\ncoords = np.array(coords)\nprint(coords.shape)\nprint(coords)", "(100, 2)\n[[10.81194303 10.2782721 ]\n [10.2272346 10.57933775]\n [10.71443143 10.20152613]\n [10.35033842 10.25212478]\n [10.20975874 10.84128657]\n [10.95105842 10.63475055]\n [10.30622227 10.15519208]\n [10.80051988 10.93149124]\n [10.94350749 10.71635695]\n [10.81590407 10.08810949]\n [10.88962759 10.79324012]\n [10.24611323 10.75086399]\n [10.21368677 10.1634206 ]\n [10.88761142 10.24808044]\n [10.80346655 10.15263401]\n [10.31645727 10.72964023]\n [10.51615076 10.53024617]\n [10.12011306 10.06478788]\n [10.85156434 10.64428437]\n [10.39947138 10.81177609]\n [10.7648015 10.02754602]\n [10.91322268 10.47129018]\n [10.97706048 10.25230228]\n [10.35081106 10.94006996]\n [10.51599502 10.23952285]\n [10.7438848 10.24468909]\n [10.36357917 10.47797729]\n [10.86858166 10.18419673]\n [10.42856976 10.20728353]\n [10.198331 10.53309349]\n [10.52332103 10.29800457]\n [10.86975521 10.43734795]\n [10.6810594 10.24163851]\n [10.9765458 10.70065718]\n [10.26531446 10.13369747]\n [10.21644017 10.80569074]\n [10.22534119 10.84107786]\n [10.14488574 10.87515812]\n [10.40291627 10.53483551]\n [10.76195758 10.00399211]\n [10.46587122 10.96778556]\n [10.38110297 10.83157211]\n [10.45353514 10.50724252]\n [10.77328031 10.81042912]\n [10.77809425 10.80988464]\n [10.88407448 10.0244113 ]\n [10.60305262 10.83010832]\n [10.81127363 10.87978149]\n [10.95161871 10.87062787]\n [10.58364758 10.02267114]\n [10.43202563 10.68557112]\n [10.99967017 10.64916169]\n [10.86189758 10.69639184]\n [10.96316755 10.39999029]\n [10.54148964 10.59318645]\n [10.19233848 10.02659851]\n [10.19757071 10.60410697]\n [10.55005496 10.15802054]\n [10.91552603 10.05353591]\n [10.32614685 10.33166193]\n [10.79179889 10.50802149]\n [10.36525091 10.12524965]\n [10.62187812 10.73051786]\n [10.44796969 10.79414199]\n [10.64495529 10.95398743]\n [10.73132177 10.24343747]\n [10.56519258 10.36043825]\n [10.14204155 10.38484683]\n [10.21091127 10.92510419]\n [10.12675124 10.73580792]\n [10.92251477 10.3891961 ]\n [10.39259413 10.52650001]\n [10.8414177 10.32836883]\n [10.06575541 10.97288761]\n [10.62794318 10.79258799]\n [10.69964774 10.78697735]\n [10.46701278 10.15515811]\n [10.77257862 10.93701651]\n [10.93985108 10.13291516]\n [10.47027624 10.26086041]\n [10.7644671 10.77870809]\n [10.51311156 10.26412324]\n [10.69589002 10.31620083]\n [10.69241055 10.6842504 ]\n [10.56932864 10.96014839]\n [10.34585254 10.65715219]\n [10.24764912 10.09193595]\n [10.80958004 10.89018085]\n [10.35204505 10.39280494]\n [10.97677673 10.54618398]\n [10.50943626 10.35898214]\n [10.65771984 10.66151774]\n [10.32198851 10.08506213]\n [10.15192092 10.06643246]\n [10.4263815 10.01062579]\n [10.39588118 10.00833168]\n [10.91098666 10.56622029]\n [10.16541752 10.62766737]\n [10.07981385 10.23805631]\n [10.7426537 10.20861656]]\n" ], [ "plt.scatter(coords[:,0], coords[:,1])\nplt.show()", "_____no_output_____" ], [ "mean, std = np.mean(coords, axis=0), np.std(coords, axis=0)\n\nprint(mean.shape, std.shape)\nplt.scatter(coords[:,0], coords[:,1])\nplt.scatter(mean[0], mean[1], c='g')\nplt.scatter(std[0], std[1], c='r')\nplt.show()", "(2,) (2,)\n" ], [ "candidates = np.apply_along_axis(lambda x: np.linalg.norm(x, ord=2),axis=1,arr=coords-mean)\nprint(candidates)\nmean_distance, std_distance = np.mean(candidates), np.std(candidates)\nprint(mean_distance, std_distance)\nprint(np.linalg.norm(std, ord=2))", "[0.32100363 0.35057003 0.31846399 0.31410634 0.50442753 0.41591586\n 0.41651374 0.50763191 0.4458081 0.46773021 0.45018664 0.41623992\n 0.47376898 0.39938294 0.40738112 0.35002363 0.06782606 0.60944174\n 0.32997442 0.36833916 0.49705056 0.34929576 0.47267734 0.50504324\n 0.24747211 0.29794218 0.20056655 0.42609818 0.30657531 0.36928212\n 0.18872817 0.30893965 0.26755426 0.46669346 0.45910629 0.47483172\n 0.49345428 0.5745414 0.16953685 0.51759235 0.49534813 0.39433055\n 0.11334865 0.38915544 0.39130813 0.55858986 0.3500081 0.46808838\n 0.54862492 0.46001893 0.24243153 0.46644718 0.36678305 0.40746272\n 0.11319304 0.58808761 0.38624684 0.32455803 0.55436207 0.28161269\n 0.22915025 0.40866564 0.25487859 0.33278743 0.47859238 0.29155853\n 0.1218412 0.43315727 0.56642753 0.50552233 0.3703039 0.17711591\n 0.31716282 0.6993198 0.31681291 0.33349225 0.34121991 0.50025366\n 0.51307027 0.24047257 0.35779876 0.22403076 0.2120122 0.23928628\n 0.47790245 0.27966766 0.50249893 0.4760758 0.23015641 0.41759642\n 0.134867 0.20221983 0.46518332 0.58550489 0.49134392 0.50291083\n 0.3568998 0.42436604 0.5423793 0.32675715]\n0.383089864635655 0.12622918692655552\n0.40335053243885\n" ] ] ]

[ "code" ]

[ [ "code", "code", "code", "code", "code" ] ]