Spaces:
Sleeping
Sleeping
Commit
·
9d912f9
1
Parent(s):
e9dfeb1
First commit
Browse files- __pycache__/resnet.cpython-312.pyc +0 -0
- __pycache__/utils.cpython-312.pyc +0 -0
- app.py +83 -0
- cat.jpg +0 -0
- dog.jpg +0 -0
- model.ckpt +3 -0
- requirements.txt +116 -0
- resnet.py +122 -0
- utils.py +237 -0
__pycache__/resnet.cpython-312.pyc
ADDED
Binary file (7.95 kB). View file
|
|
__pycache__/utils.cpython-312.pyc
ADDED
Binary file (9.6 kB). View file
|
|
app.py
ADDED
@@ -0,0 +1,83 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch, torchvision
|
2 |
+
from torchvision import transforms
|
3 |
+
import numpy as np
|
4 |
+
import gradio as gr
|
5 |
+
from PIL import Image
|
6 |
+
from pytorch_grad_cam import GradCAM
|
7 |
+
from pytorch_grad_cam.utils.image import show_cam_on_image
|
8 |
+
from resnet import ResNet18
|
9 |
+
import gradio as gr
|
10 |
+
|
11 |
+
model = ResNet18()
|
12 |
+
model.load_state_dict(torch.load("model.ckpt", map_location=torch.device('cpu')), strict=False)
|
13 |
+
|
14 |
+
inv_normalize = transforms.Normalize(
|
15 |
+
mean=[-0.50/0.23, -0.50/0.23, -0.50/0.23],
|
16 |
+
std=[1/0.23, 1/0.23, 1/0.23]
|
17 |
+
)
|
18 |
+
classes = ('plane', 'car', 'bird', 'cat', 'deer',
|
19 |
+
'dog', 'frog', 'horse', 'ship', 'truck')
|
20 |
+
|
21 |
+
def resize_image_pil(image, new_width, new_height):
|
22 |
+
|
23 |
+
# Convert to PIL image
|
24 |
+
img = Image.fromarray(np.array(image))
|
25 |
+
|
26 |
+
# Get original size
|
27 |
+
width, height = img.size
|
28 |
+
|
29 |
+
# Calculate scale
|
30 |
+
width_scale = new_width / width
|
31 |
+
height_scale = new_height / height
|
32 |
+
scale = min(width_scale, height_scale)
|
33 |
+
|
34 |
+
# Resize
|
35 |
+
resized = img.resize((int(width*scale), int(height*scale)), Image.NEAREST)
|
36 |
+
|
37 |
+
# Crop to exact size
|
38 |
+
resized = resized.crop((0, 0, new_width, new_height))
|
39 |
+
|
40 |
+
return resized
|
41 |
+
|
42 |
+
def inference(input_img, transparency = 0.5, target_layer_number = -1):
|
43 |
+
input_img = resize_image_pil(input_img, 32, 32)
|
44 |
+
|
45 |
+
input_img = np.array(input_img)
|
46 |
+
org_img = input_img
|
47 |
+
input_img = input_img.reshape((32, 32, 3))
|
48 |
+
transform = transforms.ToTensor()
|
49 |
+
input_img = transform(input_img)
|
50 |
+
input_img = input_img
|
51 |
+
input_img = input_img.unsqueeze(0)
|
52 |
+
outputs = model(input_img)
|
53 |
+
softmax = torch.nn.Softmax(dim=0)
|
54 |
+
o = softmax(outputs.flatten())
|
55 |
+
confidences = {classes[i]: float(o[i]) for i in range(10)}
|
56 |
+
_, prediction = torch.max(outputs, 1)
|
57 |
+
target_layers = [model.layer2[target_layer_number]]
|
58 |
+
cam = GradCAM(model=model, target_layers=target_layers)
|
59 |
+
grayscale_cam = cam(input_tensor=input_img, targets=None)
|
60 |
+
grayscale_cam = grayscale_cam[0, :]
|
61 |
+
visualization = show_cam_on_image(org_img/255, grayscale_cam, use_rgb=True, image_weight=transparency)
|
62 |
+
return classes[prediction[0].item()], visualization, confidences
|
63 |
+
|
64 |
+
title = "CIFAR10 trained on ResNet18 Model with GradCAM"
|
65 |
+
description = "A simple Gradio interface to infer on ResNet model, and get GradCAM results"
|
66 |
+
examples = [["cat.jpg", 0.5, -1], ["dog.jpg", 0.5, -1]]
|
67 |
+
demo = gr.Interface(
|
68 |
+
inference,
|
69 |
+
inputs = [
|
70 |
+
gr.Image(width=256, height=256, label="Input Image"), gr.Slider
|
71 |
+
(0, 1, value = 0.5, label="Overall Opacity of Image"),
|
72 |
+
gr.Slider(-2, -1, value = -2, step=1, label="Which Layer?")
|
73 |
+
],
|
74 |
+
outputs = [
|
75 |
+
"text",
|
76 |
+
gr.Image(width=256, height=256, label="Output"),
|
77 |
+
gr.Label(num_top_classes=3)
|
78 |
+
],
|
79 |
+
title = title,
|
80 |
+
description = description,
|
81 |
+
examples = examples,
|
82 |
+
)
|
83 |
+
demo.launch()
|
cat.jpg
ADDED
![]() |
dog.jpg
ADDED
![]() |
model.ckpt
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:940f21f828787740b7b275a45b29051806977b52570b4e2afbb50a3f1dd04cab
|
3 |
+
size 89492032
|
requirements.txt
ADDED
@@ -0,0 +1,116 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
aiofiles==23.2.1
|
2 |
+
aiohttp==3.9.5
|
3 |
+
aiosignal==1.3.1
|
4 |
+
altair==5.3.0
|
5 |
+
annotated-types==0.6.0
|
6 |
+
anyio==4.3.0
|
7 |
+
asttokens==2.4.1
|
8 |
+
attrs==23.2.0
|
9 |
+
certifi==2024.2.2
|
10 |
+
charset-normalizer==3.3.2
|
11 |
+
click==8.1.7
|
12 |
+
colorama==0.4.6
|
13 |
+
comm==0.2.2
|
14 |
+
contourpy==1.2.1
|
15 |
+
cycler==0.12.1
|
16 |
+
debugpy==1.8.1
|
17 |
+
decorator==5.1.1
|
18 |
+
executing==2.0.1
|
19 |
+
fastapi==0.110.2
|
20 |
+
ffmpy==0.3.2
|
21 |
+
filelock==3.13.1
|
22 |
+
fonttools==4.51.0
|
23 |
+
frozenlist==1.4.1
|
24 |
+
fsspec==2024.2.0
|
25 |
+
grad-cam==1.5.0
|
26 |
+
gradio==4.28.3
|
27 |
+
gradio_client==0.16.0
|
28 |
+
h11==0.14.0
|
29 |
+
httpcore==1.0.5
|
30 |
+
httpx==0.27.0
|
31 |
+
huggingface-hub==0.22.2
|
32 |
+
idna==3.7
|
33 |
+
importlib_resources==6.4.0
|
34 |
+
intel-openmp==2021.4.0
|
35 |
+
ipykernel==6.29.4
|
36 |
+
ipython==8.24.0
|
37 |
+
jedi==0.19.1
|
38 |
+
Jinja2==3.1.3
|
39 |
+
joblib==1.4.0
|
40 |
+
jsonschema==4.21.1
|
41 |
+
jsonschema-specifications==2023.12.1
|
42 |
+
jupyter_client==8.6.1
|
43 |
+
jupyter_core==5.7.2
|
44 |
+
kiwisolver==1.4.5
|
45 |
+
lightning==2.2.3
|
46 |
+
lightning-utilities==0.11.2
|
47 |
+
markdown-it-py==3.0.0
|
48 |
+
MarkupSafe==2.1.5
|
49 |
+
matplotlib==3.8.4
|
50 |
+
matplotlib-inline==0.1.7
|
51 |
+
mdurl==0.1.2
|
52 |
+
mkl==2021.4.0
|
53 |
+
mpmath==1.3.0
|
54 |
+
multidict==6.0.5
|
55 |
+
nest-asyncio==1.6.0
|
56 |
+
networkx==3.2.1
|
57 |
+
numpy==1.26.3
|
58 |
+
opencv-python==4.9.0.80
|
59 |
+
orjson==3.10.1
|
60 |
+
packaging==24.0
|
61 |
+
pandas==2.2.2
|
62 |
+
parso==0.8.4
|
63 |
+
pillow==10.2.0
|
64 |
+
platformdirs==4.2.1
|
65 |
+
prompt-toolkit==3.0.43
|
66 |
+
psutil==5.9.8
|
67 |
+
pure-eval==0.2.2
|
68 |
+
pydantic==2.7.1
|
69 |
+
pydantic_core==2.18.2
|
70 |
+
pydub==0.25.1
|
71 |
+
Pygments==2.17.2
|
72 |
+
pyparsing==3.1.2
|
73 |
+
python-dateutil==2.9.0.post0
|
74 |
+
python-multipart==0.0.9
|
75 |
+
pytorch-lightning==2.2.3
|
76 |
+
pytz==2024.1
|
77 |
+
pywin32==306
|
78 |
+
PyYAML==6.0.1
|
79 |
+
pyzmq==26.0.2
|
80 |
+
referencing==0.35.0
|
81 |
+
requests==2.31.0
|
82 |
+
rich==13.7.1
|
83 |
+
rpds-py==0.18.0
|
84 |
+
ruff==0.4.2
|
85 |
+
scikit-learn==1.4.2
|
86 |
+
scipy==1.13.0
|
87 |
+
semantic-version==2.10.0
|
88 |
+
setuptools==69.5.1
|
89 |
+
shellingham==1.5.4
|
90 |
+
six==1.16.0
|
91 |
+
sniffio==1.3.1
|
92 |
+
stack-data==0.6.3
|
93 |
+
starlette==0.37.2
|
94 |
+
sympy==1.12
|
95 |
+
tbb==2021.11.0
|
96 |
+
threadpoolctl==3.4.0
|
97 |
+
tomlkit==0.12.0
|
98 |
+
toolz==0.12.1
|
99 |
+
torch==2.3.0+cu121
|
100 |
+
torch-lr-finder==0.2.1
|
101 |
+
torchaudio==2.3.0+cu121
|
102 |
+
torchmetrics==1.3.2
|
103 |
+
torchsummary==1.5.1
|
104 |
+
torchvision==0.18.0+cu121
|
105 |
+
tornado==6.4
|
106 |
+
tqdm==4.66.2
|
107 |
+
traitlets==5.14.3
|
108 |
+
ttach==0.0.3
|
109 |
+
typer==0.12.3
|
110 |
+
typing_extensions==4.9.0
|
111 |
+
tzdata==2024.1
|
112 |
+
urllib3==2.2.1
|
113 |
+
uvicorn==0.29.0
|
114 |
+
wcwidth==0.2.13
|
115 |
+
websockets==11.0.3
|
116 |
+
yarl==1.9.4
|
resnet.py
ADDED
@@ -0,0 +1,122 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
ResNet in PyTorch.
|
3 |
+
For Pre-activation ResNet, see 'preact_resnet.py'.
|
4 |
+
|
5 |
+
Reference:
|
6 |
+
[1] Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun
|
7 |
+
Deep Residual Learning for Image Recognition. arXiv:1512.03385
|
8 |
+
"""
|
9 |
+
import os
|
10 |
+
import torch
|
11 |
+
import utils
|
12 |
+
import torch.nn as nn
|
13 |
+
import torch.nn.functional as F
|
14 |
+
|
15 |
+
from torchmetrics import Accuracy
|
16 |
+
from torchvision.datasets import CIFAR10
|
17 |
+
from pytorch_lightning import LightningModule
|
18 |
+
from torch.utils.data import DataLoader, random_split
|
19 |
+
|
20 |
+
|
21 |
+
class BasicBlock(nn.Module):
|
22 |
+
expansion = 1
|
23 |
+
|
24 |
+
def __init__(self, in_planes, planes, stride=1):
|
25 |
+
super(BasicBlock, self).__init__()
|
26 |
+
self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
|
27 |
+
self.bn1 = nn.BatchNorm2d(planes)
|
28 |
+
self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False)
|
29 |
+
self.bn2 = nn.BatchNorm2d(planes)
|
30 |
+
|
31 |
+
self.shortcut = nn.Sequential()
|
32 |
+
if stride != 1 or in_planes != self.expansion*planes:
|
33 |
+
self.shortcut = nn.Sequential(
|
34 |
+
nn.Conv2d(in_planes, self.expansion*planes, kernel_size=1, stride=stride, bias=False),
|
35 |
+
nn.BatchNorm2d(self.expansion*planes)
|
36 |
+
)
|
37 |
+
|
38 |
+
def forward(self, x):
|
39 |
+
out = F.relu(self.bn1(self.conv1(x)))
|
40 |
+
out = self.bn2(self.conv2(out))
|
41 |
+
out += self.shortcut(x)
|
42 |
+
out = F.relu(out)
|
43 |
+
return out
|
44 |
+
|
45 |
+
|
46 |
+
class ResNet(LightningModule):
|
47 |
+
def __init__(self, block, num_blocks, num_classes=10, loss='cross_entropy', learning_rate=2e-4, momentum=0.9, optimizer="SGD",
|
48 |
+
epochs=20):
|
49 |
+
super(ResNet, self).__init__()
|
50 |
+
self.in_planes = 64
|
51 |
+
|
52 |
+
self.conv1 = nn.Conv2d(3, 64, kernel_size=3, stride=1, padding=1, bias=False)
|
53 |
+
self.bn1 = nn.BatchNorm2d(64)
|
54 |
+
self.layer1 = self._make_layer(block, 64, num_blocks[0], stride=1)
|
55 |
+
self.layer2 = self._make_layer(block, 128, num_blocks[1], stride=2)
|
56 |
+
self.layer3 = self._make_layer(block, 256, num_blocks[2], stride=2)
|
57 |
+
self.layer4 = self._make_layer(block, 512, num_blocks[3], stride=2)
|
58 |
+
self.linear = nn.Linear(512*block.expansion, num_classes)
|
59 |
+
self.accuracy = Accuracy(task="multiclass", num_classes=num_classes)
|
60 |
+
self.learning_rate = learning_rate
|
61 |
+
self.optimizer = optimizer
|
62 |
+
self.momentum = momentum
|
63 |
+
self.loss = utils.get_criterion(loss)
|
64 |
+
self.epochs = epochs
|
65 |
+
|
66 |
+
def _make_layer(self, block, planes, num_blocks, stride):
|
67 |
+
strides = [stride] + [1]*(num_blocks-1)
|
68 |
+
layers = []
|
69 |
+
for stride in strides:
|
70 |
+
layers.append(block(self.in_planes, planes, stride))
|
71 |
+
self.in_planes = planes * block.expansion
|
72 |
+
return nn.Sequential(*layers)
|
73 |
+
|
74 |
+
def forward(self, x):
|
75 |
+
out = F.relu(self.bn1(self.conv1(x)))
|
76 |
+
out = self.layer1(out)
|
77 |
+
out = self.layer2(out)
|
78 |
+
out = self.layer3(out)
|
79 |
+
out = self.layer4(out)
|
80 |
+
out = F.avg_pool2d(out, 4)
|
81 |
+
out = out.view(out.size(0), -1)
|
82 |
+
out = self.linear(out)
|
83 |
+
return out
|
84 |
+
|
85 |
+
def training_step(self, batch, batch_idx):
|
86 |
+
x, y = batch
|
87 |
+
loss = self.loss(self(x), y)
|
88 |
+
return loss
|
89 |
+
|
90 |
+
def validation_step(self, batch, batch_idx):
|
91 |
+
x, y = batch
|
92 |
+
logits = self(x)
|
93 |
+
loss = self.loss(logits, y)
|
94 |
+
preds = torch.argmax(logits, dim=1)
|
95 |
+
self.accuracy(preds, y)
|
96 |
+
|
97 |
+
# Calling self.log will surface up scalars for you in TensorBoard
|
98 |
+
self.log("val_loss", loss, prog_bar=True)
|
99 |
+
self.log("val_acc", self.accuracy, prog_bar=True)
|
100 |
+
return loss
|
101 |
+
|
102 |
+
def test_step(self, batch, batch_idx):
|
103 |
+
# Here we just reuse the validation_step for testing
|
104 |
+
return self.validation_step(batch, batch_idx)
|
105 |
+
|
106 |
+
def configure_optimizers(self):
|
107 |
+
optimizer = utils.get_optimizer(self, lr=self.learning_rate, momentum=self.momentum, optimizer_type="SGD")
|
108 |
+
max_lr = utils.get_learning_rate(self, optimizer, self.loss, self.trainer.datamodule.train_dataloader())
|
109 |
+
scheduler = utils.get_OneCycleLR_scheduler(optimizer, max_lr=max_lr, epochs=self.epochs,
|
110 |
+
steps_per_epoch=len(self.trainer.datamodule.train_dataloader()), max_at_epoch=5,
|
111 |
+
anneal_strategy = 'linear', div_factor=10,
|
112 |
+
final_div_factor=1)
|
113 |
+
return [optimizer],[{"scheduler": scheduler, "interval": "step", "frequency": 1}]
|
114 |
+
|
115 |
+
def ResNet18(loss='cross_entropy', learning_rate=2e-4, momentum=0.9, optimizer="SGD", epochs=20):
|
116 |
+
return ResNet(BasicBlock, [2, 2, 2, 2], loss=loss, learning_rate=learning_rate, momentum=momentum,
|
117 |
+
optimizer=optimizer, epochs=epochs)
|
118 |
+
|
119 |
+
|
120 |
+
def ResNet34(loss='cross_entropy', learning_rate=2e-4, momentum=0.9, optimizer="SGD", epochs=20):
|
121 |
+
return ResNet(BasicBlock, [3, 4, 6, 3], loss=loss, learning_rate=learning_rate, momentum=momentum,
|
122 |
+
optimizer=optimizer, epochs=epochs)
|
utils.py
ADDED
@@ -0,0 +1,237 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
#!/usr/bin/env python3
|
2 |
+
"""
|
3 |
+
Utility Script containing functions to be used for training
|
4 |
+
Author: Shilpaj Bhalerao
|
5 |
+
"""
|
6 |
+
# Standard Library Imports
|
7 |
+
import math
|
8 |
+
from typing import NoReturn
|
9 |
+
|
10 |
+
# Third-Party Imports
|
11 |
+
import numpy as np
|
12 |
+
import matplotlib.pyplot as plt
|
13 |
+
import torch
|
14 |
+
from torchsummary import summary
|
15 |
+
from torchvision import transforms
|
16 |
+
from pytorch_grad_cam import GradCAM
|
17 |
+
from pytorch_grad_cam.utils.image import show_cam_on_image
|
18 |
+
|
19 |
+
import torch.optim as optim
|
20 |
+
import torch.nn.functional as F
|
21 |
+
from torch_lr_finder import LRFinder
|
22 |
+
|
23 |
+
|
24 |
+
def get_summary(model, input_size: tuple) -> NoReturn:
|
25 |
+
"""
|
26 |
+
Function to get the summary of the model architecture
|
27 |
+
:param model: Object of model architecture class
|
28 |
+
:param input_size: Input data shape (Channels, Height, Width)
|
29 |
+
"""
|
30 |
+
use_cuda = torch.cuda.is_available()
|
31 |
+
device = torch.device("cuda" if use_cuda else "cpu")
|
32 |
+
network = model.to(device)
|
33 |
+
summary(network, input_size=input_size)
|
34 |
+
|
35 |
+
|
36 |
+
def get_misclassified_data(model, device, test_loader):
|
37 |
+
"""
|
38 |
+
Function to run the model on test set and return misclassified images
|
39 |
+
:param model: Network Architecture
|
40 |
+
:param device: CPU/GPU
|
41 |
+
:param test_loader: DataLoader for test set
|
42 |
+
"""
|
43 |
+
# Prepare the model for evaluation i.e. drop the dropout layer
|
44 |
+
model.eval()
|
45 |
+
model.to(device)
|
46 |
+
|
47 |
+
# List to store misclassified Images
|
48 |
+
misclassified_data = []
|
49 |
+
|
50 |
+
# Reset the gradients
|
51 |
+
with torch.no_grad():
|
52 |
+
# Extract images, labels in a batch
|
53 |
+
for data, target in test_loader:
|
54 |
+
|
55 |
+
# Migrate the data to the device
|
56 |
+
data, target = data.to(device), target.to(device)
|
57 |
+
|
58 |
+
# Extract single image, label from the batch
|
59 |
+
for image, label in zip(data, target):
|
60 |
+
|
61 |
+
# Add batch dimension to the image
|
62 |
+
image = image.unsqueeze(0)
|
63 |
+
|
64 |
+
# Get the model prediction on the image
|
65 |
+
output = model(image)
|
66 |
+
|
67 |
+
# Convert the output from one-hot encoding to a value
|
68 |
+
pred = output.argmax(dim=1, keepdim=True)
|
69 |
+
|
70 |
+
# If prediction is incorrect, append the data
|
71 |
+
if pred != label:
|
72 |
+
misclassified_data.append((image, label, pred))
|
73 |
+
return misclassified_data
|
74 |
+
|
75 |
+
|
76 |
+
# -------------------- GradCam --------------------
|
77 |
+
def display_gradcam_output(data: list,
|
78 |
+
classes: list[str],
|
79 |
+
inv_normalize: transforms.Normalize,
|
80 |
+
model,
|
81 |
+
target_layers,
|
82 |
+
targets=None,
|
83 |
+
number_of_samples: int = 10,
|
84 |
+
transparency: float = 0.60):
|
85 |
+
"""
|
86 |
+
Function to visualize GradCam output on the data
|
87 |
+
:param data: List[Tuple(image, label)]
|
88 |
+
:param classes: Name of classes in the dataset
|
89 |
+
:param inv_normalize: Mean and Standard deviation values of the dataset
|
90 |
+
:param model: Model architecture
|
91 |
+
:param target_layers: Layers on which GradCam should be executed
|
92 |
+
:param targets: Classes to be focused on for GradCam
|
93 |
+
:param number_of_samples: Number of images to print
|
94 |
+
:param transparency: Weight of Normal image when mixed with activations
|
95 |
+
"""
|
96 |
+
# Plot configuration
|
97 |
+
fig = plt.figure(figsize=(10, 10))
|
98 |
+
x_count = 5
|
99 |
+
y_count = 1 if number_of_samples <= 5 else math.floor(number_of_samples / x_count)
|
100 |
+
|
101 |
+
# Create an object for GradCam
|
102 |
+
cam = GradCAM(model=model, target_layers=target_layers, use_cuda=True)
|
103 |
+
|
104 |
+
# Iterate over number of specified images
|
105 |
+
for i in range(number_of_samples):
|
106 |
+
plt.subplot(y_count, x_count, i + 1)
|
107 |
+
input_tensor = data[i][0]
|
108 |
+
|
109 |
+
# Get the activations of the layer for the images
|
110 |
+
grayscale_cam = cam(input_tensor=input_tensor, targets=targets)
|
111 |
+
grayscale_cam = grayscale_cam[0, :]
|
112 |
+
|
113 |
+
# Get back the original image
|
114 |
+
img = input_tensor.squeeze(0).to('cpu')
|
115 |
+
img = inv_normalize(img)
|
116 |
+
rgb_img = np.transpose(img, (1, 2, 0))
|
117 |
+
rgb_img = rgb_img.numpy()
|
118 |
+
|
119 |
+
# Mix the activations on the original image
|
120 |
+
visualization = show_cam_on_image(rgb_img, grayscale_cam, use_rgb=True, image_weight=transparency)
|
121 |
+
|
122 |
+
# Display the images on the plot
|
123 |
+
plt.imshow(visualization)
|
124 |
+
plt.title(r"Correct: " + classes[data[i][1].item()] + '\n' + 'Output: ' + classes[data[i][2].item()])
|
125 |
+
plt.xticks([])
|
126 |
+
plt.yticks([])
|
127 |
+
|
128 |
+
|
129 |
+
def get_optimizer(model, lr, momentum=0, weight_decay=0, optimizer_type='SGD'):
|
130 |
+
"""Method to get object of stochastic gradient descent. Used to update weights.
|
131 |
+
|
132 |
+
Args:
|
133 |
+
model (Object): Neural Network model
|
134 |
+
lr (float): Value of learning rate
|
135 |
+
momentum (float): Value of momentum
|
136 |
+
weight_decay (float): Value of weight decay
|
137 |
+
optimizer_type (str): Type of optimizer SGD or ADAM
|
138 |
+
|
139 |
+
Returns:
|
140 |
+
object: Object of optimizer class to update weights
|
141 |
+
"""
|
142 |
+
if optimizer_type == 'SGD':
|
143 |
+
optimizer = optim.SGD(model.parameters(), lr=lr, momentum=momentum)
|
144 |
+
elif optimizer_type == 'ADAM':
|
145 |
+
optimizer = optim.Adam(model.parameters(), lr=lr, weight_decay=weight_decay)
|
146 |
+
return optimizer
|
147 |
+
|
148 |
+
def get_StepLR_scheduler(optimizer, step_size, gamma):
|
149 |
+
"""Method to get object of scheduler class. Used to update learning rate
|
150 |
+
|
151 |
+
Args:
|
152 |
+
optimizer (Object): Object of optimizer
|
153 |
+
step_size (int): Period of learning rate decay
|
154 |
+
gamma (float): Number to multiply with learning rate
|
155 |
+
|
156 |
+
Returns:
|
157 |
+
object: Object of StepLR class to update learning rate
|
158 |
+
"""
|
159 |
+
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=step_size, gamma=gamma, verbose=True)
|
160 |
+
return scheduler
|
161 |
+
|
162 |
+
def get_ReduceLROnPlateau_scheduler(optimizer, factor, patience):
|
163 |
+
"""Method to get object of scheduler class. Used to update learning rate
|
164 |
+
|
165 |
+
Args:
|
166 |
+
optimizer (Object): Object of optimizer
|
167 |
+
factor (float): Number to multiply with learning rate
|
168 |
+
patience (int): Number of epoch to wait
|
169 |
+
|
170 |
+
Returns:
|
171 |
+
object: Object of StepLR class to update learning rate
|
172 |
+
"""
|
173 |
+
scheduler = optim.lr_scheduler.ReduceLROnPlateau(optimizer, factor=factor, patience=patience, verbose=True)
|
174 |
+
return scheduler
|
175 |
+
|
176 |
+
def get_OneCycleLR_scheduler(optimizer, max_lr, epochs, steps_per_epoch, max_at_epoch, anneal_strategy, div_factor, final_div_factor):
|
177 |
+
"""Method to get object of scheduler class. Used to update learning rate
|
178 |
+
|
179 |
+
Args:
|
180 |
+
optimizer (Object): Object of optimizer
|
181 |
+
max_lr (float): Maximum learning rate to reach during training
|
182 |
+
epochs (float): Total number of epoch
|
183 |
+
steps_per_epoch (int): Total steps in an epoch
|
184 |
+
max_at_epoch (int): Epoch to reach maximum learning rate
|
185 |
+
anneal_strategy (string): Strategy to interpolate between minimum and maximum lr
|
186 |
+
div_factor (int): Divisive factor to calculate intial learning rate
|
187 |
+
final_div_factor (int): Divisive factor to calculate minimum learning rate
|
188 |
+
|
189 |
+
Returns:
|
190 |
+
object: Object of StepLR class to update learning rate
|
191 |
+
"""
|
192 |
+
scheduler = optim.lr_scheduler.OneCycleLR(optimizer, max_lr=max_lr, epochs=epochs,
|
193 |
+
steps_per_epoch=steps_per_epoch,
|
194 |
+
pct_start=max_at_epoch/epochs,
|
195 |
+
anneal_strategy=anneal_strategy,
|
196 |
+
div_factor=div_factor,
|
197 |
+
final_div_factor=final_div_factor)
|
198 |
+
return scheduler
|
199 |
+
|
200 |
+
def get_criterion(loss_type='cross_entropy'):
|
201 |
+
"""Method to get loss calculation ctiterion
|
202 |
+
|
203 |
+
Args:
|
204 |
+
loss_type (str): Type of loss 'nll_loss' or 'cross_entropy' loss
|
205 |
+
|
206 |
+
Returns:
|
207 |
+
object: Object to calculate loss
|
208 |
+
"""
|
209 |
+
if loss_type == 'nll_loss':
|
210 |
+
criterion = F.nll_loss
|
211 |
+
elif loss_type == 'cross_entropy':
|
212 |
+
criterion = F.cross_entropy
|
213 |
+
return criterion
|
214 |
+
|
215 |
+
def get_learning_rate(model, optimizer, criterion, trainloader):
|
216 |
+
"""Method to find learning rate using LR finder.
|
217 |
+
|
218 |
+
Args:
|
219 |
+
model (Object): Object of model
|
220 |
+
optimizer (Object): Object of optimizer class
|
221 |
+
criterion (Object): Loss function
|
222 |
+
trainloader (Object): Object of dataloader class
|
223 |
+
|
224 |
+
Returns:
|
225 |
+
float: Learning rate suggested by lr finder
|
226 |
+
"""
|
227 |
+
# Create object and perform range test
|
228 |
+
lr_finder = LRFinder(model, optimizer, criterion)
|
229 |
+
lr_finder.range_test(trainloader, end_lr=100, num_iter=100)
|
230 |
+
|
231 |
+
# Plot result and store suggested lr
|
232 |
+
plot, suggested_lr = lr_finder.plot()
|
233 |
+
|
234 |
+
# Reset model and optimizer
|
235 |
+
lr_finder.reset()
|
236 |
+
|
237 |
+
return suggested_lr
|