Upload folder using huggingface_hub

README.md

@@ -1,98 +1,48 @@
-*Note: You may find it easier to read about Intel Transfer Learning tool, follow the Get
-Started guide, and browse the API material from our published documentation site
-https://intelai.github.io/transfer-learning.*
+# Setting parameters and configurations
 
-<!-- SkipBadges -->
+Please set the following in the finetuning.yaml file:
 
-# Intel® Transfer Learning Tool
+* **num_workers:** number of sub-processes or threads to use for data loading. Setting the argument num_workers as a positive integer will turn on multi-process data loading. (Default=32)
 
-Transfer learning workflows use the knowledge learned by a pre-trained model on
-a large dataset to improve the performance of a related problem with a smaller
-dataset.
+* **precision:** precision of data type in which model to be fine-tuned. Choices are [float32, bfloat16]
 
-## What is Intel® Transfer Learning Tool
+* **fine_tune:** set 'True' to run SimSiam or CutPaste self-supervised learning using Intel Transfer Learning Tool APIs. Set 'False' to run a pre-trained backbone by providing a model path under 'model_path' category
 
-Intel® Transfer Learning Tool makes it easier and faster for you to
-create transfer learning workflows across a variety of AI use cases. Its
-open-source Python\* library leverages public pretrained model hubs,
-Intel-optimized deep learning frameworks, and your custom dataset to efficiently
-generate new models optimized for Intel hardware.
+* **output_path:** path to save the checkpoints or final model
 
-This project documentation provides information, resource links, and instructions for the Intel
-Transfer Learning Tool as well as Jupyter\* notebooks and examples that
-demonstrate its usage.
+* **tlt_wf_path:** set by default to point to the workflow in the Intel Transfer Learning Tool
 
-**Features:**
-* Supports PyTorch\* and TensorFlow\*
-* Select from over [100 image classification and text classification models](Models.md) from
-  Torchvision, PyTorch Hub, TensorFlow Hub, Keras, and Hugging Face
-* Use your own custom dataset or get started quickly with built-in datasets
-* Automatically create a trainable classification layer customized for your dataset
-* Pre-process your dataset using scaling, cropping, batching, and splitting
-* Use APIs for prediction, evaluation, and benchmarking
-* Export your model for deployment or resume training from checkpoints
+* **dataset:**
+  * **root_dir:** path to the root directory of MVTEC dataset
+  * **category_type:** category type within MVTEC dataset, e.g.: hazelnut or all (for running all categories in MVTEC)
+  * **batch_size:** batch size for inference (Default=32)
+  * **image_size:** each image resized to this size (Default=224x224)
 
-**Intel Optimizations:**
-* Boost performance with Intel® Optimization for TensorFlow and Intel® Extension for PyTorch
-* Quantize to INT8 to reduce model size and speed up inference using Intel® Neural Compressor
-* Optimize model for FP32 inference using Intel Neural Compressor
-* Reduce training time with auto-mixed precision for select hardware platforms
-* Further reduce training time with multinode training for PyTorch
+* **model:** Options to select when running with a pre-trained backbone, no fine-tuning on custom dataset
+  * **name:** pretrained backbone model E.g.: resnet50, resnet18
+  * **layer:** intermediate layer from which features will be extracted
+  * **pool:** pooling kernel size for average pooling
+  * **feature_extractor:** select the type of modelling and subsequent feature extractor. Options are:
+    * pretrained -  No fine-tuning on custom dataset, features will be extracted from pretrained model which is set in model/name
+    * simsiam - SimSiam self-supervised training on custom dataset
+    * cutpaste - CutPaste self-supervised training on custom dataset 
 
-## How the Intel Transfer Learning Tool Works
+* **simsiam:** Set when 'feature_extractor' is set to simsiam. For details about simsiam method, please refer to https://arxiv.org/abs/2011.10566
+  * **batch_size:** batch size for fine-tuning (Default=64)
+  * **epochs:** number of epochs to fine-tune the model
+  * **optim:** optimization algorithm E.g.: sgd, adam
+  * **model_path:** path to save the checkpoints or final model
+  * **ckpt:** flag to specify whether intermediate checkpoints should be saved or not
 
-The Intel Transfer Learning Tool lets you train AI models with TensorFlow or
-PyTorch using either no-code command line interface (CLI) commands at a bash
-prompt, or low-code application programming interface (API) calls from a Python
-script.
+* **cutpaste:** Set when 'feature_extractor' is set to cutpaste. For details about cutpaste method, please refer to https://arxiv.org/abs/2104.04015
+  * **cutpaste_type:**  type of image augmentation for cutpaste fine-tuning, choices are ['normal', 'scar', '3way', 'union'].
+  * **head_layer:**     number of fully-connected layers on top of average pooling layer followed by the last linear layer of backbone network
+  * **freeze_resnet:**  number of epochs till only head layers will be trained. After this, complete network will be trained.
+  * **batch_size:** batch size for fine-tuning (Default=64)
+  * **epochs:** number of epochs to fine-tune the model
+  * **optim:** optimization algorithm E.g.: sgd, adam
+  * **model_path:** path to save the checkpoints or final model
+  * **ckpt:** flag to specify whether intermediate checkpoints should be saved or not
 
-Use your own dataset or select an existing image or text classification dataset listed in the
-[public datasets](DATASETS.md) documentation. Construct your own CLI or API commands for training, evaluation,
-and optimization using the TensorFlow or PyTorch framework, and finally export
-your saved model optimized for inference on Intel CPUs.
+* **pca_thresholds:** percentage of variance ratio to be retained. Number of PCA components are selected according to it
 
-An overview of the Intel Transfer Learning Tool flow is shown in this
-figure:
-
-<p align="center"><b>Intel Transfer Learning Tool Flow</b></p>
-
-<img alt="Intel Transfer Learning Tool Flow" title="Intel Transfer Learing Tool Flow" src="images/TLT-tool_flow.svg" width="600">
-
-## Get Started
-
-The [Get Started](GetStarted.md) guide walks you through the steps to check
-system requirements, install, and then run the tool with a couple of examples
-showing no-code CLI and low-code API approaches. After that, you can check out
-these additional CLI and API [Examples](examples/README.md).
-
-<!-- ExpandGetStarted-Start -->
-As described in the [Get Started](GetStarted.md) guide, once you have a Python
-3.9 environment set up, you do a basic install of the Intel Transfer Learning
-Tool using:
-
-```
-pip install intel-transfer-learning-tool
-```
-
-Then you can use the Transfer Learning Tool CLI interface (tlt) to train a
-TensorFlow image classification model (resnet_v1_50), download and use an
-existing built-in dataset (tf_flowers), and save the trained model to
-`/tmp/output` using this one command:
-
-```
-tlt train --framework tensorflow --model-name resnet_v1_50 --dataset-name tf_flowers \
-   --output-dir /tmp/output --dataset-dir /tmp/data
-```
-
-Use `tlt --help` to see the list of CLI commands.  More detailed help for each
-command can be found using, for example, `tlt train --help`.
-
-<!-- ExpandGetStarted-End -->
-
-## Support
-
-The Intel Transfer Learning Tool team tracks bugs and enhancement requests using
-[GitHub issues](https://github.com/IntelAI/transfer-learning-tool/issues). Before submitting a
-suggestion or bug report, search the existing GitHub issues to see if your issue has already been reported.
-
-See [Legal Information](Legal.md) for Disclaimers, Trademark, and Licensing information.

eval.yaml

@@ -0,0 +1,33 @@
+num_workers: 32
+precision: float32
+fine_tune: false
+output_path: ./output
+tlt_wf_path: ./transfer-learning/workflows/vision_anomaly_detection
+dataset:
+  root_dir: ./data
+  category_type: all
+  batch_size: 32
+  image_size: 224
+model:
+  name: resnet50
+  layer: layer3
+  pool: 2
+  feature_extractor: simsiam
+simsiam:
+  batch_size: 64
+  epochs: 2
+  optim: sgd
+  model_path: ./output
+  ckpt: true
+  initial_ckpt: null
+cutpaste:
+  cutpaste_type: 3way
+  head_layer: 2
+  freeze_resnet: 20
+  batch_size: 64
+  epochs: 2
+  optim: sgd
+  model_path: ./output
+  ckpt: true
+pca:
+  pca_thresholds: 0.99

finetuning.yaml

@@ -0,0 +1,36 @@
+num_workers: 32 # number of parallel threads to process the data
+precision: float32 # supported options are [float32, bfloat16]
+fine_tune : true
+output_path: ./output/  # path to save the checkpoints or final model
+tlt_wf_path: ./transfer-learning/workflows/vision_anomaly_detection
+dataset:
+  root_dir: ./data/  # full path of root directory of MVTEC dataset
+  category_type: hazelnut # category type within MVTEC dataset, i.e. hazelnut or all (for running all categories in MVTEC)
+  batch_size: 32 # inference batch size
+  image_size: 224 # each image resized to this size (224x224)
+model: 
+  name: resnet50 # pretrained backbone model ..choices are [resnet50, resnet18]
+  layer: layer3  # intermediate layer from which features will be extracted
+  pool: 2 # pooling kernel size for average pooling
+  feature_extractor: cutpaste # choices are [pretrained, cutpaste, simsiam]
+#pretrained -  No fine-tuning on custom dataset, features will be extracted from pretrained ResNet model
+#simsiam - fine-tune resnet model on custom dataset using simsiam self-supervised technique
+#cutpaste - fine-tune resnet model on custom datset using cutpaste self-supervised technique
+simsiam:
+  batch_size: 64 # fine-tuning batch size
+  epochs: 2 # number of epochs to fine-tune the model
+  optim: 'sgd' # optimizer
+  model_path: './output' # path to save the checkpoints or final model
+  ckpt: true  # flag for whether intermediate checkpoints would be saved or not
+  initial_ckpt:
+cutpaste:
+  cutpaste_type: '3way' # choices are ['normal', 'scar', '3way', 'union'] for image augmentation
+  head_layer: 2 # number of perceptron layers appended towards the end of ResNet layers
+  freeze_resnet: 20 # number of epochs till resnet layers will be frozen and only head layers will be trained
+  batch_size: 64 # fine-tuning batch size
+  epochs: 1 # number of epochs to fine-tune the model
+  optim: 'sgd' # optimizer
+  model_path: './output' # path to save the checkpoints or final model
+  ckpt: true # flag for whether intermediate checkpoints would be saved or not
+pca:
+  pca_thresholds: 0.99 # PCA select number of components such that it ensures to retain the variance ratio specified