Awesome

CutPaste

CutPaste: image from paper

Unofficial implementation of Google's "CutPaste: Self-Supervised Learning for Anomaly Detection and Localization" in PyTorch

Installation

To rerun experiments or try on your own dataset, first clone the repository and install requirements.txt.

$ git clone https://github.com/LilitYolyan/CutPaste.git
$ cd CutPaste
$ pip install -r requirements.txt

Self-supervised training

Run train.py to train self-supervised model on MVTec dataset

For 3 way classification head

$ python train.py --dataset_path path/to/your/dataset/ --num_class 3

For binary classification head

$ python train.py --dataset_path path/to/your/dataset/ --num_class 2

For feature extractor any torchvision model can be used. For example to use EfficientNet(B4)

$ python train.py --dataset_path path/to/your/dataset/ --encoder efficientnet_b4

To track training process with TensorBoard

tensorboard --logdir logdirs

Anomaly Detection

To run anomaly detection for MVTec with Gaussian Density Estimator

$ python anomaly_detection.py --checkpoint path/to/your/weights --data path/to/mvtec

TODO

• Self-supervised model
• Gaussian Density Estimator
• EfficientNet Implementation
• t-SNE visualization of representations
• Localization
• Grad-CAM visualization
• Implement GDE in PyTorch (too slow with sklearn)

Any contribution is appreciated!

Experiment Results

For more experiment results go to "experiments.md"

To train self-supervised model we used same hyperparameters as was used in paper:

AUC comparison of our code and paper results

ROC curves using embeddings from binary classification for self-supervised learning

t-SNE visualisation of embeddings