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:

Hyperparameter	Value
Number of epochs	265
Batch size	32
Learning rate	0.03
Input size	256

AUC comparison of our code and paper results

Defect Name	CutPaste binary (ours)	CutPaste binary (paper's)	CutPaste 3way (ours)	CutPaste 3way (paper's)
tile	84.1	95.9	78.9	93.4
wood	89.5	94.9	89.2	98.6
pill	88.7	93.4	78.7	92.4
leather	98.7	99.7	84.8	100.0
hazelnut	98.8	91.3	80.8	97.3
screw	89.2	54.4	56.6	86.3
cable	83.3	87.7	75.7	93.1
toothbrush	94.7	99.2	78.6	98.3
capsule	80.2	87.9	70.8	96.2
carpet	57.9	67.9	26.1	93.1
zipper	99.5	99.4	85.7	99.4
metal_nut	91.5	96.8	89.7	99.3
bottle	98.5	99.2	75.7	98.3
grid	99.9	99.9	73.0	99.9
transistor	84.4	96.4	85.5	95.5

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

t-SNE visualisation of embeddings