jaeminyoo / ssl-ad

Data Augmentation is a Hyperparameter: Cherry-picked Self-Supervision for Unsupervised Anomaly Detection is Creating the Illusion of Success (TMLR)

Data Augmentation is a Hyperparameter for Self-supervised Anomaly Detection

This is our implementation for Data Augmentation is a Hyperparameter: Cherry-picked Self-Supervision for Unsupervised Anomaly Detection is Creating the Illusion of Success, published in Transactions on Machine Learning Research (TMLR).

Prerequisites

Our implementation is based on Python 3.8.12 and PyTorch 1.10.1. Refer to requirements.txt for the required packages.

How to Run

You can type the following command to train a denoising autoencoder (DAE) model with a specified augmentation function, which is flip in this case:

cd src ; python main.py --data cifar10 --normal-class 0 --augment flip 

The --normal-class option chooses a class to consider as normal. The script uses only the chosen class as training data, treating the rest as anomalous. The training is done only once for each normal class, but the evaluation is done in the one-vs-one scheme; we take the AUC value from each pair of classes and use it to run the comprehensive Wilcoxon test.

The script automatically downloads the specified dataset and prints the following log during the training (assuming that the dataset is already downloaded):

Files already downloaded and verified
Files already downloaded and verified
Parameters: 4159235
[epoch   0] [0.0980] [0.4849]                                                                             
[epoch   1] [0.0748] [0.5054]                                                                             
[epoch   2] [0.0475] [0.6054]                                                                             
[epoch   3] [0.0358] [0.6607]                                                                             
[epoch   4] [0.0299] [0.6529]  
...

The first value at each line represents the training loss, which is the reconstruction error, and the second value is the AUC measured in the one-vs-rest scheme. No early stopping is used, since we have no validation data. The result is stored at the out-tmp directory. You can change the arguments of main.py to run experiments in other settings and configurations.

Reference

Please cite our paper if you utilize our code in your research:

@article{
yoo2023data,
title={Data Augmentation is a Hyperparameter: Cherry-picked Self-Supervision for Unsupervised Anomaly Detection is Creating the Illusion of Success},
author={Jaemin Yoo and Tiancheng Zhao and Leman Akoglu},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2023},
url={https://openreview.net/forum?id=HyzCuCV1jH},
note={}
}