PANDA
Official PyTorch implementation of “PANDA: Adapting Pretrained Features for Anomaly Detection and Segmentation” (CVPR 2021).
Virtual Environment
Use the following commands:
cd path-to-PANDA-directory
virtualenv venv --python python3
source venv/bin/activate
pip install -r requirements.txt --find-links https://download.pytorch.org/whl/torch_stable.html
Data Preparation
Use the following commands:
cd path-to-PANDA-directory
mkdir data
Download:
Extract these files into path-to-PANDA-directory/data and unzip tiny.zip
Experiments
To replicate the results on CIFAR10, FMNIST for a specific normal class with EWC:
python panda.py --dataset=cifar10 --label=n --ewc --epochs=50
python panda.py --dataset=fashion --label=n --ewc --epochs=50
To replicate the results on CIFAR10, FMNIST for a specific normal class with early stopping:
python panda.py --dataset=cifar10 --label=n
python panda.py --dataset=fashion --label=n
Where n indicates the id of the normal class.
To run experiments on different datasets, please set the path in utils.py to the desired dataset.
OE Experiments
To replicate the results on CIFAR10 for a specific normal class:
python outlier_exposure.py --dataset=cifar10 --label=n
Where n indicates the id of the normal class.
Further work
See our new paper “Mean-Shifted Contrastive Loss for Anomaly Detection” which achieves state-of-the-art anomaly detection performance on multiple benchmarks including 97.5% ROC-AUC on the CIFAR-10 dataset.
Citation
If you find this useful, please cite our paper:
@inproceedings{reiss2021panda,
title={PANDA: Adapting Pretrained Features for Anomaly Detection and Segmentation},
author={Reiss, Tal and Cohen, Niv and Bergman, Liron and Hoshen, Yedid},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={2806--2814},
year={2021}
}