Awesome
AnomalyDetection-SoftPatch/SoftPatch+
This repository contains codes for the official implementation in PyTorch of NeurIPS 2022 paper "SoftPatch: Unsupervised Anomaly Detection with Noisy Data" and its improved version SoftPatch+.
Quick Start
Requirement
Our results were computed using Python 3.8 with packages and respective version noted in requirements.txt.
MVTec-AD
- Default Training. To train SoftPatch on MVTec AD with 0.1 additional noise samples, run
datapath=/path_to_mvtec_folder/mvtec
datasets=('bottle' 'cable' 'capsule' 'carpet' 'grid' 'hazelnut'
'leather' 'metal_nut' 'pill' 'screw' 'tile' 'toothbrush' 'transistor' 'wood' 'zipper')
dataset_flags=($(for dataset in "${datasets[@]}"; do echo '-d '$dataset; done))
python main.py --dataset mvtec --data_path ../../MVTec --noise 0.1 "${dataset_flags[@]}" --gpu 0
The default setting in run_mvtec.sh runs with 224x224 image size using a WideResNet50-backbone pretrained on ImageNet.
- Expected Performance. Training on 1 GPU (NVIDIA Tesla V100 32GB) results in following performance.
| Row Names | image_auroc | pixel_auroc |
|---|---|---|
| mvtec_bottle | 1.0000 | 0.9878 |
| mvtec_cable | 0.9904 | 0.9862 |
| mvtec_capsule | 0.9654 | 0.9883 |
| mvtec_carpet | 0.9965 | 0.9920 |
| mvtec_grid | 1.0000 | 0.9939 |
| mvtec_hazelnut | 1.0000 | 0.9906 |
| mvtec_leather | 1.0000 | 0.9931 |
| mvtec_metal_nut | 0.9987 | 0.9845 |
| mvtec_pill | 0.9562 | 0.9798 |
| mvtec_screw | 0.9526 | 0.9944 |
| mvtec_tile | 0.9866 | 0.9645 |
| mvtec_toothbrush | 0.9931 | 0.9860 |
| mvtec_transistor | 0.9974 | 0.9064 |
| mvtec_wood | 0.9854 | 0.9714 |
| mvtec_zipper | 0.9753 | 0.9892 |
| Mean | 0.9865 | 0.9805 |
- Parameter Setting.
To choose other noise discriminator, use the --weight_method argument with 'lof', 'nearest', 'gaussian' or 'lof_gpu'. 'lof_gpu' supports computing LOF using the GPU which usually faster.
To
BTAD
To train SoftPatch on BTAD, run:
datapath=/path_to_btad_folder/BTAD
datasets=('01' '02' '03')
dataset_flags=($(for dataset in "${datasets[@]}"; do echo '-d '$dataset; done))
python main.py --dataset btad --data_path ../../BTAD --noise 0 "${dataset_flags[@]}" --seed 0 \
--gpu 1 --resize 512 --imagesize 512 --sampling_ratio 0.01
The default setting in run_btad.sh runs with 512x512 image size using a WideResNet50-backbone pretrained on ImageNet.
| Row Names | image_auroc | pixel_auroc |
|---|---|---|
| btad_01 | 0.9981 | 0.9761 |
| btad_02 | 0.9343 | 0.9662 |
| btad_03 | 0.9969 | 0.9935 |
| Mean | 0.9764 | 0.9786 |
Comments
- Our codebase for the coreset construction builds heavily on PatchCore codebase. Thanks for open-sourcing!
Citation
Please cite the following paper if this dataset helps your project:
@misc{xisoftpatch,
title={SoftPatch: Unsupervised Anomaly Detection with Noisy Data},
author={Xi, Jiang and Liu, Jianlin and Wang, Jinbao and Nie, Qiang and Kai, WU and Liu, Yong and Wang, Chengjie and Zheng, Feng},
booktitle={Advances in Neural Information Processing Systems}
}
License
This project is licensed under the Apache-2.0 License.