ShuaiLYU / REB

REB：Reducing Biases in Representation for Industrial Anomaly Detection

REB

official source code for REB：Reducing Biases in Representation for Industrial Anomaly Detection

Install

$ git clone https://github.com/ShuaiLYU/REB
$ cd REB
$ conda create -n reb python==3.10.6
$ conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.3 -c pytorch

$ pip install -r requirements.txt

Training on Mvtec AD

Run commands below to reproduce results on Mvtec AD

1. you are advised to download Mvtec ad dataset with saliency from BaiduNetdisk (code: 1234) or OneDrive.
   you can also generate saliency by yourself with teh EDN saliency model

2. modify the dataset path and OUTPUT path in global_param.py according to your personal config.
   the best K for Mvtec is 9. set K value in argparse or global_param.py

3. Run commands below to train

$ cd  REB/projects/reb_mvtec
#  Self-supervied learning  (fine-turning ImageNet-pretrained model with DefectMaker))  Resnet18
python main.py  -exp_name res18_dm_com6_bs1024_epo300   -K 9 -run_name run1  -run_mode 0

# only run LDKNN (after fine-turning ImageNet-pretrained model with DefectMaker)  Resnet18
$ python main.py  -exp_name res18_dm_com6_bs1024_epo300    -K 9 -run_name run1  -run_mode 1


# DefectMaker + LDKNN  (REB)  Resnet18
$ python main.py  -exp_name res18_dm_com6_bs1024_epo300  -K 9  -run_name run1  -run_mode 2

Training on Mvtec LOCO

Run commands below to reproduce results on Mvtec AD LOCO

1. we don't use saliency for Mvtec LOCO

2. modify the dataset path and OUTPUT path in global_param.py according to your personal config.
   the best K for Mvtec LOCO is 45. set K value in argparse or global_param.py

3. Run commands below to train

$ cd  REB/projects/reb_mvtec_loco
#  Self-supervied learning  (fine-turning ImageNet-pretrained model with DefectMaker))  Resnet18
python main.py  -exp_name res18_dm_com6_bs1024_epo300  -K 45 -run_name run1  -run_mode 0

# only run LDKNN (after fine-turning ImageNet-pretrained model with DefectMaker)  Resnet18
$ python main.py  -exp_name res18_dm_com6_bs1024_epo300 -K 45  -run_name run1  -run_mode 1

# only run LDKNN ( directly use ImageNet-pretrained model)   Resnet18
$ python main.py  -exp_name res18_imagenet -K 45 -run_name run1  -run_mode 1

# DefectMaker + LDKNN  (REB)   Resnet18
$ python main.py  -exp_name res18_dm_com6_bs1024_epo300  -K 45 -run_name run1  -run_mode 2


# only run LDKNN ( directly use ImageNet-pretrained model)  Resnet18  
$ python main.py  -exp_name res18_imagenet   -K 45 -run_name run1  -run_mode 1

# only run LDKNN ( directly use ImageNet-pretrained model)  wideRes50 
$ python main.py  -exp_name wr50_imagenet   -K 45 -run_name run1  -run_mode 1

# only run LDKNN ( directly use ImageNet-pretrained model)  wideRes101 
$ python main.py  -exp_name wr101_imagenet   -K 45  -run_name run1  -run_mode 1

Tutorials

Bezier_gen

• visualize Bezier Curve

DefectMaker

• wrap multiple DefectMaker combinations with your data