terrlo / DS2

Dual Scale Dual Similarity (DS2)

Supported platform:

Linux

Package installation:

conda install the following packages in the recommended order:

• pytorch==1.7.1 torchvision==0.8.2 cudatoolkit=11.0 -c pytorch
• scikit-learn
• python=3.8.8
• opencv -c conda-forge
• termcolor -c omnia
• pillow=9.1.0

MVTec dataset preparation:

Follow the steps to create mvtec dataset (for evaluation stage) and mvtec_train dataset (for pretraining stage)

• create the dataset folder inside the project root: mkdir dataset
• move into the folder: cd dataset
• create the mvtec folder: mdkir mvtec
• move into the folder: cd mvtec
• Download MVTec AD dataset: wget https://www.mydrive.ch/shares/38536/3830184030e49fe74747669442f0f282/download/420938113-1629952094/mvtec_anomaly_detection.tar.xz
• unzip: tar -xf mvtec_anomaly_detection.tar.xz
• remove zip file: rm mvtec_anomaly_detection.tar.xz
• move to parent folder (dataset/): cd ..
• create the mvtec_train folder: mkdir mvtec_train
• move to parent folder (project root): cd ..
• run the command ./tools/make_mvtec_train.sh to create the mvtec_train dataset for pretraining (Note: replace the $PROJ_ABS_PATH to the absolute path of your project on your local machine)

MVTec LOCO dataset preparation:

• visit https://www.mvtec.com/company/research/datasets/mvtec-loco, and fill out the form required in the website to download the dataset
• once downloaded, unzip the file, rename the outermost folder name to mvtecloco, and move the files to folder DS2/dataset/, such that the breakfast_box category is located at DS2/dataset/mvtecloco/breakfast_box

KSDD2 dataset preparation:

• download the dataaset: wget https://go.vicos.si/kolektorsdd2 -O KSDD2.zip
• unzip the file: unzip KSDD2.zip
• move the train and test folders to DS2/dataset/KSDD2/

MTD dataset preparation:

• download the dataset: git clone https://github.com/abin24/Magnetic-tile-defect-datasets..git
• rename the folder name from Magnetic-tile-defect-datasets. to MTD
• move the MTD folder to DS2/dataset/

Run pretraining code for DS2 on MVTec (Stage 1):

• To run the pretraining code for DS2, execute ./tools/ds2_pretrain_mvtec.sh. The default setting requires two GPUs (preferably A100-40GB and above). The seed range is [1,5]
• The output log (including model checkpoints) will be stored in folder output/mvtec_$TIMESTAMP/

Run evaluation code for DS2 on MVTec (Stage 2):

• After pretraining, execute ./tools/ds2_eval_mvtec.sh to perform anomaly detection on test split. The default setting requires one GPU.
  • Inside file ./tools/ds2_eval_mvtec.sh, set pretrained_model_dir to the checkpoint models' folder output/mvtec_$TIMESTAMP/
• The evaluation output will be stored in folder logs/

Run pretraining code for CutPaste_(3-way, one-for-all) on MVTec:

• To run the pretraining code for CutPaste, execute ./tools/cutpaste_pretrain_mvtec.sh. The default setting requires two GPUs (preferably A100-40GB and above). The seed range is [1,5]
• The output log (including model checkpoints) will be stored in folder output/mvtec_$TIMESTAMP_cutpaste/

Run evaluation code for CutPaste_(3-way, one-for-all) on MVTec:

• After pretraining, execute ./tools/cutpaste_eval_mvtec.sh to perform anomaly detection on test split. The default setting requires one GPU.
  • Inside file ./tools/cutpaste_eval_mvtec.sh, set pretrained_model_dir to the checkpoint models' folder output/mvtec_$TIMESTAMP_cutpaste/
• The evaluation output will be stored in folder logs/

Run evaluation code for DS2 on MVTec LOCO:

• After pretraining, execute ./tools/ds2_eval_loco.sh. The default setting requires one GPU.
  • Inside file ./tools/ds2_eval_loco.sh, set pretrained_model_dir to the checkpoint models' folder output/mvtec_$TIMESTAMP/
• The evaluation output will be stored in folder logs/

Run evaluation code for DS2 on KSDD2:

• After pretraining, execute ./tools/ds2_eval_ksdd2.sh. The default setting requires one GPU.
  • Inside file ./tools/ds2_eval_ksdd2.sh, set pretrained_model_dir to the checkpoint models' folder output/mvtec_$TIMESTAMP/
• The evaluation output will be stored in folder logs/

Run evaluation code for DS2 on MTD:

• After pretraining, execute ./tools/ds2_eval_mtd.sh. The default setting requires one GPU.
  • Inside file ./tools/ds2_eval_mtd.sh, set pretrained_model_dir to the checkpoint models' folder output/mvtec_$TIMESTAMP/
• The evaluation output will be stored in folder logs/

Acknowledgement

The main architecture is adapted from https://github.com/zdaxie/PixPro (Propagate Yourself: Exploring Pixel-Level Consistency for Unsupervised Visual Representation Learning)

The implementation of DistAug and RotPred is adapted from https://github.com/google-research/deep_representation_one_class (LEARNING AND EVALUATING REPRESENTATIONS FOR DEEP ONE-CLASS CLASSIFICATION)

The implementation of CutPaste is adapted from https://github.com/Runinho/pytorch-cutpaste