MCL on (CIFAR-10, CIFAR-100)

Official code for MCL. (Masked Contrastive Learning for Anomaly Detection IJCAI-2021) This code includes SEI eval on CIFAR-100. (AUROC around 94.) Some codes are from SimCLR-CIFAR10.

Dependencies

• pytorch >=1.2
• torchvision >=0.4.0
• hydra >=0.11.3
• tqdm >=4.45.0

Install Hydra

Hydra is a python framework to manage the hyperparameters during training and evaluation. Install with:

pip install hydra-core --upgrade

Usage

Dataset download and preprocess (4-way rotations augmented dataset with rotation label) python utils/data_preprocess.py

Train MCL: python mcl_main.py

Use the following prefix to train MCL with single GPU : CUDA_VISIBLE_DEVICES="GPU_number"

All the hyperparameters are available in mcl_config.yml, which could be overrided from the command line.

Evaluate trained model.

Download MCL trained model.

place trained model (epoch_800.pt) in following directory.

./logs/model_name(default : MCL)/ckpt/epoch_800.pt

To evaluate pretrained model, set evaluate flag to True and load_epoch to 800 in mcl_config.yml.

Then run python mcl_main.py.

Experimental results are stored in ./logs/model_name(default : MCL)/SEI performance_summary.txt file.

Notes

Current version of SEI is not optimized enough, thus it might be slow.