Elsa: Energy-based Learning forSemi-supervised Anomaly Detection

Official PyTorch implementation of "Elsa: Energy-based Learning forSemi-supervised Anomaly Detection" (BMVC 2021) by Sungwon Han*, Hyeonho Song*, Seungeon Lee, Sungwon Park, Meeyoung Cha.

Requirements

Environment

• python == 3.7
• torch == 1.7.1
• torchvision == 0.8.2
• CUDA == 10.1
• scikit-learn == 0.23.2
• tensorboard == 2.4.0
• tensorboardX == 2.4
• torchlars == 0.1.2
• diffdist == 0.1
• soyclustering == 0.2.0

Pretraining

Please refer following repositories for original codes.

simCLR: https://github.com/google-research/simclr

CSI: https://github.com/alinlab/CSI

To pretrain simCLR or CSI on CIFAR-10, try the following commands.

cd elsa_finetune

simCLR (ELSA)

python train.py --dataset cifar10 --model resnet18 \
--mode simclr --one_class_idx 0 --ratio_pollution 0.1 \
--batch_size 512 --epochs 1000 --single_device 0

mv ./logs0/cifar10_resnet18_unsup_simclr_one_class_0 ../pretrained_result

CSI (ELSA++)

python train.py --dataset cifar10 --model resnet18 \
--mode simclr_CSI --one_class_idx 0 --ratio_pollution 0.1 \
--batch_size 128 --epochs 1000 --shift_trans_type rotation --single_device 0

mv ./logs0/cifar10_resnet18_unsup_simclr_CSI_shift_rotation_one_class_0 ../pretrained_result

FineTuning

To pretrain the model on CIFAR-10, try the following commands.

• ELSA uses pretrained model of simclr and ELSA++ uses pretrained model of simclr_CSI.
• Argument "known_normal" and "ratio_pollution" of finetuning must be same to "one_class_idx" and "ratio_pollution" in pretraining.

ELSA

python ELSA.py --save_dir ./ --load_path ../pretrained_result/last.model \
--n_known_outlier 1 --known_normal 0 --known_outlier 1 \
--ratio_known_normal 0.1 --ratio_known_outlier 0.1 --ratio_pollution 0.1 \
 --batch_size 64 --n_cluster 50 --optimizer adam --lr 1e-4 --weight_decay 0.0 --n_epochs 50

ELSA++

python ELSApp.py --save_dir ./ --load_path ../pretrained_result/last.model \
--n_known_outlier 1 --known_normal 0 --known_outlier 1 \
--ratio_known_normal 0.1 --ratio_known_outlier 0.1 --ratio_pollution 0.1 \
 --batch_size 64 --n_cluster 50 --optimizer adam --lr 1e-4 --weight_decay 0.0 --n_epochs 50

License

Distributed under the MIT License. See LICENSE.txt for more information.