f-AnoGAN: Fast unsupervised anomaly detection with generative adversarial networks

Code for reproducing f-AnoGAN training and anomaly scoring presented in “f-AnoGAN: Fast Unsupervised Anomaly Detection with Generative Adversarial Networks”, implemented in Pytorch.

Requirements

scikit-learn==0.21.2

torch==1.4.0

torchvision==0.5.0

Run

Generative adversarial training

$ python ./gan_training.py --n_epochs   50 --dataset    mnist --latent_dim 128 --img_size   32 --channels   1 --abnormal_class 0 --device     cuda --out        ckpts

Encoder training

$ python ./enc_training.py --n_epochs   5 --dataset    mnist --latent_dim 128 --img_size   32 --channels   1 --abnormal_class 0 --device     cuda --out        ckpts --G_path     ckpts/G_epoch49.pt --D_path     ckpts/D_epoch49.pt