f-AnoGAN is a GAN for anomaly detection. One of the features of this GAN is that two adversarial networks (Generator and Discriminator) and Encoder are trained separately. In addition, an anomaly score is computed by both a discriminator feature residual error and an image reconstruction error.

Japanese version

Papar

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

Github

• tSchlegl/f-AnoGAN: Code for reproducing f-AnoGAN training and anomaly scoring
• PyTorch-GAN/wgan_gp.py at master · eriklindernoren/PyTorch-GAN
• PyTorch-GAN/dcgan.py at master · eriklindernoren/PyTorch-GAN

mnist/model.py, fanogan/train_wgangp.py and fanogan/train_encoder_izif.py are modified eriklindernoren's wgan_gp.py for f-AnoGAN.
mvtec_ad/model.py is modified eriklindernoren's dcgan.py for f-AnoGAN.

Python 3.6 or later
PyTorch 1.x

Matplotlib
Numpy
pandas
Pillow
scikit-learn

Please run below in order on the CPU.

python setup.py install
cd mnist

python train_wgangp.py --training_label 1 --seed 2 --n_epochs 20

python train_encoder_izif.py --training_label 1 --seed 2 --n_epochs 20

python test_anomaly_detection.py --training_label 1

After Step: 3, score.csv will be generated in the directory results.

See f-AnoGAN_MNIST.ipynb about data visualization for score.csv.

python save_compared_images.py --seed 4 --n_iters 0 --n_grid_lines 10

Compared images are saved under f-AnoGAN/mnist/results/images_diff.

Please run below in order.

python setup.py install
cd your_own_dataset

Add your own dataset under f-AnoGAN/your_own_dataset

Please replace your_own_dataset_dir_name/train_dir_name with a relative path of your own training data.

python train_wgangp.py "your_own_dataset_dir_name/train_dir_name"

Please replace your_own_dataset_dir_name/train_dir_name with a relative path of your own training data.

python train_encoder_izif.py "your_own_dataset_dir_name/train_dir_name"

Please replace your_own_dataset_dir_name/test_dir_name with a relative path of your own test data.

python test_anomaly_detection.py "your_own_dataset_dir_name/test_dir_name"

After Step: 3, score.csv will be generated in the directory results.

Please replace your_own_dataset_dir_name/test_dir_name with a relative path of your own test data.

python save_compared_images.py "your_own_dataset_dir_name/test_dir_name" --n_iters 0 --n_grid_lines 10

Compared images are saved under f-AnoGAN/your_own_dataset/results/images_diff.

f-AnoGAN_MNIST.ipynb
f-AnoGAN_MVTecAD.ipynb