f-AnoGAN
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.
References
Papar
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.
Requirements
Python 3.6 or later
PyTorch 1.x
Matplotlib
Numpy
pandas
Pillow
scikit-learn
Usage for training and test the MNIST dataset
Please run below in order on the CPU.
Step: 0
python setup.py install
cd mnist
Step: 1
python train_wgangp.py --training_label 1 --seed 2 --n_epochs 20
Step: 2
python train_encoder_izif.py --training_label 1 --seed 2 --n_epochs 20
Step: 3
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.
Step: 4
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.
Visualization
