tSchlegl / f-AnoGAN

Code for reproducing f-AnoGAN training and anomaly scoring
MIT License
247 stars 64 forks source link

Mode colapse #9

Open ChristianEschen opened 5 years ago

ChristianEschen commented 5 years ago

Hi

I am trying to use your code on cifar10, in which I exclude as the abnormal class and use the rest of the class as normal classes.

After 7 epoch the result looks like this: samples_epoch7-3410

I continue training and after 26 epochs the results looks like this samples_epoch26-12665

The ground truth looks like this: samples_groundtruth

The discriminator loos looks like this: train_disc_cost

Finally, the generator loss looks like this train_gen_cost

As seen in the posted image, the model tends to mode collapse. How do you troubleshoot this model