confused with the train process

[lzwhard]

Dear samet,

I think your ganomaly has no adversarial train.

you trained the Discriminator with |f(x) - f(x')|, (not the real or fake).

so I doubt that your D can not judge input is real or fake.

and your Generator using the real/fake (D have no idea about it) to adversarial the the autoencoder.

I have read your skip-ganomaly paper, there is the actual adversarial train.

[gdjmck]

I have the same confusion as well, it seems that the classifier of the NetD are never modified and just stay as what it was initialized?

[lzzlxxlsz]

hello,have you run the code successfully ？

[samet-akcay]

Hi @lzwhard and @gdjmck,
As you pointed out, the previous version had some issues with the discriminator training. The updated code trains the discriminator in an adversarial setting. The feature matching loss |f(x) - f(x')| from discriminator is added to the generator training, which overall improves the performance of the model. Please check out the latest version, and let me know if you have any concerns.

[samet-akcay]

I'm closing this issue due to inactivity. If needed, the discussion could be continued in #28.