0
$\begingroup$

Would it be possible to update the loss of my generator using a loss function different from binary cross entropy?

I have a multi-class labeled data-set and I want to train my discriminator to learn how to classify these images, while my generator is trained to make fake images so that the discriminator can't predict the right class.

Excuse me if I ask something obvious, but i would appreciate any help!!!

$\endgroup$
0
$\begingroup$

An alternative is Wasserstein loss, which supposedly helps with mode collapse. See this paper for details

Are you training a separate GAN for each label? AFAIK the discriminator is usually a binary classifier. It's trying to predict whether an input is either real or generated.

I found an interesting post on semi-supervised GAN training, but I've never attempted it myself. Hope that is helpful.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ thank you so much for your response! No I dont have a separate GAN for each label :( Also I would like to get rid of the real fake classification part in general (which is included in semi supervised GANs)!! $\endgroup$ – kkk Mar 10 '19 at 20:03

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.