Backward about improved-wgan-pytorch HOT 4 CLOSED

Hi Tim,
Can you share an example?

from improved-wgan-pytorch.

Yes, I mean, in your fils, let's see in congan_train.py, when training G, you use only one backward(), which is g_cost.backward(), and when training D, there is also only one (disc_cost + ACGAN_SCALE*disc_acgan).backward(). But I had seen many others' codes that do 3 backwards(), I modify my code based on their code and find that the GAN does not work, so I'm not sure is where the problem is. The problem link can be found here: https://stackoverflow.com/questions/69116199/wgan-loss-always-increase. If you have free time, you may check it, thx a lot! Note that this is WGAN and contain 2 backwards, if it is WGAN-GP, there should be one more GP backwards......

from improved-wgan-pytorch.

You mean here?
https://github.com/jalola/improved-wgan-pytorch/blob/master/train.py#L192

disc_cost = disc_fake - disc_real + gradient_penalty
disc_cost.backward()

This equals:

disc_fake.backward()
(-disc_real).backward()
gradient_penalty.backward()

from improved-wgan-pytorch.

Thx, it seems I still cannot find the problem... I think my network is all OK...

from improved-wgan-pytorch.