loss_D about wgan

[jeejeelee]

Hi:
Thanks for sharing your impressive repo, I found the loss_D in wan.py is :
loss_D = -torch.mean(discriminator(real_imgs)) + torch.mean(discriminator(fake_imgs)) ,but the wgan's paper described the loss_D as follow:

So is it a bug or other reasons？

[jeejeelee]

I have understood the loss_D,but in my personal opinion,the loss_D should be wrote as follow:
loss_D = -torch.mean(discriminator(real_imgs)) + torch.mean(discriminator(fake_imgs)) -1
is it right ?

[eriklindernoren]

I believe that since they add the gradient to the weights during update they perform gradient ascent instead of descent, and because of this in order to do gradient descent you will have to multiply the gradient with minus one. Because of this the loss will be:

-1 * (torch.mean(discriminator(real_imgs)) - torch.mean(discriminator(fake_imgs))) = 
-torch.mean(discriminator(real_imgs)) + torch.mean(discriminator(fake_imgs))

[jeejeelee]

@eriklindernoren thank you