Missing Classification loss in Discriminator for fake Images

[arshagarwal]

According to the following code (line 248-252 at solver.py), classification loss for fake images is not computed and hence backpropogated, is there any specific reason behind this or is it a possible bug?

# Compute loss with fake images.
            x_fake = self.G(x_real, c_trg)
            out_src, out_cls = self.D(x_fake.detach())
            d_loss_fake = torch.mean(out_src)

Any help is highly appreciated.

[segalon]

To my understanding it is computed at line 281 at solver.py. Classification loss for fake images is a part of the generator objective, not of the discriminator.

[arshagarwal]

@segalon Thanks for replying, I think I understand this now. In-fact, I even tried the model by adding classification loss for fake images for the discriminator, and found that the results deprecated. I reckon this happens because if we employ classification loss on fake images our ground truth label is often wrong, especially in the beginning when fake images have really poor quality.