Missing Classification loss in Discriminator for fake Images

[arshagarwal]

Missing Classification loss in Discriminator for fake Images

[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.