Can someone explain the rationale of this statement from GitHub - soumith/ganhacks: starter from "How to Train a GAN?" at NIPS2016
if loss of generator steadily decreases, then it’s fooling D with garbage (says martin)
2 Likes
I’m not sure if it’s exactly this observation, but in some of my experiments the generator learned to fool the discriminator by “garbage” images like completely blue images, which could be “sky images”, but were of course complete useless.
The generator thus exploited a weakness of the discriminator and fooled it until no learning was possible.
But as I said, it’s my interpretation of this statement. 
1 Like
I think what you pointed was what Martin wanted to mean in the Soumith’s GANHacks. In the case when generator learns to exploit a weakness of discriminator like generating a blue image (sky), what do you do to counteract this failure? Can you please share some tips and tricks around this issue?
I think I’m the wrong one to answer this question properly, since I’m just trying out different hyperparameters and hope for the best.
There are certainly other GAN artists in this discussion board with more useful advice. 
1 Like
Hi @smth! Can you shed some light on what is meant by “if loss of generator steadily decreases, then it’s fooling D with garbage”? And if possible can you please give us some workaround and techniques to tackle the issue?
1 Like