Cost of generative model not reducing,

[aashish-kumar]

The cost of discriminate model converges fast to low value and then no perceivable change to generative model. Can you please suggest what the issue might be.
this.mp4.tar.gz

[yuluntian]

I'm having the same issue.

[MalteFlender]

I too have this issue. A cheap workaround is to change the Data Distribution to self.mu = -2 or something like this.

[akanimax]

@MalteFlender How is that even a workaround let alone cheap? How can one change the original data distribution in any scenario? It's like saying my generator is not creating MNIST samples and is instead creating random noise, so I'd just drop MNIST dataset and instead say that my original dataset itself is noise and so my generator works! 😮
Btw, has anyone investigated the issue?

[akanimax]

@aashish-kumar , @yuluntian , @MalteFlender , if you guys are still interested, I found the same diagram as in @aashish-kumar 's video in this paper -> "MANY PATHS TO EQUILIBRIUM: GANS DO NOT NEED TO DECREASE A DIVERGENCE AT EVERY STEP" [see Fig 2. (page no. 5)]. I am currently reading this paper. You guys can check it out if interested.

[akanimax]

Wasserstein GAN is the solution to the problem ... https://www.alexirpan.com/2017/02/22/wasserstein-gan.html