ll_categorical_objective tends to saturate very quickly

JonathanRaiman / tensorflow-infogan

:dolls: InfoGAN: Interpretable Representation Learning

154 stars 63 forks source link

ll_categorical_objective tends to saturate very quickly #1

Closed coventry closed 5 years ago

coventry commented 7 years ago

Any idea why this might be happening? It doesn't with the openai version. For example:

evolution_of_categorical_objective

JonathanRaiman commented 7 years ago

I'm not sure. I imagine this has something to do with weighing the objectives differently, or perhaps reducing the expressive power of the discriminator a bit. Have you found a good explanation?

coventry commented 7 years ago

No, not yet.

JonathanRaiman commented 7 years ago

Ok. I would be curious to know if you find out :)