Support for categorical latent code

[kaihuchen]

@marcbelmont
Thank you for sharing this code. I have been trying to use InfoGAN on datasets other than MNIST, but couldn't get it to work since the code is not yet generalized to handle non-MNIST input. Same thing with AC-GAN. Very happy to find out that your code readily work for other datasets.

Here is a question for you. My interest is mainly about InfoGAN's ability to disentangle the representation, as such I want to play with the latent_c in various ways, including injecting categorical latent code as described in the original InfoGAN paper. I see that your code currently supports only continuous latent codes, and I wonder if you have plan to support the categorical latent code at some point. Thanks!

[marcbelmont]

Categorical latent code is a next logical step. However, I do not have time to work on that at the moment. I'll be happy to merge any pull request with this change.