Add "Progressive Growing of GANs" (ProGAN) model

[seovchinnikov]

I've decided to add a new model to the framework - an implementation of the paper "Progressive Growing of GANs": https://arxiv.org/abs/1710.10196.
It's basically a GAN that generates images from a random noise with a support of high resolution images by training G and D nets progressively. I've decided to post a PR because found it easily to implement it inside this repo's framework so it fits it very well (despite the name of repo :)

Model requires dataset of type dataset_mode='single', generator netG='progan', discriminator netD='progan'.
Please note that opt.crop_size (default 256) == 4 * 2 ** opt.max_steps (default max_steps is 6).
ngf and ndf controlls dimensions of the backbone (128-512).
Network G is a master-generator (accumulates weights for eval) and network C (stands for current) is a
current trainable generator.

See also:
    https://github.com/tkarras/progressive_growing_of_gans
    https://github.com/odegeasslbc/Progressive-GAN-pytorch

Please note that PR contains changes from PR #1090 with NVIDIA apex because larger batch sizes are prefered.

[junyanz]

Thanks for your contribution. Have you been able to reproduce the FID number on the ProgGAN datasets?

[seovchinnikov]

Sorry for a late response, I've only tested it on our private dataset, need help to evaluate it on Celeba

[aashana94]

I am trying to run metrics using the original implementation of Progressive Growing of GAN by tkarras but I am not able to, has anyone any luck? or can help me with it?

https://github.com/tkarras/progressive_growing_of_gans

[ridhachahed]

Hey ! You use 2 generators with one being the running average of the other and set adam parameter beta1 to 0. Why this choice ? Why not use the same generator for training and testing ?

[maxmax1992]

May I ask why are you in the ProGanModel doing generation from torch.randn and not from source Image A? It seems also that you don't even save the image A as input, this contradicts with the whole cycle consistency of the CycleGan as there's no loss between source image A and reconstruction of A.

[junyanz]

We will leave it as is. It might introduce too many changes to add a new GAN to the current repo. But if anyone is interested in progressive gans, you can refer to this PR.