Add WGAN-GP example

[jerry73204]

I'd like to present the Wasserstein GAN with gradient penalty example. It is slightly modified from the gan example. It removes batch norms in discriminator as the paper suggested, and replace them with instance norms.

I tested against the MNIST images. It works fine on single channel (grayscale) image. https://imgur.com/k5nQryZ

However, I found the image is bulrring if three-channel RGB images are provided. https://imgur.com/GuCVnEy

The code is almost ready. If anyone can tell if it is an implementation or fine-tuning issue, I could fix it soon and merge.

[jerry73204]

I tested WGAN-GP and it also generated blurry images. I found sources that experienced with this issue (1, 2).

I see we already have a Realtivistic DCGAN example. WGAN is just a plus. I not sure about the plan. Perhaps we can

• Discard this PR
• Accept this PR but without RGB blurry example
• Accept this PR with RGB blurry example.

[LaurentMazare]

Closing old issue as no activity in a while.