Go to index
WGAN
Reading time : ~1 min
by Subhaditya Mukherjee
Paper notes for the paper
[12] WGAN
 Arjovsky, M., Chintala, S., & Bottou, L. (2017). Wasserstein gan. arXiv preprint arXiv:1701.07875.
Paper
 Infinite anime faces
 Dataset Link
Notes
 No requirement of maintaining balance between discriminator and generator training
 mode collapse is reduced
 Use EM distance instead of KL divergence
 alpha = .00005, c = .01, m = 64, ncrit = 5
 Distances:
 KL divergence
 Wasserstein Distance
 EM distance is continuous and differentiable a.e. means that
we can (and should) train the critic till optimality.
 The argument is simple, the
more we train the critic, the more reliable gradient of the Wasserstein we get, which
is actually useful by the fact that Wasserstein is differentiable almost everywhere.
 improved stability of the optimization process
Examples
Architecture
Discriminator

Generator

Network
