Index page

WGAN

[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
  • is a way of measuring the matching between two distributions
  • math D​KL​​(p∣∣q)​∑​​​p(x​i​​)⋅(log p(x​i​​)−log q(x​i​​))
• 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