homeganwassersteingradient_penalty

View code on Github

#

Gradient Penalty for Wasserstein GAN (WGAN-GP)

This is an implementation of Improved Training of Wasserstein GANs.

WGAN suggests clipping weights to enforce Lipschitz constraint on the discriminator network (critic). This and other weight constraints like L2 norm clipping, weight normalization, L1, L2 weight decay have problems:

1. Limiting the capacity of the discriminator 2. Exploding and vanishing gradients (without Batch Normalization).

The paper Improved Training of Wasserstein GANs proposal a better way to improve Lipschitz constraint, a gradient penalty.

labml.ai