Wasserstein GAN is another variant of GANs that solve an issue that can happen when training GANs, called mode collapse. Moreover, it aims to give a metric that indicates when the GAN has converged, in other words, a loss function where the value has a meaning.
Important changes are to remove log from loss and clip the discriminator weights.
Also, follow these steps:
- Train discriminator more than generator
- Clip the weights of discriminator
- Use RMSProp instead of Adam
- Use low learning rates (0.0005)
A disadvantage of WGANs is that they are slower to train :
The image results produced by WGAN are still not that great, but this model does manage to help solve the mode collapse issue.