Wasserstein Generative Adversarial NetworksHas several advantages:

• Solve the problem of unstable GAN training, without having to carefully balance the training degree of the generator and the discriminator;
• Basically solve the Collapse Mode problem and ensure the diversity of generated samples;
• During the training process, there are values such as cross entropy and accuracy that indicate the progress of the training. The smaller the value, the better the GAN training is, which indicates that the image quality produced by the generator is higher.
• No elaborate network architecture is required, only the simplest fully connected network is needed.

Compared with GAN, Wasserstein GAN has the following differences:

• The last layer of the discriminator cancels Sigmoid;
• The Loss of the generator and discriminator is not Log;
• Each time the discriminator parameters are updated, their absolute values are truncated to be lower than a fixed constant c.

Related words: GAN