A Method to Accelerate the Training of WGAN

Generative adversarial network draw much attention in recent years because of its impressive results. The idea of adversary was cleverly introduced to the training of the network. And the result was used in the value function to evaluate the quality of the network. During the adversarial process, the discriminator was trained to separate the training data from the fake data, and the generator learn to estimate the real data distribution from the feedback of the discriminator. Both sides of the network learn from the back-propagation and finally reach a state of balance. After the training, the generator can produce very visually appealing samples. In this paper, we review the main route of the GANs, and we find the weakness of the value function from the experiments and propose new optimal method to improve the training of the GAN.