Training strategy?

[lantaoyu]

Training strategy?

[LantaoYu]

Theoretically, in generative adversarial networks, the generator can perfectly fit the underlying data distribution only when the discriminator is optimal while we are updating the generator. Another way to see it is that GANs is a minimax game, as introduced in the Eq. 1 of the paper, but in implementation we kind of optimize the loss function in a (Min-FixedMax and Max-FixedMin) iteration, note that (Min-FixedMax and Max-FixedMin) are different from the original minimax game and to approximate the min-max training scheme to constantly provide meaningful training signal for the generator, each time we need to train the discriminator as fully as possible. However, there is always a tradeoff between the training time and the performance. In a nutshell, to guarantee the convergence, we normally need to train the discriminator more sufficiently.