GAN training

http://blog.otoro.net/2016/04/01/generating-large-images-from-latent-vectors/

Like a real counterfeit painter, G has a much tougher job compared to D. Creating content is hard, and even with the error gradients advising G how to defeat D, there’s no real guarantee that G can do a much better job at fooling D. Sometimes the gradient can just be stuck at zero, when D has gotten good enough at winning that any small change in G’s strategy will not beat D.

One of the most important points of training a GAN network is to not let the discriminator network become that much better than the generator, so a lot of thought should be placed on structuring the architecture, and size of both networks so that they are fair. For example, the learning rate for G’s weights are larger than the learning rate for D’s weights.

There are some more tricks to slowdown D’s training so G can have a good chance to always catch up to D. This includes running gradient descent learning on G, N times for every time gradient descent learning is run on D, during every batch. We have experimented with setting N between 4 and 8 times.

最后

以上就是漂亮毛衣最近收集整理的关于GAN training的全部内容，更多相关GAN内容请搜索靠谱客的其他文章。