Open wjx2 opened 6 years ago
GAN convergence is a very difficult problem in practice. At the moment, just look at TensorBoard and stop training whenever you "feel" it enough. For this task, maybe 150k steps is enough. There are many articles online that explain briefly about GAN, e.g. https://medium.com/@jonathan_hui/gan-why-it-is-so-hard-to-train-generative-advisory-networks-819a86b3750b If you have time, take a look at GAN tutorial from Goodfellow (the author of GAN) to understand deeper: https://sites.google.com/view/cvpr2018tutorialongans/
Hi, thank you for your replay! Looking at the TensorBoard is an effective way to judge the convergence, and I found that saving images every few times to observe the generated images' quality is also a good way.
Hi . Thanks for the wonderful explanation. I am new in GAN. And I have a question. When I train cycleGAN, I found that the discriminator loss decreases and converges, but the generator loss always change not much. So, How to judge whether cycleGAN converges?