search

LMescheder / GAN_stability

Code for paper "Which Training Methods for GANs do actually Converge? (ICML 2018)"
MIT License
918 stars 114 forks source link

Weight average #5

Open Blade6570 opened 5 years ago

Blade6570 commented 5 years ago

Hi, it is really a great method to stabilize the GANs. I am just curious to know that how important the weight average is for stabilizing the GANs (in this case). You have mentioned in the paper that it increases the performance slightly. is it a viable option to remove it completely? I can check it experimentally but just looking for your insights.

LMescheder commented 5 years ago

The weight average reduces the noise but it also works without it. That said, for more varied data sets like Imagenet, the noise can be quite big. You can alternatively anneal the learning rate (we did that in an earlier version of this paper), but it's a bit annoying to find a good schedule and we obtained better results with the moving average.