LMescheder / GAN_stability

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

Spectral Normalization #3

Open oleksandrlazariev opened 5 years ago

oleksandrlazariev commented 5 years ago

I'm curious if you've tried the Spectral Normalization technic for your approach? At least at resolution 128^2 it didn't work for me

LMescheder commented 5 years ago

We haven't tried to combine spetral norm with our method, but depending on the spectral norm implementation your problems might be related to how our code handles trainable parameters. Here and here we set all parameters of the generator / discriminator to trainable, which might also affect the u and v vectors of the spectral norm (which shouldn't be trainable). You can use buffers instead, but then you have to make sure that the running average still works, which currently only operates on the parameters.

oleksandrlazariev commented 5 years ago

@LMescheder thanks for the quick reply. I'll give it a try accordingly to your insights

Johnson-yue commented 5 years ago

@oleksandrlazariev Hi, did you work fine with SN ??