search

mit-han-lab / gan-compression

[CVPR 2020] GAN Compression: Efficient Architectures for Interactive Conditional GANs
Other
1.1k stars 150 forks source link

distilling on higer resolution #105

Closed youjin-c closed 2 years ago

youjin-c commented 2 years ago

Hello, I am training a CycleGAN with a repo based on this paper. I found distilling on 256x256 works fine and very well but not on 512x512. This is my thread about it. I found this repo very recently, so was thinking of investigating the paper and repo, but posting this question first to get any hints for distilling. Are there any ways to prevent collapsing with fine-tuning? or should I change the student structure?

lmxyy commented 2 years ago

What student architecture did you use?

youjin-c commented 2 years ago

Hello @lmxyy, Thanks for checking my question. They are using an inception distiller they implemented, which is not part of your code. I think I would go back to the author of the distiller or test with your repo separately. Thanks again for checking this, and I will get back when I have more firm questions.