Open FrancisYu2020 opened 3 years ago
@FrancisYu2020 Authors of StyleGAN2 have implemented a fused bias+activation operation (those under gan2shape/stylegan2/stylegan2-pytorch/op
), which is used in StyleGAN2 and is faster than using the original Pytorch's implementation. So compiling this part would make the training faster.
Hi, I am trying to familiarize myself with your code and could you please explain why there is an optional step that compiles StyleGAN2 to make the training faster?