Kim Seonghyeon

In terms of convergence speed, it is actually slower than previous version.

FID was like this: Iterations | StyleGAN 2 (old) | SWAGAN (prev) | SWAGAN (revised) -----------|-------------------:|------------------:|---------------------: 70k | 9.41 | 12.70 | 12.91 80k | 8.67 | 11.06 | 12.10...

@Otje89 I have updated the code at bef283a.

As ddp is depend on torch.distributed.launch, it is related on that module. I think you can try to use export CUDA_VISIBLED_DEVICES=4,5,6,7.

I think you may need to check your checkpoints is valid stylegan2 checkpoint.

@Quasimondo Oh, thank you for the test! I will check and update the readme.

@FlashlightET Did you set `--repo` arguments to the path that contains dnnlib directory?

@zhouyuezhou1 You need to clone official tensorflow repository and set --repo to the path of it.

The approach of official implementations is slightly different from this repository. I choosed to initialize bias 1, official implementation uses bias initialized 0 and use + 1 for applying biases....

Currently coverter for stylegan 1 was not implemented.