Cuda

[linuxreitt]

Is there anywhere pointing out Cuda/Cudnn Version requirements? Any chance of a workaround for using AMD GPU at these early stages?

[marwanmusa]

As you can see in conv2d_gradfix.py it will use backend torch in version 1.7 and 1.8

otherwise it'll use torch.nn.functional.conv2d()

[marwanmusa]

It is recommended to using conda environment. You can download torch and cuda version 1.8 with this command

# CUDA 11.1
conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge

hope this help

[WeiquanWa]

I also have noticed same thing as @marwanmusa. Under the current version of stylegan2 this project is using, pytorch 1.8 is most recommended.