Open ssxxx1a opened 3 years ago
Hi @ssxxx1a! The op extensions were taken from rosinality's stylegan2-pytorch implementation. It seems like rosinality updated the extensions since then (for example, I see cpu support in his new version).
While the current files are the ones I conducted my experiments on, I will look into updating the op extensions.
Thanks for letting me know about this! Omer
thanks for your reply! Due to memory issues, I tested model parallelism, which means that the encoder and decoder are placed on different GPUs. Then I found that there have somes errors (data conflicts) in this expansion, but if the op expansion here is replaced with rosinality version, there will be no problems. i guess is that the communication between different GPUs may involve the cpu? So rosinality version can resolve this conflict.
I found that the op extension you use here is different from the original one. Is there any difference here? What is the purpose?