search

rosinality / style-based-gan-pytorch

Implementation A Style-Based Generator Architecture for Generative Adversarial Networks in PyTorch
Other
1.1k stars 232 forks source link

num_workers=1 in DataLoader #108

Open YoelShoshan opened 4 years ago

YoelShoshan commented 4 years ago

Wondering - is there any reason that num_workers is hardcoded to 1, and there's no usage of more workers?

rosinality commented 4 years ago

It is due to this: https://github.com/rosinality/style-based-gan-pytorch/pull/48/commits/211cb45d5aeac1da1e2f29702f2f89cc94ab6c7c I remember using 1 workers is more efficient for low resolutions.

YoelShoshan commented 4 years ago

I see - makes sense since the multi-process communication becomes heavier than the inter process communication overhead, due to the extremely small work that a single worker needs to do in small resolutions. A possible solution is to reinitialize DataLoader with bigger workers_num as the resolution increases but I assume it's not high priority.

rosinality commented 4 years ago

Yes. And using distributed training will be even better.