amazon-science/bigdetection

why set workers_per_gpu=0

czczup opened this issue · 2 comments

czczup commented

Hi, thanks for your awesome work! I have a question about the dataloader of BigDetection, why set workers_per_gpu=0 in the dataloader? The current code requires 0.2-0.3s to load data in my experiment, I hope to use more workers to improve training efficiency, but set workers_per_gpu>0 seems to hang the code.

bryanyzhu commented

Thank you, the hanging issue could be caused by multiple reasons, which you might find solutions in MMdet issues. In general, it is likely because small shared memory and multi-threading. I suggest to install the latest MMdet to see if the issue persists.

czczup commented

Thank you, the hanging issue could be caused by multiple reasons, which you might find solutions in MMdet issues. In general, it is likely because small shared memory and multi-threading. I suggest to install the latest MMdet to see if the issue persists.

I solved it. Thank you for your reply.

  • 👍1