Closed TalonX1 closed 2 years ago
I try to close the multiprocessing dataloader,it works. But i still want to know why i happened and want to use multiprocessing.
Hi @TalonX1!
This is most likely a Windows-specific problem (with Python). It worked when I tried it the last time. I sometimes had strange errors when the dataloader was shutting down its workers after the training. I can't reproduce this error.
Are you using the code without any modifications? Windows had / has problems with pickling lambdas.
Please try again with the package versions specified in the p2s.yml. You might need to replace the ">=" with "==". The Python version (3.7) is the most important thing here.
It might be a disguised out-of-memory error caused by Windows' inefficient process creation. Please try again with only 1 worker to check.
You can read up on spawn vs fork processes on Windows. I'm now training in WSL 2 because it has a proper fork and avoids a lot of overhead, especially virtual memory of DLLs.
Hi @TalonX1!
This is most likely a Windows-specific problem (with Python). It worked when I tried it the last time. I sometimes had strange errors when the dataloader was shutting down its workers after the training. I can't reproduce this error.
Are you using the code without any modifications? Windows had / has problems with pickling lambdas.
Please try again with the package versions specified in the p2s.yml. You might need to replace the ">=" with "==". The Python version (3.7) is the most important thing here.
It might be a disguised out-of-memory error caused by Windows' inefficient process creation. Please try again with only 1 worker to check.
Thank you for your rapid reply,I will try like what you said,and close the issue temporarily.
When i try to run it,error happens.
pytorch 1.10.2 py3.8_cuda11.3_cudnn8_0 pytorch