I want to transfer my data preprocession into GPU, and I have writen a series of code to do something to a torch.cuda.tensor before feeding it into network, the code was proved to be true. But when I actually use this way to train my model, the cuda runtime initialization error always occure. I can’t understand where the problem is.
The error massage is
RuntimeError: cuda runtime error (3) : initialization error at /home/ayu/pytorch/aten/src/THC/THCCachingAllocator.cpp:508
Before I fed data to model, the defined model has been transfered to GPU, I don’t know if this is the problem.
Thank you for your advice!
Are you using multiple workers in your DataLoader
?
Could you share some code of your Dataset
and the instantiation of your DataLoader
?
Your error might be related to some multi-processing / CUDA issue.
Yes I used multiple workers to get data for network. Now I choose to get data from DataLoader and then do some transform for a batch of data to bypass this problem.
Thank you for your reply.