Closed HendrikSchmidt closed 1 year ago
As we're rewriting the dataloader anyways, I will not try to solve this issue directly, but see whether our new dataloader might already circumvent it.
Either a very similar or exactly the same stack trace on my Windows machine. Perhaps we might be able to solve this and make the benchmark multi-platform. But I have not looked into it.
With the new updates to the environment, CUDA backend can be used on windows. MPS on Mac works for some models, but there still are some issues with the implementation in pytorch, which lead to failures in the LSTM model.
The transformer works on CUDA, but not on MPS so far.
Right now, when trying to run the training of the DL models, the multiprocessing throws an error on my machine (MacBook Pro with M1 Pro).
The issue might be the dataloader / way the H5 file is opened, a possible solution is described here: https://github.com/pytorch/pytorch/issues/11929#issuecomment-649760983. Ideally, the training would work on all different architectures and not only Linux to facilitate development speed.