Data loader crashes during training. Something to do with multiprocessing in docker

I had the same problem. I think it can sometimes be hardware related. It disappeared when I decreased the overclock of the CPU and the RAM. So maybe memory instability could explains some cases

For people who randomly end up getting this error, there is a chance for this error
[ConnectionRefusedError: [Errno 111] Connection refused] to occur when there is corrupted data (for my case, it was an image) that is being fed to the dataset/loader.

Worth to have a look.