PyTorch uses shared memory to efficiently share tensors between its dataloader workers and its main process. However in a docker container the default size of the shared memory (a tmpfs file system mounted at /dev/shm) is 64MB, which is too small to ...
