Dataloader shared list across workers with ddp

[Pietro_Cicalese]

This was done in DP. See here:

Communicating with Dataloader workers

I looked everywhere and worked on this problem for a few days before solving it. Basically, data structures instantiated using the multiprocessing package do not behave reliably. For example, things break predictably when you nest shared memory structures within other nested shared memory structures. The ideal way to have asynchronous communication between PyTorch dataloader workers is to use process Queues, which shuttle active child process state information to the next active worker which the…