I want to have a model with shared memory (same model parameter, no training) across all processes, so it is necessary to use
torch.multiprocessing.Process. And in addition, I want to have duplex communicator between master and all workers, can I combine PyTorch Process and python Pipe together ?
i.e. sending pipe connection heads as argument to PyTorch Process
from torch.multiprocessing import Process from multiprocessing import Pipe def worker(master_connection, worker_connection): do something def master(): master_connection, worker_connection = Pipe() process = Process(target=worker, args=[master_connection, worker_connection]) process.start() process.join()