I have an 8-GPU machine and successfully used DataParallel to train my network. At the end of the epoch I’m attempting to evaluate the model on a fairly large dev set, and I’d like to parallelize that operation. For some reason only one GPU is utilized during my prediction operations. Does setting model.eval() have some impact on how DataParallel works?
It looks like my issue is that the sample code I’m working with sets a new varialble mnetwork=DataParallel(network), and the training code operates on mnetwork but prediction code operates on network. Verifying the fix now.