pose-hg-train
pose-hg-train copied to clipboard
Add choice to use multi-gpu (data parallel)
The original code just occupy multiple GPUs but actually did use only one GPU. I modified model.lua and made it possble to use multi-GPU by introducing nn.DataParallelTable.
Thanks for the work on multi-gpu training, but I find that it doesn't reduce the training time, why is this happening?
I'm very sorry to report that 'nngraph' library is not compatible to multi-gpu mode, so the modification does not work. I would like you to close this pull request.