Opennmt-py: Anybody use multi-gpu for this systerm ?

[KelleyYin]

This systerm is very faster in only one gpu . However , we want to train the corpus of WMT which will cost too much times. Futhermore , we seek to spend less time for waiting experimental result .

[srush]

It's the next thing we plan on working on. We were waiting for some pytorch features.

[alphadl]

maybe you can utilize nn.DataParallel function in pytorch v0.3, usage like this torch.nn.DataParallel(net, device_ids=[0, 1, 2]).cuda()

[srush]

Yeah, we had that originally, but it didn't give speed improvements. We are coming back to it now.

[mcjoshi]

Multi GPU still not working

[vince62s]

if you're in a hurry, try this fork:
https://github.com/Ubiqus/OpenNMT-py

[sbmaruf]

@vince62s : thank you for the link.