sent-conv-torch icon indicating copy to clipboard operation
sent-conv-torch copied to clipboard

Published 20 hours ago • verified publisher icon harvardnlp

Too slow

Open ox1ygen opened this issue 7 years ago • 1 comments

Hello! I am running this code on amazon p2.xlarge instance, and it is working much slower on GPU than on CPU. That is the most time-consuming place:

trainer.lua -- renormalize linear row weights local w = layers.linear.weight for j = 1, w:size(1) do renorm(w[j]) end

The version of Lua I use is 5.2. CUDA is 7.5, and cuDNN is 5.1 Do you have any suggestions?

ox1ygen avatar Apr 23 '17 14:04 ox1ygen

I had same issue as well, can someone tell me why?

clairett avatar Oct 20 '17 22:10 clairett