Way to run operations on CPU, thusly RAM instead of GPU/VRAM?

Question

Way to run operations on CPU, thusly RAM instead of GPU/VRAM?

mxchinegod opened this issue 8 years ago · 5 comments

I'm limited by the Amazon instance. I can buy a server with a K80 and 12GB of VRAM but I'd also be willing to pay more for slower processing on the cloud if I can get the CPU working.

I'm going to look at the code after this but in the event that the answer is obvious to the community, I wanted to ask before I go digging.

Thanks so much! Hope everyone's results have been fun.

Answer 1 · 2016-07-18T17:35:08.000Z

Just do not enable CUDNN and the script will run on a CPU. But it will take you hours if not days to get a picture.

Answer 2 · 2016-07-18T17:43:59.000Z

The CPU implementation of the convolution operation is too slow for any practical use. You would have to implement something faster first.

Answer 3 · 2016-07-18T17:56:59.000Z

I believe you that it's slow and I appreciate both very quick responses!

Since our needs may be slightly different, by what magnitude is it impractical? We talking days or weeks for 1MP + resolution?

Thanks!

Answer 4 · 2016-07-18T22:18:41.000Z

fyi https://github.com/jcjohnson/neural-style/

Answer 5 · 2016-07-19T16:31:44.000Z

I've used it, the results even with gradient smoothing are not ideal whatsoever. I've done enormous amounts of tweaking and attempted to isolate what I like about it, it just isn't as good. I will perhaps come back to it anyways because it's more fleshed out.