nicholas-leonard/dp

Feature Request: OpenCL Support

cathalgarvey opened this issue · 3 comments

cathalgarvey commented

This code appears to have lots of explicit calls to :cuda(), which rules out using it with torch-cl as far as I'm currently aware?

As OpenCL is getting generally much better support in ML lately, and particularly in Caffe and Torch7, it'd be nice to see this key torch library getting OpenCL support, also.

I'm aware that there's an upstreaming angle here, too; it would be nice if Torch7 presented a common GPU interface that papered over the difference between cuda and opencl tensors, so library maintainers didn't have to handle both.

Thanks!

nicholas-leonard commented

I don't support dp anymore. Most of the other packages are adding support to opencl though.

cathalgarvey commented

That's a shame! Is there any active supported fork of dp out there, or are ther other packages that supercede dp's functionality?

nicholas-leonard commented

torch-net is another fully-fledged deep learning package built over torch/nn. But personally, I recommend just using dataload + nn (+ any other package you see fit).