GPU Load 5% avg about style-transfer HOT 3 OPEN

That is unfortunately due to the expensive CPU-to-GPU memory copies employed by the code in the master branch.

If you'd like a bit more speed (on par with the Torch implementations out there), try this version of Caffe: https://github.com/fzliu/caffe/tree/gram-layer, along with the gram-layer branch of this repository. That Caffe contains dpaiton's code for Gramian computation, which should speed things up significantly and better utilize the GPU.

from style-transfer.

Hi, ok I see. I'm wondering if everything shouldn't be done in CUDA or C++ AMP ? Python & lua are good for prototyping but if someone wants the algorithm to run really fast one has to port every piece of code to the GPU, no ?.

I'll try to compile the gram-layer Caffe and keep you up to date. Thank you for the code :) !

from style-transfer.

@fzliu I have tried what you suggest about gram layer, very huge speedup. Thank you!
But I find that when using python profiler to dig into it, Scipy optimizer costs quite a lot of times even though numpy now is compiled with multi-threads support.
Any suggestion about this performance issue?

Thank you.

from style-transfer.