Run this without CUDA? about waifu2x

[nagadomi]

No.

[an-electric-sheep]

It may be possible to run this on a CPU by changing all instances of CudaTensor to FloatTensor, replacing cudnn module with nn and replacing various :cuda() method calls in the source.

With additional substitutions OpenCL might work for applying the model but not for training it.

But this will be extremely slow.

[nagadomi]

You are right. In officially, waifu2x supports only cuDNN. but,
You can run on CPU If replace cudnn.SpatialConvolution with nn.SpatialConvolutionMM in lib/srcnn.lua and use https://github.com/soumith/fakecuda .

[nagadomi]

and this code is cool.

waifu2x upscaling, done in 35 lines of Python: https://mrcn.st/t/waifu2x.py  
https://mrcn.st/t/scale2.0x_model.json … PyOpenCL version next? 

from https://twitter.com/marcan42/status/601456859556417536

[darkdragon-001]

from https://twitter.com/marcan42/status/601456859556417536

OpenCL implementation from the same guy: https://github.com/marcan/cl-waifu2x