fast-neural-style
fast-neural-style copied to clipboard
Training new model with OpenCL backend
Does it work? I've encountered following error and I'm not sure if it is fault of my configuration, fast-neural-style or cltorch
} (15): nn.SpatialFullConvolution(128 -> 64, 3x3, 2,2, 1,1, 1,1) (16): nn.InstanceNormalization (17): nn.ReLU (18): nn.SpatialFullConvolution(64 -> 32, 3x3, 2,2, 1,1, 1,1) (19): nn.InstanceNormalization (20): nn.ReLU (21): nn.SpatialConvolution(32 -> 3, 9x9, 1,1, 4,4) (22): nn.Tanh (23): nn.MulConstant (24): nn.TotalVariation } /home/lukasz/torch-cl/install/bin/luajit: ./fast_neural_style/GramMatrix.lua:44: Not implemented at /home/lukasz/torch-cl/opencl/cltorch/src/lib/THClTensorMathBlas.cpp:427 stack traceback: [C]: in function 'bmm' ./fast_neural_style/GramMatrix.lua:44: in function 'forward' ./fast_neural_style/StyleLoss.lua:40: in function 'updateOutput' .../lukasz/torch-cl/install/share/lua/5.1/nn/Sequential.lua:44: in function 'forward' ./fast_neural_style/PerceptualCriterion.lua:78: in function 'setStyleTarget' train.lua:134: in function 'main' train.lua:327: in main chunk [C]: in function 'dofile' ...z/torch-cl/install/lib/luarocks/rocks/trepl/scm-1/bin/th:145: in main chunk [C]: at 0x00405ea0
I experienced a similar error. I solved it by updating my CUDA installation (from 7.5 to 8.0), and then reinstalling cutorch and cunn. I also updated cuDNN. For some reason cuDNN v5.1 for CUDA 8.0 caused a similar error to show up, but when I downgraded to cuDNN v5.0 for CUDA 8.0, everything worked out.
Seems that cltorch has not implemented this function yet (and they are not really working on this, according to this timeline).
I am facing the same issue. @lukaszarczynski or Anyone figured out a way to get it working? @Runescaped I am running with openCL backend on mac book with Intel 4000 HD graphics card. So, I am guessing CUDA is not the problem, as I believe CUDA is related to Nvidia Graphics card. Please correct me if I am missing something.
When it was last discussed, there was some needed support missing in cltorch https://github.com/jcjohnson/fast-neural-style/issues/45
So unless this has changed, it simply cannot work.