Need help with CPU only build for windows

[santanpereira]

Hi, I need help with CPU only build. Currently i am running "https://github.com/yajiedesign/mxnet/blob/master/setup-utils/install-mxnet-windows-python.bat" file to create build which creates the "libmxnet.dll" in the release folder. I am doing this on CPU machine without CUDA, CUDNN installation on that machine. (Here i have added operators for Deformable-ConvNets to work with CPU)

To test the mxnet build I run the demo from "https://github.com/msracver/Deformable-ConvNets/blob/master/rfcn/demo.py" (modified to use CPU). On GPU machine it worked fine for me and detected all objects with good accuracy. But when i run the demo on CPU machine it gives me "CUDA driver version is insufficient for CUDA runtime version" warning and detects only limited objects.

To create CPU only build is there any other code changes or flag which needs to be set ? output on GPU machine :

output on CPU machine :

console o/p on CPU machine : 'pretrained_epoch': 0}, 'output_path': './output/rfcn', 'symbol': 'resnet_v1_101_rfcn'} CUDA driver version is insufficient for CUDA runtime version CUDA driver version is insufficient for CUDA runtime version CUDA driver version is insufficient for CUDA runtime version CUDA driver version is insufficient for CUDA runtime version CUDA driver version is insufficient for CUDA runtime version CUDA driver version is insufficient for CUDA runtime version CUDA driver version is insufficient for CUDA runtime version testing COCO_test2015_000000000891.jpg 7.9530s

[yajiedesign]

two way. 1.use install-mxnet-windows-python.bat in not install cuda machine. 2.build with https://github.com/yajiedesign/mxnet/blob/master/docs/install/windows_setup.md but not checked USE_CUDA in cmake,

[santanpereira]

I have modified the bat file lines as below https://github.com/yajiedesign/mxnet/blob/0013ca14cb3d009f11dff20e4ba8efdc1293fa6f/setup-utils/install-mxnet-windows-python.bat#L246

REM SET CMAKE_OPT=%CMAKE_OPT% -DUSE_CUDA=%MXNET_SETUP_HAS_CUDA% REM SET CMAKE_OPT=%CMAKE_OPT% -DUSE_CUDNN=%MXNET_SETUP_HAS_CUDNN% REM if %MXNET_SETUP_HAS_CUDNN%=="1" SET CMAKE_OPT=%CMAKE_OPT% -DCUDNN_ROOT="%CUDNN_ROOT%" SET CMAKE_OPT=%CMAKE_OPT% -DBLAS="%MXNET_BLAS%" if "%MXNET_BLAS%"=="Open" SET CMAKE_OPT=%CMAKE_OPT% -DOpenBLAS_HOME="%MXNET_CONDA_LIBRARY%" -DOpenBLAS_INCLUDE_DIR="%OpenBLAS_INCLUDE_DIR%" -DOpenBLAS_LIB="%OpenBLAS_LIB%" REM if "%MXNET_BLAS%"=="MKL" SET CMAKE_OPT=%CMAKE_OPT% -DMKL_HOME="%INTEL_MKL_DIR%" -DMKL_INCLUDE_DIR="%MKL_INCLUDE_DIR%" -DMKL_LIB="%MKL_LIB%"

SET CMAKE_OPT=%CMAKE_OPT% -DOpenCV_LIB_PATH="%MXNET_CONDA_LIBRARY%\lib\" -DOpenCV_INCLUDE_DIRS="%MXNET_CONDA_LIBRARY%\include\" -DOpenCV_CONFIG_PATH="%MXNET_CONDA_LIBRARY%"

cmake -Wno-dev %CMAKE_OPT% -DUSE_CUDA=0 -DUSE_CUDNN=0 -DUSE_NVRTC=0 -DUSE_OPENCV=1 -DUSE_OPENMP=1 -DUSE_BLAS=open -DUSE_LAPACK=1 -DUSE_DIST_KVSTORE=0 -DCMAKE_PREFIX_PATH="%MXNET_CONDA_LIBRARY%" -G "Visual Studio 14 2015 Win64" -DUSE_PROFILER=1 -DCMAKE_BUILD_TYPE=Release -H. -Bbuild if errorlevel 1 goto :FAIL msbuild build\mxnet.sln /t:Clean,Build /p:Configuration=Release;Platform=x64 /m if errorlevel 1 goto :FAIL

Also, I am running the build on the machine where there is no cuda. Still i am getting warning "CUDA driver version is insufficient for CUDA runtime version".

[yajiedesign]

look like no problem.you try use cmake-gui direct.

[santanpereira]

Thank you, @yajiedesign. I figured out the problem. I had to modify the 'Deformable-ConvNets' repo at few places to make it work in CPU mode and got rid of the "CUDA driver version is insufficient for CUDA runtime version" warning.