How to Build TensorFlow for Windows

Environemt

• TensorFlow: ver. 1.10.0
  • https://github.com/tensorflow/tensorflow
  • https://github.com/tensorflow/tensorflow/archive/v1.10.0.tar.gz
• Python: Python 3.6.5 (Miniconda3 ver. 4.5.4 / 64bit)
  • https://conda.io/miniconda.html
  • https://github.com/tensorflow/tensorflow/archive/v1.10.0.tar.gz
• CMake: ver. 3.7.2
  • https://cmake.org/
• Compiler: Visual Studio 2017 (15.4.5)
  • https://visualstudio.microsoft.com/
  • Note: Be careful of compatibility with CUDA NVCC if you are building with GPU support; Latest VS may be incompatilble with CUDA NVCC and cause compilation error.
• SWIG: ver. 3.0.12
  • http://www.swig.org/
• git: Portable Git ver. 2.16.2
  • https://git-scm.com/
  • https://git-scm.com/download/win

• CUDA: ver 9.2.148

• cuDNN: 7.2.1

---

Setup

• Open Visual Studio command Prompt (2017 / Win64).

• Start “Anaconda Prompt 3” from the VS command prompt.

• Set path:

  set Path=<CMake folder>\bin;<SWIG folder>;<Portable git folder>\bin;%Path%
  

• Modify CMakeLists.txt (GPU version only):

  • Insert appropriate CUDA compute capability into arch=compute_XX,code="sm_XX,compute_XX" and -DTF_EXTRA_CUDA_CAPABILITIES=X.X

  • Go to https://developer.nvidia.com/cuda-gpus to look up Compute Capability.

  set(CUDA_NVCC_FLAGS ${CUDA_NVCC_FLAGS};-gencode arch=compute_61,code=\"sm_61,compute_61\";)
  
  ...
  
  if (WIN32)
    add_definitions(-DGOOGLE_CUDA=1 -DTF_EXTRA_CUDA_CAPABILITIES=6.1)
  else (WIN32)
  

• Go to output directory.

• Configuration:

  • GPU version:

    • Change “Visual Studio 15 2017 Win64” according to the compiler you are using.

  cmake ^
  <TF source>\tensorflow\contrib\cmake ^
  -G"Visual Studio 15 2017 Win64" ^
  -DCMAKE_BUILD_TYPE=Release ^
  -Dtensorflow_DISABLE_EIGEN_FORCEINLINE=ON ^
  -Dtensorflow_BUILD_CC_TESTS=OFF ^
  -Dtensorflow_BUILD_MORE_PYTHON_TESTS=OFF ^
  -Dtensorflow_ENABLE_GPU=OFF
  

  • CPU-only version

    • Change “tensorflow_CUDA_VERSION=9.2” according to your CUDA environment.
    • “CUDA_HOST_COMPILER” must be specified when using Visual Studio 2017; otherwise compilation will not complete.

  cmake ^
  <TF source>\tensorflow\contrib\cmake ^
  -G"Visual Studio 15 2017 Win64" ^
  -DCUDA_HOST_COMPILER="C:/Program Files (x86)/Microsoft Visual Studio/2017/Community/VC/Tools/MSVC/14.11.25503/bin/Hostx64/x64" ^
  -DCMAKE_BUILD_TYPE=Release ^
  -Dtensorflow_DISABLE_EIGEN_FORCEINLINE=ON ^
  -Dtensorflow_BUILD_CC_TESTS=OFF ^
  -Dtensorflow_BUILD_MORE_PYTHON_TESTS=OFF ^
  -Dtensorflow_ENABLE_GPU=ON ^
  -Dtensorflow_CUDA_VERSION=9.2
  

• Apply patch:

  • Reference:
    • https://github.com/tensorflow/tensorflow/issues/19198
    • http://eigen.tuxfamily.org/bz/attachment.cgi?id=834

• Build Python wheel:

  cmake --build . --config Release --target tf_python_build_pip_package
  

---

Install

• GPU version:

  pip install tensorflow_gpu-1.10.0-cp36-cp36m-win_amd64.whl
  

• CPU-only version:

  pip install tensorflow-1.10.0-cp36-cp36m-win_amd64.whl
  

---