** * Windows builds are officially available from 0.12! ** **
Release 0.12.0/Major Features and Improvements
Recently (probably from around 2016/10), TensorFlow can be built on Windows. The procedure for building using CMake is published below (however, it seems that Migrate to bazel). Masu).
tensorflow/tensorflow/contrib/cmake/README.md
It's written fairly carefully, so basically you can build it, generate a Python wrapper, and use Python x Tensorflow natively on Windows. However, it takes a lot of time to build, so if you want to use it quickly, the one I built is listed below, so please use it. We have confirmed that the code in Introduction works.
icoxfog417/tensorflow-windows-build
In addition, build on Windows is still handled as an alpha version as of November 2016, and it seems that some functions are not supported. Therefore, if you want to use it on Windows for normal use, you should install it using bash on Windows (although there is an advantage that you can use GPU if you can build Windows natively).
Below is a description of the environment I built. If you want to try it yourself, please refer to it.
With miniconda / Anaconda, you can easily create an environment for building.
conda create -n tensorflow-win swig numpy zlib
Actually, cmake is also included in conda install, but the version of cmake entered here is 3.3.1, and a build error occurred in this case. Therefore, for cmake, I dropped the latest version of 3.6 from here and used it. The build script looks like this:
cmake .. -A x64 -DCMAKE_BUILD_TYPE=Release ^
-DSWIG_EXECUTABLE=<your_conda_env_path>\Library\bin\swig.exe ^
-DPYTHON_EXECUTABLE=<your_conda_env_path>\python.exe ^
-DPYTHON_LIBRARIES=<your_conda_env_path>\libs\python35.lib
More? -Dtensorflow_ENABLE_GPU=ON ^
More? -DCUDNN_HOME="<cudnn extraction directory>\cuda"
The following resources are required for building.
2G (quites eating)1 ~ 1.5 time takes to build (8G memory / Core i7)Please try TensorFlow life on Windows.
Recommended Posts