Tensorflow was built first and foremost as a Python API in a Unix-like environment. But there are some projects where using Windows and C++ is unavoidable. This post will show how to write a simple C++ program in Visual Studio 2015 that links to Tensorflow. This build was done on Windows 8 without GPU support.

Building Tensorflow in Visual Studio with CMake

The first step is to build Tensorflow into a static library that our program can eventually link to. Google generally builds their code using Bazel, but Bazel support on Windows is experimental. The Tensorflow group has also provided a Windows CMake build with fairly detailed instructions (which is also experimental). I found the CMake build easier to work with than the Bazel build. The instructions in this section outline the more detailed instructions provided by the Tensorflow team above, but with a few minor modifications I had to make to get the build to work.

Prerequisites

Make sure the following are installed:

• CMake
• Git
• SWIG

Additionally, make sure that the CMake and git executables are in your %PATH%. Then, prepare your environment by running

C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\bin\amd64\vcvars64.bat

Note that the Tensorflow group’s instructions call for running vcvarsall.bat instead, but this script doesn’t exist on my system. Running vcvars64.bat appears to have worked for me.

Download the Tensorflow source

Decide where you want the Tensorflow repository to live. Something like C:\Users\%USERNAME%\bin\ is good. cd to this directory and download the Tensorflow repository with

C:\...> git clone https://github.com/tensorflow/tensorflow.git

Prepare the Visual Studio project files with CMake

All the Tensorflow libraries that we build will be inside the cmake directory in the Tensorflow repository. We’ll make a separate build directory in the cmake directory like so:

C:\...> cd tensorflow\tensorflow\contrib\cmake
C:\...> mkdir build
C:\...> cd build

Now run CMake as follows, making sure to set the build type as Release (the Debug build is not currently supported). Only Python 3.5 is supported — Python 3.6 is not yet supported, nor is Python 2. If you run Python through Anaconda this will not be a problem. You may have to change the directories to match wherever the SWIG and Python executables and libraries are on your system. (Note that I’m using backticks to continue the command in PowerShell. If using the command prompt, end each line with ^. Or just type everything on one line.)

C:\...> cmake .. -A x64 -DCMAKE_BUILD_TYPE=Release `
>> -DSWIG_EXECUTABLE=C:\Users\%USERNAME%\bin\swig\swigwin-3.0.12\swig.exe `
>> -DPYTHON_EXECUTABLE=C:\Users\%USERNAME%\Anaconda3\python.exe `
>> -DPYTHON_LIBRARIES=C:\Users\%USERNAME%\Anaconda3\libs\python35.lib
>>

After a little while, CMake will have generated a number of Visual Studio project files.

Build Tensorflow in Visual Studio

Now in Visual Studio open the Tensorflow solution tensorflow.sln in tensorflow\tensorflow\contrib\cmake\build. The ALL_BUILD project should be highlighted as the startup project. Make sure that you have the 64-bit build and Release version set in the Configuration Manager. Then build ALL_BUILD. This should take some time (about two hours on my machine). Note that if (like me) you’re running Windows through Parallels, you may have to substantially increase the memory of the virtual machine (I needed to set the memory to more than 12 GB — the default was 1 GB). Otherwise you may find that Visual Studio throws a Fatal Error C1060 “compiler is out of heap space” after an hour and a half.

To check whether everything worked, there is a small program called tf_tutorials_example_trainer which iteratively finds the largest eigenvalue of a matrix:

C:\...> Release\tf_tutorials_example_trainer.exe

This should generate a bunch of lines that look something like this:

000000/000004 lambda = 2.000000 x = [0.894427 -0.447214] y = [1.788854 -0.894427]