Tensorflow C++ SDK Generator

Overview

It is quite strange that a project of status like tensorflow does not provide an offically generated development package.

Here is an attempt to generate a sort of SDK. I have made few assumptions and the directory structure of the generated SDK is based on the include folder that ships with the tensorflow python package.

Building tensorflow on your machine

This is well documented here - https://www.tensorflow.org/install/install_sources

Generate the SDK

You need to provide two arguments to the script (gen_sdk.sh)

• The path to the tensorflow github repository that you cloned
• The path where you want to generate the SDK

Here is an example command -

./gen_sdk.sh -r ~/Desktop/Dev/3rdparty-compiled/tensorflow -o example/third-party/tensorflow-sdk

The tensorflow-sdk contains 3 main items -:

• include directory that contains the header files
• libs directory that contains libtensorflow_cc.so & libtensorflow_framework.so shared libraries
• CMakeLists.txt file that exposes an INTERFACE only CMake module with an alias third-party::tensorflow-cc

Example usage

The example folder contains a project structure that I generally use. Notice that the above example generation command created the SDK in the example/third-party folder.

This example project uses CMake as a build system.

Here are the steps to compile it -:

# traditional cmake based build
cd example
mkdir build
cd build
cmake ..
cmake --build .

Running the sample application -

Since the tensorflow shared libraries are not in the path you would have to either move them in /usr/local/lib or update your LD_LIBRARY_PATH environment variable.

# set LD_LIBRARY_PATH
export LD_LIBRARY_PATH=<full_path_to_repo>/example/third-party/tensorflow-sdk/libs

# run it
cd build
./example