chickensouple / tensorflow_cpp_wrapper

A C++ wrapper for the Tensorflow C API

A C++ wrapper for the Tensorflow C API

This will allow you to take a network trained in python and run it in a c++ program. This wraps the Tensorflow C API in a class (which is supposedly more stable?).

To use this, there are two steps:

1. Build a shared tensorflow library
2. Build static wrapper library that links to shared tensorflow library

To build tensorflow shared library:

1. clone tensorflow: https://github.com/tensorflow/tensorflow
2. checkout the version you want
3. cd into tensorflow/ root directory
4. run "./configure"
5. run "bazel build tensorflow:libtensorflow.so"

To build static wrapper library:

1. cd into tensorflow_cpp_wrapper/ root folder
2. set the "tensorflow_path" variable in the CMakeLists.txt to point to the root directory of tensorflow
3. mkdir build
4. cd build
5. cmake..
6. make

================================================

Now you can link against the static wrapper library to create c++ code that can load saved binary protobuf tensorflow models. (you may also have to link your c++ code with the shared tensorflow library)

This wrapper only works with tensorflow models that have constant variables (it won't restore a checkpoint). Use tf.graph_util.convert_variables_to_constants() and tf.train.write_graph() to save tensorflow models in python.

See test.cpp for example usage.

You can test if the wrapper is working by going to the main tensorflow_cpp_wrapper/ folder and

1. run xor_network.py to train a simple MLP that learns the xor function
2. run "build/test_exe"

================================================

some of this code is based off of: https://stackoverflow.com/questions/44305647/segmentation-fault-when-using-tf-sessionrun-to-run-tensorflow-graph-in-c-not-c