TensorFlow CMake

This repository provides pre-built TensorFlow for C/C++ (headers + libraries) and CMake.

Maintainer: Vassilios Tsounis
Affiliation: Robotic Systems Lab, ETH Zurich
Contact: tsounisv@ethz.ch

Overview

This repository provides TensorFlow libraries with the following specifications:

Provided versions: 1.13.2 (Default)
Supports Ubuntu 18.04 LTS (GCC >=7.4).
Provides variants for CPU-only and Nvidia GPU respectively.
All variants are built with full CPU optimizations available for amd64 architectures.
GPU variants are built to support compute capabilities: 5.0, 6.1, 7.0, 7.2, 7.5

NOTE: This repository does not include or bundle the source TensorFlow repository.

Install

First clone this repository:

git clone https://github.com/leggedrobotics/tensorflow-cpp.git

or if using SSH:

git clone git@github.com:leggedrobotics/tensorflow-cpp.git

Eigen

To install the special version of Eigen requried by TensorFlow that we also bundle in this repository:

cd tensorflow/eigen
mkdir build && cd build
cmake -DCMAKE_INSTALL_PREFIX=~/.local -DCMAKE_BUILD_TYPE=Release ..
make install -j

NOTE: We recommend installing to ~/.local in order to prevent conflicts with other version of Eigen which may be installed via apt. Eigen exports its package during the build step, so CMake will default to finding the one we just installed unless a HINT is used or CMAKE_PREFIX_PATH is set to another location.

TensorFlow

These are the options for using the TensorFlow CMake package:

Option 1 (Recommended): Installing into the (local) file system

cd tensorflow/tensorflow
mkdir build && cd build
cmake -DCMAKE_INSTALL_PREFIX=~/.local -DCMAKE_BUILD_TYPE=Release ..
make install -j

NOTE: The CMake will download the pre-built headers and binaries at build time and should only happen on the first run.

Option 2 (Advanced): Create symbolic link to your target workspace directory:

ln -s /<SOURCE-PATH>/tensorflow/tensorflow <TARGET-PATH>/

For example, when including as part of larger CMake build or in a Catkin workspace

ln -s ~/git/tensorflow/tensorflow ~/catkin_ws/src/

Use

TensorFlow CMake can be included in other projects either using the find_package command:

...
find_package(TensorFlow CONFIG REQUIRED)
...

or alternatively included directly into other projects using the add_subdirectory command

...
add_subdirectory(/<SOURCE-PATH>/tensorflow/tensorflow)
...

NOTE: By default the CMake package will select the CPU-only variant of a given library version and defining/setting the TF_USE_GPU option variable reverts to the GPU-enabled variant.

User targets such as executables and libraries can now include the TensorFlow::TensorFlow CMake target using the target_link_libraries command.

add_executable(tf_hello src/main.cpp)
target_link_libraries(tf_hello PUBLIC TensorFlow::TensorFlow)
target_compile_features(tf_hello PRIVATE cxx_std_14)

NOTE: For more information on using CMake targets please refer to this excellent article.

A complete example is included in this repository to provide boilerplate CMake for developers of dependent projects and packages.

Customize

If a specialized build of TensorFlow (e.g. different verion of CUDA, NVIDIA Compute Capability, AVX etc) is required, then the following steps can be taken:

Follow the standard instructions for installing system dependencies.
NOTE: For GPU-enabled systems, additional steps need to be taken.
View and/or modify our utility script for step-by-step instructions for building, extracting and packaging all headers and libraries generated by Bazel from building TensorFlow.
Set the TENSORFLOW_ROOT variable with the name of the resulting directory:

cmake -DTENSORFLOW_ROOT=~/.tensorflow/lib -DCMAKE_INSTALL_PREFIX=~/.local -DCMAKE_BUILD_TYPE=Release ..

License

Apache License 2.0

whidbey/tensorflow-cpp