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: [email protected]

Overview

This repository provides TensorFlow libraries with the following specifications:

• Provided versions: 1.15.2 (Default) and 1.13.2
• Supported for Ubuntu 18.04 LTS.
• Built with GCC>=7.5.
• Built with support for C++14.
• 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 the TensorFlow source files.

NOTE: As each pre-built distribution of TensorFlow is quite large (~1GB), the tensorflow/CMakeLists.txt CMake script will automatically download and unpack the archive the first time the package is built.

A complete CMake example example is provided for demonstrating how to write dependent packages.

Moreover, we provide additional scripts and tooling for:

• Downloading, patching and installing Eigen.
• Building tensorflow from source and extracting all library binaries and headers.

Install

First clone this repository:

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

or if using SSH:

git clone [email protected]:leggedrobotics/tensorflow-cpp.git

Eigen

Each distribution of tensorflow>=r1.13 requires a special patched version of the Eigen header-only library. As of v0.2.0 of this repository, the aforementioned patched header files of Eigen are already included in the the headers downloaded by tensorflow/CMakeLists.txt. However, in certain cases, code in some package A using tensorflow-cpp might interface with some other code in an external package B that also uses Eigen. Thus, in order to ensure that A and B work together properly, we must build both packages using the same version of Eigen.

For such cases, we provide an bash script in tensorflow-cpp/eigen/install.sh.

To download, unpack and patch Eigen:

cd tensorflow-cpp/eigen
install.sh

To additionally build and install Eigen, the --run-cmake argument can be used:

cd tensorflow-cpp/eigen
install.sh --run-cmake

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.

Please refer to our complete example for details.

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:

1. Follow the standard instructions for installing system dependencies.
   NOTE: For GPU-enabled systems, additional steps need to be taken.
2. 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.
3. 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 ..

Issues

If experiencing any issues please first take a look at our ISSUES.md file. If you are experiencing something we have not accounted for please create a new repository issue.

License

Apache License 2.0