libtensorflow_cc

We provide a pre-built library and a Docker image for easy installation and usage of the TensorFlow C++ API.

In order to, e.g., run TensorFlow models from C++ source code, one usually needs to build the C++ API in the form of the libtensorflow_cc.so library from source. There is no official release of the library and the build from source is only sparsely documented.

We try to remedy this current situation by providing two main components:

1. We provide the pre-built libtensorflow_cc.so including accompanying headers as a one-command-install deb-package. The package is available for both x86_64/amd64 and arm64 machines running Ubuntu. See Installation.
2. We provide a pre-built Docker image based on the official TensorFlow Docker image. Our Docker image has both TensorFlow Python and TensorFlow C++ installed. The Docker images support both x86_64/amd64 and arm64 architectures. The arm64 version is specifically targeted at NVIDIA Jetson Orin. See Docker Images.

[!IMPORTANT]
This repository is open-sourced and maintained by the Institute for Automotive Engineering (ika) at RWTH Aachen University.
Deep Learning is one of many research topics within our Vehicle Intelligence & Automated Driving domain.
If you would like to learn more about how we can support your deep learning or automated driving efforts, feel free to reach out to us!
     Timo Woopen - Manager Research Area Vehicle Intelligence & Automated Driving
     +49 241 80 23549
     [email protected]

If you want to use the TensorFlow C++ API to load, inspect, and run saved models and frozen graphs in C++, we suggest that you also check out our helper library tensorflow_cpp.

• Demo
• Installation
  • CMake Integration
• Docker Images
• Build
• Supported TensorFlow Versions
  • Version Matrix

Demo

Run the following from the Git repository root to mount, build and run the example application in the pre-built Docker container.

# libtensorflow_cc$
docker run --rm \
    --volume $(pwd)/example:/example \
    --workdir /example \
    rwthika/tensorflow-cc:latest \
        ./build-and-run.sh

# Hello from TensorFlow C++ 2.13.0!
#
# A = 
# 1 2
# 3 4
#
# x = 
# 1
# 2
#
# A * x = 
#  5
# 11

Installation

The pre-built libtensorflow_cc.so library and accompanying headers are packaged as a deb-package that can be installed as shown below. The deb-package can also be downloaded manually from the Releases page. We provide versions with and without GPU support for x86_64/amd64 and arm64 architectures.

GPU

wget https://github.com/ika-rwth-aachen/libtensorflow_cc/releases/download/v2.13.0/libtensorflow-cc_2.13.0-gpu_$(dpkg --print-architecture).deb
sudo dpkg -i libtensorflow-cc_2.13.0-gpu_$(dpkg --print-architecture).deb
ldconfig

CPU

wget https://github.com/ika-rwth-aachen/libtensorflow_cc/releases/download/v2.13.0/libtensorflow-cc_2.13.0_$(dpkg --print-architecture).deb
sudo dpkg -i libtensorflow-cc_2.13.0_$(dpkg --print-architecture).deb
ldconfig

CMake Integration

Use find_package() to locate and integrate the TensorFlow C++ API into your CMake project.

# CMakeLists.txt
find_package(TensorFlow REQUIRED)
# ...
add_executable(foo ...) # / add_library(foo ...)
# ...
target_include_directories(foo PRIVATE ${TensorFlow_INCLUDE_DIRS})
target_link_libraries(foo PRIVATE ${TensorFlow_LIBRARIES})

Docker Images

Instead of installing the TensorFlow C++ API using our deb-package, you can also run or build on top the pre-built Docker images in our Docker Hub repository. If supported, we offer CPU-only and GPU-supporting images. Starting with TensorFlow 2.9.2, we offer multi-arch images, supporting x86_64/amd64 and arm64 architectures.

The amd64 images are based on the official TensorFlow Docker images. The arm64 images are based on NVIDIA's official L4T TensorFlow images for Jetson, targeted at NVIDIA Jetson Orin. Our provided images only install the TensorFlow C++ API on top of those. The resulting images therefore enable you to run TensorFlow in both Python and C++.

Build

If you would like to build the deb-package and Docker images yourself, use the Makefile as instructed below.

All make targets support the flags TF_VERSION (defaults to 2.13.0), GPU (defaults to 1), and ARCH (defaults to host architecture) in order to build a specific TensorFlow version in CPU/GPU mode for a specific architecture. The flag BUILD_PIP_PACKAGE (defaults to 0) enables you to also build the pip-package alongside.

All make targets listed below also have a counterpart named <target>-all, which can be used to build multiple TensorFlow versions one after the other using the TF_VERSIONS flag like so:

make 0-download-official-dockerfiles-all TF_VERSIONS="2.9.0 2.8.0 2.7.0"

0. Download Dockerfiles from TensorFlow repository

This downloads the directory tensorflow/tools/dockerfiles/ from the TensorFlow repository.

make 0-download-official-dockerfiles

1. Build TensorFlow Development Image

This builds a TensorFlow development image tensorflow/tensorflow:X.Y.Z-devel[-gpu] based on the Dockerfile downloaded from the TensorFlow repository.

make 1-build-official-devel-image

2. Build TensorFlow C++ Image

Based on the development image, this builds the TensorFlow C++ library libtensorflow_cc.so and installs it in a runtime image rwthika/tensorflow-cc:X.Y.Z[-gpu], including both Python and C++ TensorFlow. In an intermediate Docker build stage, a deb-package libtensorflow-cc_X.Y.Z[-gpu].deb is created.

make 2-build-cpp-image

3. Export TensorFlow C++ Library Installation Package

This exports the deb-package libtensorflow-cc_X.Y.Z[-gpu].deb to the libtensorflow-cc output folder.

make 3-export-libtensorflow-cc

4. Test TensorFlow C++

This installs builds and runs the example application inside a container of the runtime image.

make 4-test-libtensorflow-cc

5. Print Versions of Build Tools

This prints the exact version numbers of all tools involved in the build process.

make 5-print-versions

Supported TensorFlow Versions

Version Matrix

Acknowledgements

This work is accomplished within the projects 6GEM (FKZ 16KISK038) and UNICARagil (FKZ 16EMO0284K). We acknowledge the financial support for the projects by the Federal Ministry of Education and Research of Germany (BMBF).

Notice

This repository is not endorsed by or otherwise affiliated with TensorFlow or Google. TensorFlow, the TensorFlow logo and any related marks are trademarks of Google Inc. TensorFlow is released under the Apache License 2.0.