A Catkin-friendly package for utilizing the C++ API of Tensorflow.
Get Tensorflow C++ API into ROS as easy as
find_package(catkin REQUIRED COMPONENTS
... your other packages ...
tensorflow_ros_cpp
)
See the usage example at [https://github.com/tradr-project/tensorflow_ros_test].
New: Support for Noetic and TensorFlow 2.x!
You can choose either one of the following options to install Tensorflow.
- As a pip (Python) package: The easiest way on Ubuntu 14.04 and 20.04, Just
pip install tensorflowand that's it. GPU version supported! Can be used on Ubuntu 16.04 and 18.04, too, but with important limitations. - Using
tensorflow_catkinpackage: Easily compile Tensorflow for your platform. A convenient way on newer systems. Supports GPU version. - Using a custom build of Tensorflow built by bazel: The least comfortable, yet most powerful way. Compile Tensorflow yourself using bazel and tell this package where to find it.
See below for more detail about each of the installation types and how to set them up.
If you're managing dependencies via rosdep, it is likely that you do not want it to try to install the optional dependencies (currently python-tensorflow-pip and tensorflow_catkin). In such case, add the following to the rosdep call:
rosdep install ... --skip-keys=tensorflow_catkin --skip-keys=python-tensorflow-pip
If you successfully used this package on an untested configuration (marked with ?), please, tell us.
| TF | CUDA | CUDNN | pip tensorflow | pip tensorflow-gpu | bazel (CPU) | bazel (GPU) | tensorflow_catkin (CPU) | tensorflow_catkin (GPU) |
|---|---|---|---|---|---|---|---|---|
| 0.12.1 | no | no | ✓ | ✓ | ? | ? | N/A | N/A |
| 1.0.0 | no | no | ✓ | ✓ | ? | ? | N/A | N/A |
| 1.1.0 | 8 | 5 | ✓ | ✓ | ? | ? | N/A | N/A |
| 1.2.0 | 8 | 5 | ✓ | ✓ | ? | ? | N/A | N/A |
| 1.3.0 | 8 | 6 | ✓ | ✓ | ? | ? | N/A | N/A |
| 1.4.0 | 8 | 6 | ✓ | N/A (wants CUDA 9) | ? | ? | N/A | N/A |
| 1.5.0 | 8 | 6 | ✓ | N/A (wants CUDA 9) | ? | ? | N/A | N/A |
| 1.6.0 | 8 | 6 | ✓ | N/A (wants CUDA 9) | ? | ? | N/A | N/A |
| 1.7.0 | 8 | 6 | ✓ | N/A (wants CUDA 9) | ✓ | ✓ | ✓ | ✓ |
| 1.8.0 | 8 | 6 | ✓ | N/A (wants CUDA 9) | ? | ? | N/A | N/A |
| 1.14.0 | 8 | 6 | ✓ | N/A (wants CUDA 10) | ? | ? | N/A | N/A |
| 1.15.0 | 8 | 6 | X (link error) | N/A (wants CUDA 10) | ? | ? | N/A | N/A |
Had to set TF_PYTHON_LIBRARY manually since CMake was only finding Python 2.7 libraries.
| TF | CUDA | pip3 tensorflow | pip3 tensorflow-gpu |
|---|---|---|---|
| 1.8.0 | 8 | ✓ | N/A (wants CUDA 9) |
| TF | CUDA | CUDNN | pip tensorflow | pip tensorflow-gpu | bazel (CPU) | bazel (GPU) | tensorflow_catkin (CPU) | tensorflow_catkin (GPU) |
|---|---|---|---|---|---|---|---|---|
| 0.12.1 | 8.0 | no | ✓, see ABI difference problems | ✓, see ABI difference problems | ? | ? | N/A | N/A |
| 1.0.0 | 8.0 | 5.0 | ✓, see ABI difference problems | ✓, see ABI difference problems | ? | ? | N/A | N/A |
| 1.1.0 | 8.0 | 5.0 | ✓, see ABI difference problems | ✓, see ABI difference problems | ? | ? | N/A | N/A |
| 1.2.0 | 8.0 | 5.0 | ✓, see ABI difference problems | ✓, see ABI difference problems | ? | ? | N/A | N/A |
| 1.3.0 | 8.0 | 6.0 | ✓, see ABI difference problems | ✓, see ABI difference problems | ? | ? | N/A | N/A |
| 1.4.0 | 8.0 | 6.0 | ✓, see ABI difference problems | ✓, see ABI difference problems | ? | ? | N/A | N/A |
| 1.5.0 | 9.0 | 7.0 | ✓, see ABI difference problems | ✓, see ABI difference problems | ? | ? | N/A | N/A |
| 1.6.0 | 9.0 | 7.0 | ✓, see ABI difference problems | ✓, see ABI difference problems | ? | ? | N/A | N/A |
| 1.7.0 | 9.0 | 7.0 | ✓, see ABI difference problems | ✓, see ABI difference problems | ✓ | ✓ | ✓ | ✓ |
| 1.8.0 | 9.0 | 7.0 | ✓, see ABI difference problems | ✓, see ABI difference problems | ✓ | ✓ | N/A | N/A |
| 1.14.0 | 9.0 | 7.0 | ✓, see ABI difference problems | N/A (wants CUDA 10) | ? | ? | N/A | N/A |
| 1.15.0 | 9.0 | 7.0 | ✓, see ABI difference problems | N/A (wants CUDA 10) | ? | ? | N/A | N/A |
Had to set TF_PYTHON_LIBRARY manually since CMake was only finding Python 2.7 libraries.
| TF | CUDA | CUDNN | pip3 tensorflow | pip3 tensorflow-gpu |
|---|---|---|---|---|
| 1.8.0 | 9.0 | 7.0 | ✓, see ABI difference problems | ✓, see ABI difference problems |
| TF | CUDA | CUDNN | pip tensorflow | pip tensorflow-gpu | bazel (CPU) | bazel (GPU) | tensorflow_catkin (CPU) | tensorflow_catkin (GPU) |
|---|---|---|---|---|---|---|---|---|
| 1.0.0 | 8.0 | 5.0 | ✓, see ABI difference problems | ? | ? | ? | N/A | N/A |
| 1.1.0 | 8.0 | 5.0 | ✓, see ABI difference problems | ? | ? | ? | N/A | N/A |
| 1.2.0 | 8.0 | 5.0 | ✓, see ABI difference problems | ? | ? | ? | N/A | N/A |
| 1.3.0 | 8.0 | 6.0 | ✓, see ABI difference problems | ? | ? | ? | N/A | N/A |
| 1.4.0 | 8.0 | 6.0 | ✓, see ABI difference problems | ? | ? | ? | N/A | N/A |
| 1.5.0 | 9.0 | 7.1 | ✓, see ABI difference problems | ✓, see ABI difference problems | ? | ? | N/A | N/A |
| 1.6.0 | 9.0 | 7.1 | ✓, see ABI difference problems | ✓, see ABI difference problems | ? | ? | N/A | N/A |
| 1.7.0 | 9.0 | 7.1 | ✓, see ABI difference problems | ✓, see ABI difference problems | ? | ? | ✓ | ✓ |
| 1.7.0 | 9.1 | 7.1 | N/A | N/A | ? | ? | ✓ | ✓ |
| 1.8.0 | 9.0 | 7.1 | ✓, see ABI difference problems | ✓, see ABI difference problems | ✓ | ✓ | N/A | N/A |
| 1.8.0 | 9.1 | 7.1 | N/A | N/A | ✓ | ✓ | N/A | N/A |
| 1.14.0 | 10.0 | 7.4 | ✓, see ABI difference problems | ✓, see ABI difference problems | ✓ | ? | ? | ? |
| 1.14.0 | 10.0 | 7.4 | ✓, see ABI difference problems | ? | ? | ? | ? | ? |
Had to set TF_PYTHON_LIBRARY manually since CMake was only finding Python 2.7 libraries.
| TF | CUDA | CUDNN | pip3 tensorflow | pip3 tensorflow-gpu |
|---|---|---|---|---|
| 1.8.0 | 9.0 | 7.0 | ✓, see ABI difference problems | ✓, see ABI difference problems |
| 1.14.0 | 10.0 | 7.4 | ✓, see ABI difference problems | ✓, see ABI difference problems |
TF | CUDA | CUDNN | pip3 tensorflow --- | --- | --- | --- | --- 2.9.1 | no | no | ✓ | ✓
| TF | CUDA | CUDNN | pip tensorflow | pip tensorflow-gpu | bazel (CPU) | bazel (GPU) | tensorflow_catkin (CPU) | tensorflow_catkin (GPU) |
|---|---|---|---|---|---|---|---|---|
| 1.3.0 | 8.0 | 6 | ✓ | ✓ | ? | ? | ? | ? |
| TF | CUDA | CUDNN | pip tensorflow | pip tensorflow-gpu | bazel (CPU) | bazel (GPU) | tensorflow_catkin (CPU) | tensorflow_catkin (GPU) |
|---|---|---|---|---|---|---|---|---|
| 1.4.0 | 8.0 | 6 | C++ ABI problems | C++ ABI problems | ✓ | ? | ? | ? |
Except for the standard catkin variables (tensorflow_ros_cpp_INCLUDE_DIRS, tensorflow_ros_cpp_LIBRARIES, tensorflow_ros_cpp_DEPENDS and tensorflow_ros_cpp_CATKIN_DEPENDS), the following variables can be used from packages that find_package this package:
tensorflow_ros_cpp_USES_CXX11_ABI(bool): Whether the used Tensorflow library is built using C++11 ABI or not.tensorflow_ros_cpp_TF_LIBRARY_VERSION(int): Whether the found library is TF 1 or 2.
These variables are read when configuring the package:
TF_LIBRARY_VERSION(int): Either 1 or 2 depending on whether you want TF 1.x or 2.x. This will not modify the search behavior, but it will check whatever is found, and if it does not satisfy the major version, another search method is tried.
Assumes tensorflow or tensorflow-gpu pip package is installed. Ideally via pip, but custom installs are also supported (if they're either on PYTHON_PACKAGE_PATH or if you manually specify environment variable TF_PIP_PATH).
Install either using rosdep, or by calling
sudo pip install tensorflow
# or
pip install --user tensorflow
# or in a virtualenv
pip install tensorflow
This package uses a "hack" to link against the library that's installed by pip. I have no idea if this is currently an officially supported way of accessing the C++ API, but it seems to work.
You can change these variables either in the CMake cache file, or from commandline passing e.g. -DTF_PIP_PATH=/my/path
FORCE_TF_PIP_SEARCH:BOOL(defaultOFF): Search for the pip-installed Tensorflow even if using a system with C++11 ABI (see below for explanation).DISABLE_TF_PIP_SEARCH:BOOL(defaultOFF): Do not search for pip-installed Tensorflow at all.TF_PYTHON_VERSION:STRING(default2.7): The python version used by the Tensorflow installation.TF_PYTHON_LIBRARY:STRING(default""): If nonempty, specifies path to libpythonx.y.so. Needed in some cases where CMake finds the wrong library, e.g. when using Python 3 pip.TF_PIP_EXECUTABLE:STRING(defaultpip${TF_PYTHON_VERSION}): Path to thepipexecutable that should be used (important to distinguishpip2for Python 2 andpip3for Python 3).TF_PIP_DISABLE_SEARCH_FOR_GPU_VERSION:BOOL(defaultOFF): Only regards thetensorflowpip package, even if the GPU version is installed.TF_PIP_PATH:STRING(default""): If the automated search process fails, manually specify path to the folder(site|dist)-packages/tensorflowcontaining e.g.python/pywrap_tensorflow.py.
Ubuntu systems starting with 16.04 (Xenial) are using the new C++11 ABI for all system libraries. Until pip bug 1707002 gets resolved (if ever!), the pip-distributed Tensorflow is built againts an older C++ ABI, which is incompatible with the C++11 ABI. This means linking to the Tensorflow library will fail on such systems and there's no way around it. This only affects TF 1.x, because TF 2.x is always built with C++11 ABI.
The only "workaround" is to wrap all tensorflow code into a library that exposes its functions using C API (so no std::strings, std::vectors or Eigen types) and which does not link to any system library with a C++ API.
Such library can then be built separately with -D_GLIBCXX_USE_CXX11_ABI=0.
Then you can freely link the rest of your code using ROS, Eigen and so on to this library.
An example of this approach can be found at kinetic-devel branch of tensorflow_ros_test.
A library created this way is "backwards compatible" with older systems, so it can be used everywhere.
However, separating all Tensorflow code can be a pain in the ***, so we only suggest it as a last resort option.
This is why searching for the pip-installed Tensorflow is disabled by default on systems using the new C++11 ABI (you can force it using -DFORCE_TF_PIP_SEARCH=ON).
Add package tensorflow_catkin to your Catkin workspace.
Read its readme to correctly set the required CMake variables (no setup needed for the CPU version).
Be prepared that the compilation will eat up a lot of RAM and take a long time. On my system with 8 concurrent build jobs, it ate around 20 GBs in the peak.
You can add BUILD_COMMAND make -j2 and INSTALL_COMMAND make -j2 install to its CMakeLists.txt/ExternalProject_Add to limit the number of concurrent build jobs, which also decreases memory requierements.
FORCE_TF_CATKIN_SEARCH:BOOL(defaultOFF): Search fortensorflow_catkineven if Tensorflow has already been found.DISABLE_TF_CATKIN_SEARCH:BOOL(defaultOFF): Do not search fortensorflow_catkinat all.
Follow the Installing TensorFlow from Sources guide up to "Configure the installation" (including), and build the C++ library with the following command:
bazel build --config=opt --define framework_shared_object=false tensorflow:libtensorflow_cc.so
You don't need to continue with the guide building or installing the pip package (but you might be interested, because a custom-built tensorflow can provide you with higher performance even in Python).
If you encounter memory problems during the build, you can limit the used resources using the --local_resources bazel option.
FORCE_TF_BAZEL_SEARCH:BOOL(defaultOFF): Search for bazel-compiled Tensorflow even if Tensorflow has already been found.DISABLE_TF_BAZEL_SEARCH:BOOL(defaultOFF): Do not search for bazel-compiled Tensorflow at all.TF_BAZEL_LIBRARY:STRING(default"${CATKIN_DEVEL_PREFIX}/../libtensorflow_cc.so"): Full path to the buildlibtensorflow_cc.solibrary. If you put a symlink in your Catkin workspace root, it will be picked up by the default value.TF_BAZEL_SRC_DIR:STRING(default"${CATKIN_DEVEL_PREFIX}/../tensorflow-include-base"): Path to the sources folder of Tensorflow (your clone of the Git repository). If you put a symlink namedtensorflow-include-basein your Catkin workspace root, it will be picked up by the default value.TF_BAZEL_USE_SYSTEM_PROTOBUF:BOOl(defaultOFF): Whether to use system-installed protobuf includes or those distributed with Tensorflow.
If you encounter crashes when running the TF-using application, or it errors out with Not found: No session factory registered for the given session options, try the change to tensorflow/BUILD suggested in #9 (comment), clean the bazel workspace, and build it again.
Tensorflow 1.15+ is not compatible with ROS Indigo. You will get linking errors when using the library. Please use TF 1.14 or older on Indigo.