sarlinpe / tensorflow_catkin

Catkin wrapper for a Bazel-free build of Tensorflow

[09/05/2019] tensorflow_catkin is now moving to ethz-asl/tensorflow_catkin.
This repository will no longer be supported. A usage example is provided in ethz-asl/hierarchical_loc.

tensorflow_catkin

Catkin package wrapper for Tensorflow 1.8 C++

• Uses the CMake build and does not require Bazel
• Builds the latest commit of the master branch as of 24/06/2018
• Supports GPU using the -DUSE_GPU CMake flag
  • Assumes that the root directories of CUDA, CUDNN and NCCL are provided either as environment variables CUDA_ROOT, CUDNN_ROOT and NCCL_ROOT, or as flags -DCUDA_ROOT, -DCUDNN_ROOT, -DNCCL_ROOT
  • Supports CUDA >=8.0 and CUDNN >=6.0 (-DCUDA_VERSION and -DCUDNN_VERSION flags)
  • Supports linking against the run-time version of libcuda in $CUDA_ROOT/lib64/stubs/
• Can leverage several make jobs to speed up the build (-DNUM_MAKE_JOBS flag)
• Builds without SSE and AVX accelerations by default
  • Because of memory alignment issues in Eigen, it is sometimes necessary to build all the libraries of a project with the same flags
  • Set -DCMAKE_CXX_FLAGS="-march=native" to enable all the acceleration primitives
• Tested on Ubuntu 14 and 16 and on an NVIDIA Jetson TX2
• Requires CMake >=3.5

Note: some features are disabled, such as GRPC support, Python bindings, and JPEG- and PNG-related ops like image encoding and decoding (the latter interferes with OpenCV but can be enabled by deleting the relevant patch).

Usage

package.xml

<?xml version="1.0" encoding="UTF-8"?>
<package format="2">
  <name>test_tensorflow</name>
  <version>0.0.0</version>
  <description>test_tensorflow</description>
  <maintainer email="[e-mail]">Author</maintainer>
  <license>Apache 2.0</license>

  <buildtool_depend>catkin_simple</buildtool_depend>
  <buildtool_depend>catkin</buildtool_depend>
  <depend>tensorflow_catkin</depend>
</package>

CMakeLists.txt

cmake_minimum_required(VERSION 2.8)
project(test_tensorflow)

find_package(catkin_simple REQUIRED)
catkin_simple()

cs_add_executable(example example.cpp)

cs_install()
cs_export()

example.cpp

#include <tensorflow/core/public/session.h>
#include <tensorflow/core/platform/env.h>
#include <iostream>
using namespace std;
using namespace tensorflow;

int main()
{
    Session* session;
    Status status = NewSession(SessionOptions(), &session);
    if (!status.ok()) {
        cout << status.ToString() << endl;
        return 1;
    }
    cout << "Session successfully created." << endl;
}