This CMake project is basically an install script for LibTorch (PyTorch C++ API). It downloads the binary archieve and installs headers, libraries and CMake configurations files to the corresponding INCLUDEDIR, LIBDIR and DATADIR directories.
The package can be used as part of a pure CMake catkin or colcon workspace by including:
  <buildtool_depend>cmake</buildtool_depend>
  <depend>onnx_runtime_ros</depend>
  <export>
    <build_type>cmake</build_type>
  </export>in your package.xml. The relevant files will then be installed within the workspace target folder (install or devel).
Usage in a CMake project:
find_package(Torch REQUIRED)
target_link_libraries(${PROJECT_NAME} PUBLIC torch)
set_property(TARGET ${PROJECT_NAME} PROPERTY CXX_STANDARD 14)
target_link_options(${PROJECT_NAME} PUBLIC ${TORCH_CXX_FLAGS})If you get the error No CMAKE_CUDA_COMPILER could be found., then the CUDA compiler nvcc cannot be found in the default search paths ($PATH). In this case, you have to set the path to nvcc manually:
export CUDACXX=/usr/local/cuda/bin/nvcc