PyTorch ROS Wrapper

This is a ros wrapper package for PyTorch.

Prerequisite

• Make sure ROS is all set.
• We ran the package on ROS NOETIC.

TO-DO-LIST

• Test cuda availability. Feel free to pull request, if you have done it!

Installation

• Git Clone. Run:

  mkdir ws && cd ws
  mkdir src && cd src
  git clone lala
  

• PyTorch. Please first go to here, and select Python to install PyTorch. You should be running some command like this:

  pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
  

  Then do:

  pip3 install --upgrade PyTorch
  

• libtorch. Go to here again, and select C++/Java to install libtorch. The website should give you a link. Make sure it is Download here (cxx11 ABI) not Download here (Pre-cxx11 ABI), as per issue here. Then, with the given link, do the following:

  cd ~/ws/src/torch_ros_wrapper/
  wget https://download.pytorch.org/libtorch/cpu/libtorch-cxx11-abi-shared-with-deps-2.1.1%2Bcpu.zip
  unzip libtorch-cxx11-abi-shared-with-deps-2.1.1+cpu.zip 
  cd ~/ws && catkin_make
  

  Please adjust the command according to your zip filename. And please make sure the file hierarchy is correct, as we defined the following in CMakeLists.txt:

  set(Torch_DIR "${CMAKE_CURRENT_SOURCE_DIR}/libtorch/share/cmake/Torch/")
  

Usage

This package provide 2 functions:

• Convert. You can convert your model from PyTorch trained .pt to libtorch (C++) accessible .pt by running the following (see remarks first):

  roslaunch torch_ros_wrapper convertTrainedPt.launch
  

  Remarks:

  • Put Trained .pt File. Place your .pt trained file under /launch/model/.
  • Put Trained .pt Filename. Please modify rosparam in convertTrainedPt.launch accordingly to your .pt filename.
  • Put Model Class Script. Under /scripts/models/, please place your model PyTorch script. Here, we put a tcn.py.
  • Modify Convert Script. In /scripts/ConvertCppPt.py, please first do import {your_model} at the top of the script. Then under function def load_custom_model(self):, please instantiate self.your_model.

  After running the above, you should be able to see model_cpp.pt under /launch/model/.

• RUN in ROS. We provide 2 launch files for running libtorch. 1 in rosnode, 1 in rosnodelet. Just run:

  roslaunch torch_ros_wrapper TorchRosNode.launch # or
  roslaunch torch_ros_wrapper TorchRosNodelet.launch
  

  The rosnode will try to load the model_cpp.pt that we just created, while rosnodelet does nothing. You can modify further.

Shoutout

Our wrapper is based on https://github.com/ActiveIntelligentSystemsLab/pytorch_enet_ros. Yet, the above provides more detailed description on how to setup your environment, and it is for a more general case where you can include any model.

Maintainer

pattylo @ AIRO-LAB @ RCUAS, HKPolyU