Docker ROS environment for easily using existing pre-trained PyTorch models

• CUDA 11.2
• PyTorch 1.8.1
• Torchvision 0.9.1

• Docker
• nvidia-docker2
• NVIDIA driver compatible with CUDA 11.2

make build-training

make build-ros

1. Put a dataset somehwhere.
2. Edit DATASET_DIR and LOG_DIR in Makefile.
   • DATASET_DIR : Location of the dataset
   • LOG_DIR : Location to put training log (tensorboard) and checkpoints (trained weights)
3. Run make train

1. Build ROS packages

   make catkin-build
   

2. Run ROS master

   docker-compose up master
   

3. Run usb_cam

   docker-compose up ros-usb-cam
   

• Run:

  docker-compose up ros-object-recognition
  

• Run:

  docker-compose up ros-object-detection
  

• Run:

  docker-compose up ros-semantic-segmentation
  

• Run:

  docker-compose up ros-depth-estimation
  

• Face detection / recognition