Machine Learning Containers for Jetson and JetPack

Hosted on NVIDIA GPU Cloud (NGC) are the following Docker container images for machine learning on Jetson:

• l4t-ml
• l4t-pytorch
• l4t-tensorflow

Dockerfiles are also provided for the following containers, which can be built for JetPack 4.4 or newer:

• ROS Melodic (ros:melodic-ros-base-l4t-r32.4.4)
• ROS Noetic (ros:noetic-ros-base-l4t-r32.4.4)
• ROS2 Eloquent (ros:eloquent-ros-base-l4t-r32.4.4)
• ROS2 Foxy (ros:foxy-ros-base-l4t-r32.4.4)

Below are the instructions to build and test the containers using the included Dockerfiles.

Docker Default Runtime

To enable access to the CUDA compiler (nvcc) during docker build operations, add "default-runtime": "nvidia" to your /etc/docker/daemon.json configuration file before attempting to build the containers:

{
    "runtimes": {
        "nvidia": {
            "path": "nvidia-container-runtime",
            "runtimeArgs": []
        }
    },

    "default-runtime": "nvidia"
}

You will then want to restart the Docker service or reboot your system before proceeding.

Building the Containers

To rebuild the containers from a Jetson device running JetPack 4.4 or newer, first clone this repo:

$ git clone https://github.com/dusty-nv/jetson-containers
$ cd jetson-containers

ML Containers

To build the ML containers (l4t-pytorch, l4t-tensorflow, l4t-ml), use scripts/docker_build_ml.sh - along with an optional argument of which container(s) to build:

$ ./scripts/docker_build_ml.sh all        # build all: l4t-pytorch, l4t-tensorflow, and l4t-ml
$ ./scripts/docker_build_ml.sh pytorch    # build only l4t-pytorch
$ ./scripts/docker_build_ml.sh tensorflow # build only l4t-tensorflow

You have to build l4t-pytorch and l4t-tensorflow to build l4t-ml, because it uses those base containers in the multi-stage build.

Note that the TensorFlow and PyTorch pip wheel installers for aarch64 are automatically downloaded in the Dockerfiles from the Jetson Zoo.

ROS Containers

To build the ROS containers, use scripts/docker_build_ros.sh with the name of the ROS distro to build:

$ ./scripts/docker_build_ros.sh all       # build all: melodic, noetic, eloquent, foxy
$ ./scripts/docker_build_ros.sh melodic   # build only melodic
$ ./scripts/docker_build_ros.sh noetic    # build only noetic
$ ./scripts/docker_build_ros.sh eloquent  # build only eloquent
$ ./scripts/docker_build_ros.sh foxy      # build only foxy

Note that ROS Noetic and ROS2 Foxy are built from source for Ubuntu 18.04, while ROS Melodic and ROS2 Eloquent are installed from Debian packages into the containers.

Testing the Containers

To run a series of automated tests on the packages installed in the containers, run the following from your jetson-containers directory:

$ ./scripts/docker_test_ml.sh all        # test all: l4t-pytorch, l4t-tensorflow, and l4t-ml
$ ./scripts/docker_test_ml.sh pytorch    # test only l4t-pytorch
$ ./scripts/docker_test_ml.sh tensorflow # test only l4t-tensorflow

To test ROS:

$ ./scripts/docker_test_ros.sh all       # build all: melodic, noetic, eloquent, foxy
$ ./scripts/docker_test_ros.sh melodic   # build only melodic
$ ./scripts/docker_test_ros.sh noetic    # build only noetic
$ ./scripts/docker_test_ros.sh eloquent  # build only eloquent
$ ./scripts/docker_test_ros.sh foxy      # build only foxy

https://medium.com/nvidiajapan/implementing-robotics-applications-with-ros-2-and-ai-on-jetson-platform-2-20772bb9423c