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

The following ROS containers are also provided, which can be pulled from DockerHub or built for JetPack 4.4 or newer:

• ROS Melodic (ros:melodic-ros-base-l4t-r32.5.0)
• ROS Noetic (ros:noetic-ros-base-l4t-r32.5.0)
• ROS2 Eloquent (ros:eloquent-ros-base-l4t-r32.5.0)
• ROS2 Foxy (ros:foxy-ros-base-l4t-r32.5.0)
• ROS2 Galactic (ros:galactic-ros-base-l4t-r32.5.0)

Pre-built Container Images

The following images can be pulled from NGC or DockerHub without needing to build the containers yourself:

	L4T Version	Container Tag
l4t-ml	R32.5.0*	nvcr.io/nvidia/l4t-ml:r32.5.0-py3
	R32.4.4	nvcr.io/nvidia/l4t-ml:r32.4.4-py3
	R32.4.3	nvcr.io/nvidia/l4t-ml:r32.4.3-py3
l4t-pytorch	R32.5.0*	nvcr.io/nvidia/l4t-pytorch:r32.5.0-pth1.7-py3
	R32.5.0*	nvcr.io/nvidia/l4t-pytorch:r32.5.0-pth1.6-py3
	R32.4.4	nvcr.io/nvidia/l4t-pytorch:r32.4.4-pth1.6-py3
	R32.4.3	nvcr.io/nvidia/l4t-pytorch:r32.4.3-pth1.6-py3
l4t-tensorflow	R32.5.0*	nvcr.io/nvidia/l4t-tensorflow:r32.5.0-tf1.15-py3
	R32.5.0*	nvcr.io/nvidia/l4t-tensorflow:r32.5.0-tf2.3-py3
	R32.4.4	nvcr.io/nvidia/l4t-tensorflow:r32.4.4-tf1.15-py3
	R32.4.4	nvcr.io/nvidia/l4t-tensorflow:r32.4.4-tf2.3-py3
	R32.4.3	nvcr.io/nvidia/l4t-tensorflow:r32.4.3-tf1.15-py3
	R32.4.3	nvcr.io/nvidia/l4t-tensorflow:r32.4.3-tf2.2-py3
ROS Melodic	R32.5.0*	dustynv/ros:melodic-ros-base-l4t-r32.5.0
	R32.4.4	dustynv/ros:melodic-ros-base-l4t-r32.4.4
ROS Noetic	R32.5.0*	dustynv/ros:noetic-ros-base-l4t-r32.5.0
	R32.4.4	dustynv/ros:noetic-ros-base-l4t-r32.4.4
ROS2 Eloquent	R32.5.0*	dustynv/ros:eloquent-ros-base-l4t-r32.5.0
	R32.4.4	dustynv/ros:eloquent-ros-base-l4t-r32.4.4
ROS2 Foxy	R32.5.0*	dustynv/ros:foxy-ros-base-l4t-r32.5.0
	R32.4.4	dustynv/ros:foxy-ros-base-l4t-r32.4.4
ROS2 Galactic	R32.5.0*	dustynv/ros:galactic-ros-base-l4t-r32.5.0
	R32.4.4	dustynv/ros:galactic-ros-base-l4t-r32.4.4

note: the L4T R32.5.0 containers can be run on both JetPack 4.5 (L4T R32.5.0) and JetPack 4.5.1 (L4T R32.5.1)

To download and run one of these images, you can use the included run script from the repo:

$ scripts/docker_run.sh -c nvcr.io/nvidia/l4t-pytorch:r32.5.0-pth1.7-py3

For other configurations, 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

Before proceeding, make sure you have set your Docker Default Runtime to nvidia as shown above.

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       # test if the build of ROS all was successful: 'melodic', 'noetic', 'eloquent', 'foxy'
$ ./scripts/docker_test_ros.sh melodic   # test if the build of 'ROS melodic' was successful
$ ./scripts/docker_test_ros.sh noetic    # test if the build of 'ROS noetic' was successful
$ ./scripts/docker_test_ros.sh eloquent  # test if the build of 'ROS eloquent' was successful
$ ./scripts/docker_test_ros.sh foxy      # test if the build of 'ROS foxy' was successful