Docker Guide

This guide covers how to use the Docker image for ACCV-Lab.

Setting up the Docker Image

There is a Dockerfile that can be used to set up a Docker image with all the dependencies needed to build and install ACCV-Lab, including all the contained namespace packages. The Dockerfile is located in the docker directory of the ACCV-Lab repository.

The Dockerfile uses the nvidia/cuda image (version: nvidia/cuda:12.4.1-devel-ubuntu22.04) as the base image.

The Dockerfile has the following optional build arguments:

USER_ID: The ID of the user to use for the container.
USER_NAME: The name of the user to use for the container.
GROUP_ID: The ID of the group to use for the container.
GROUP_NAME: The name of the group to use for the container.

If these are not set, the root user will be used.

The Dockerfile can be built as follows:

docker build -t accvlab .

Additional to the dockerfile, there is the run_docker_build.sh script (located in the docker directory) that can be used to build the container for the current user.

Running the Docker Image

The container can be run as follows:

docker run -it accvlab <...additional arguments...>

The following arguments are recommended to be used when running the container to ensure optimal performance: --gpus=all --cap-add SYS_NICE --privileged --shm-size=16g --ipc=host. Note that the optimal shared memory size (--shm-size) may depend on the specific use case.

To ensure that the the on-demand video decoder package can be used, the Nvidia container runtime needs to be used. The runtime can be installed by following the instructions here.

After installation, the runtime can be enabled by specifying the arguments: --runtime=nvidia -e NVIDIA_DRIVER_CAPABILITIES=all when running the container.

An example for running the container with this repository, some additional data, and a project making use of the ACCV-Lab packages could be:

docker run \
  --gpus=all --cap-add SYS_NICE --privileged --shm-size=16g --ipc=host \
  --runtime=nvidia -e NVIDIA_DRIVER_CAPABILITIES=all \
  -v /path/to/accvlab_repo:/accvlab \
  -v /path/to/your/data:/data \
  -v /path/to/your/project:/project \
  --rm -it \
  -u $(id -u):$(id -g) -w /project \
  accvlab:latest

If you want to run the container without the Nvidia container runtime, you can remove the --runtime=nvidia -e NVIDIA_DRIVER_CAPABILITIES=all arguments. However, note that this may prevent the on-demand video decoder package from working.

Note that in this example, the container is run with the current user $(id -u):$(id -g) and the working directory is set to /project.