Development Machine Setup (NVIDIA Docker)
This document contains valuable information on dockerized CARLA / TensorFlow setups. All instructions are tested for Ubuntu 20.04 with a NVIDIA RTX 3060 GPU, but should also work for pretty much any other GPU supported by NVIDIA Docker and TensorFlow's Docker images.
If you haven't tried already, we can tell you that multi-component NVIDIA GPU setups with CUDA, TensorFlow, CARLA, etc. are extremely fragile and it's a total nightmare finding out the right versions that fit together. NVIDIA Docker simplifies those kind of setups by a lot and most importantly makes the development environment significantly more stable for various hardware setups, allowing developers with different machines to work together seemlessly.
First, we just update all apt packages of your system, so everything is up-to-date.
sudo apt-get update && sudo apt-get upgrade -yNote: Don't skip this even if installing a 100% fresh Ubuntu image. It seems hilarious, but those images don't come with the latest packages installed, even if you picked "install latest updates" in the Ubuntu installation dialog, so you need to update the packages first.
Now, we pick a NVIDIA driver for our CARLA simulations and AI trainings. Ubuntu already offers us an easy option to run a driver setup for installing proprietary NVIDIA drivers that are compatible with NVIDIA Docker, TensorFlow GPU and the CARLA simulation.
Unfortunately, there's no way to set up NVIDIA drivers safely from the CLI, so some clicks in the Ubuntu UI will be required. Don't try out scripted approaches, most of them break your system and cause lots of tears. We can tell you from our own experience.
- Press the Windows key and type software
- Select the UI app for updating Ubuntu packages
- Navigate to the "Additional Drivers" tab and select a proprietary NVIDIA driver (here it's version 470 for desktop PCs)
- Once the setup is done, restart your PC, so the GPU driver changes apply
Note: Don't waste any time thinking about using the open-source driver, it's not working!!!
Before diving into the Docker section, we need to set up git for pulling the source code. This is mostly required because Ubuntu 20.04 doesn't ship with Git LFS which is required to synchronize large binary files like pre-trained AI models, training datasets, etc.
sudo apt-get update && sudo apt-get install -y git git-lfsNext, we install Docker for building images and launching containers (care, the last command reboots your system).
# install cURL, Docker and Docker-Compose
sudo apt-get update && sudo apt-get install -y curl docker.io docker-compose
# register the NVIDIA Docker PPA as apt source
distribution=$(. /etc/os-release;echo $ID$VERSION_ID) \
&& curl -s -L https://nvidia.github.io/nvidia-docker/gpgkey | sudo apt-key add - \
&& curl -s -L https://nvidia.github.io/nvidia-docker/$distribution/nvidia-docker.list \
| sudo tee /etc/apt/sources.list.d/nvidia-docker.list
# install NVIDIA Docker from the official NVIDIA PPA
sudo apt-get update && sudo apt-get install -y nvidia-docker2
# allow non-root users to work with Docker (requires a reboot)
sudo usermod -aG docker $USER && rebootIf you're not familiar with Docker commands, check out the official docs:
- Docker: https://docs.docker.com/engine/reference/commandline/docker/
- Docker-Compose: https://docs.docker.com/compose/compose-file/compose-file-v2/
See the official NVIDIA docs for further information on NVIDIA Docker.
According to a GitHub post, running a CARLA simulation can be done like this using NVIDIA Docker:
# launch a bash terminal session to a CARLA 0.9.10.1 environment (version used for this seminar)
docker run -it --gpus all \
-e DISPLAY=$DISPLAY --net=host --runtime=nvidia \
--security-opt seccomp=unconfined \
carlasim/carla:0.9.10.1 /bin/bash CarlaUE4.sh -openglNote: There is official documentation for simulating CARLA 0.9.10 - 0.9.12 with Docker, so check those out as well.
Visual Studio Code offers great flexibility and lightweight code editing for various programming languages. Some of its many upsides are the easy setup, proper Linux bash scripting and Docker support.
As already mentioned the setup is extremely easy (see the snap command below) and only takes several minutes. Installing extensions can be automated as well, making the annyoing IDE setup procedure 100% scriptable.
Following command installs the VSCode editor:
sudo snap install code --classicRecommended VSCode extensions are:
| Extension | Description |
|---|---|
| cpptools | Sets up useful CMake tools for compiling, testing, debugging, etc. |
| ms-python.python | Sets up useful Python tools for debugging, linting, running notebooks, etc. |
| ms-azuretools.vscode-docker | Sets up useful Docker tools for managing images / containers via UI |
| Gruntfuggly.todo-tree | Searches for TODO stubs in the code and lists them |
JupyterHub is a really nice solution to run Jupyter notebooks and quickly hack out some Python AIs. Gladly, TensorFlow officially provides their Docker image with all Jupyter + NVIDIA tools pre-installed, allowing devs like us to be productive without having to install our workstation PC for weeks just to make the GPU support work.
Following command launches a Docker container serving the JupyterHub website on port 8888. For access, open the signup link printed to the console and open it inside your browser. Now, you're good to go and can train TensorFlow AIs with GPU support!
# launch a JupyterHub server with GPU support
docker run -it --rm -v $(realpath ~/notebooks):/tf/notebooks -p 8888:8888 \
--runtime=nvidia --gpus all \
--security-opt seccomp=unconfined \
tensorflow/tensorflow:latest-gpu-jupyterIn case you prefer code linting, auto-complete, etc. (which is unfortunately missing in the JupyterHub server), you can also launch a local Jupyter notebook in VSCode and attach it to the JupyterHub server using the jupyter VSCode plugin. See this blog post for further information.
Info: The notebook file is kept local while executing cells inside the remote Python kernel. In case you want to read from a dataset, you need to mount it inside the Docker container hosting the JupyterHub server, so the remote kernel can see your files.
We hope this setup did work for you like a charm. If not, please let us know, so we can figure out a solution for you.