Docker is awesome — more and more people are leveraging it for development and distribution. Instant environment setup, platform independent apps, ready-to-go solutions, better version control, simplified maintenance: Docker has a lot of benefits. But when it comes to data science and deep learning, there is a certain hitch. You have to memorize all those docker flags to share ports and files between host and container, create unnecessary scripts and deal with CUDA versions and GPU sharing. If you have ever seen this error, you know the pain: run.sh $ nvidia-smi Failed to initialize NVML: Driver/library version mismatch Our goal The purpose of this small post is to introduce you a sufficient set of Docker utilities and GPU-ready boilerplate we often use in our company. So, instead of this: docker run \--rm \--device /dev/nvidia0:/dev/nvidia0 \--device /dev/nvidiactl:/dev/nvidiactl \--device /dev/nvidia-uvm:/dev/nvidia-uvm \ You will end up with this: doc up Cool, right? What do we actually want to achieve: Manage our application state (run, stop, remove) using one command Save all those run flags to a single configuration file we can commit to a git repo Forget about GPU driver version mismatch and sharing Use GPU-ready containers in production tools like Kubernetes or Rancher So here is the list of tools we highly recommend for every deep learner: 1. CUDA First, you will need . It’s an absolute must-have, if you plan to train models yourself. We recommend to use installer type instead of , because it won’t mess your dependencies in future updates. CUDA toolkit runfile deb (Optional) How to check if it works: cd /usr/local/cuda/samples/1_Utilities/deviceQuerymake./deviceQuery # Should print "Result = PASS" 2. Docker You don’t want to pollute your computer with tons of libraries and be afraid of broken versions hell. Also, you won’t have to build and install stuff yourself — usually, software is already built for you and packed in image! is simple: Installing Docker curl -sSL | sh https://get.docker.com/ 3. Nvidia Docker A must have from NVIDIA if you use Docker — it really simplifies using GPU inside Docker containers. utility Installation is really simple: wget -P /tmp https://github.com/NVIDIA/nvidia-docker/releases/download/v1.0.1/nvidia-docker_1.0.1-1_amd64.debsudo dpkg -i /tmp/nvidia-docker*.deb Now, instead of sharing nvidia devices every time like this: docker run --rm --device /dev/nvidia0:/dev/nvidia0 --device /dev/nvidiactl:/dev/nvidiactl --device /dev/nvidia-uvm:/dev/nvidia-uvm you can use a command: nvidia-docker nvidia-docker run --rm nvidia/cuda nvidia-smi Also, you can stop worrying about driver version mismatch: docker plugin from Nvidia will solve your problems. 4. Docker Compose Super useful utility that allows you to store configuration in a file and manage application state more easily. Though it was designed to “compose” multiple docker containers together, docker compose is still very useful when you only have one service. Pick the stable version : docker run here curl -L https://github.com/docker/compose/releases/download/1.15.0/docker-compose-`uname -s`-`uname -m` > /usr/local/bin/docker-composechmod +x /usr/local/bin/docker-compose 5. Nvidia Docker Compose Unfortunately, Docker Compose doesn’t know that Nvidia Docker exists. Lucky, there is a solution: a tiny that generates configuration with driver. Install it using pip: Python script nvidia-docker pip install nvidia-docker-compose Now you can use command instead of . nvidia-docker-compose docker-compose Alternative If you don’t want to use , you can . Just add those options to your : nvidia-docker-compose pass volume-driver manually docker-compose.yml # Your nvidia driver version herevolumes:nvidia_driver_375.26:external: true...volumes:- nvidia_driver_375.26:/usr/local/nvidia:ro 6. Bash aliases But is 21 characters to type! That’s too much. nvidia-docker-compose Lucky we can use bash aliases. Open (sometimes ) in your favorite editor and type those lines: ~/.bashrc ~/.bash_profile alias doc=' 'alias docl='doc logs -f --tail=100' nvidia-docker-compose Update your settings by running . source ~/.bashrc Start a TensorFlow service Now we are ready to use benefits from all those stuff above. For example, let’s run a Tensorflow GPU-enable Docker container. In a project directory create file with the following content: docker-compose.yml version: '3' services:tf:image: gcr.io/tensorflow/tensorflow:latest-gpuports:- 8888:8888volumes:- .:/notebooks Now we can start TensorFlow Jupiter with a single command: doc up is an alias for — it will generate modified configuration file with correct and then run . doc nvidia-docker-compose nvidia-docker-compose.yml volume-driver docker-compose You can manage your service using the same command: doc logsdoc stopdoc rm# ...etc Conclusion But is it worth the effort? Let’s weigh the pros and cons here. Pros Forget about GPU device sharing You don’t have to worry about Nvidia driver version anymore We got rid of command flags in favour of clean and plain configuration No more flag to manage container state --name Well-known documented and widely used utilities Your configuration is ready for orchestration tools like Kubernetes that understand docker-compose files Cons You have to install more tools Is it production-ready? Yep. In our movies recommendation service we use GPU-accelerated TensorFlow network to calculate real time film selection based on user input. Movix We have three computers with Nvidia Titan X in behind Proxy API. Configuration is stored in regular files: because of that it’s really ease to setup development environment or deploy application on a new server. So far it works perfect. Rancher cluster docker-compose.yml Be prepared for the future of ML! If you have any questions or comments, feel free to write here or on twitter @deepsystemsru .

NVIDIA

Super

Twitter

Hacking GTA V for Carvana Kaggle Challenge

Too Long; Didn't Read

Get Free Unlimited IP Geo Data With IPinfo Lite

Docker Compose + GPU + TensorFlow = ❤️

About Author

Comments

TOPICS

THIS ARTICLE WAS FEATURED IN

Related Stories

Can you solve a person detection task in 10 minutes?

It Is Okay If You Don't Know What You Like. We Do (feat. Deep Recommendation Algorithms)

10 Machine Learning, Data Science, and Deep Learning Courses for Programmers in 2020

10 Computer Vision Startups on Product Hunt with the Most Upvotes

10 Best Entry Level Machine Learning Tutorials

10 Best + Free Machine Learning Courses Collection

Can you solve a person detection task in 10 minutes?

It Is Okay If You Don't Know What You Like. We Do (feat. Deep Recommendation Algorithms)

10 Machine Learning, Data Science, and Deep Learning Courses for Programmers in 2020

10 Computer Vision Startups on Product Hunt with the Most Upvotes

10 Best Entry Level Machine Learning Tutorials

10 Best + Free Machine Learning Courses Collection

Light-Mode

Classic

Newspaper

Dark-Mode

Neon Noir

Minty

HN StartUps