Set up jupyter notebook with docker-compose and tensorflow

Using single command to set up whole environment that is isolated and self-contained is very useful not only for machine learning projects. Docker together with docker-compose can do it with very little effort from our side.

Here is how to create simple setup for environment with tensorflow gpu, python and jupyter notebooks using docker.

1. Create new directory docker_setup
2. In this directory create new file with name Dockerfile without extension and fill it with line:

   FROM tensorflow/tensorflow:latest-gpu-py3
   

   This is almost the simplest Dockerfile you can write. This will build a container from tensorflow/tensorflow:latest-gpu-py3 image without adding anything extra to it.

3. Create file docker-compose.yml and fill it with this content:

   version: '2.3'
   
   services:
     tensor:
        build: ./
        runtime: nvidia
        environment:
            - NVIDIA_VISIBLE_DEVICES=all
        volumes:
            - ./src:/notebooks
        ports:
            - 8888:8888
   

   This translates to the following:

   • Create 1 service called “tensor”
   • Build docker image from current directory
   • Use nvidia runtime (nvidia-docker)
   • Make all nvidia devices visible to the environment
   • Map “src” directory to “notebooks” directory that’s inside the docker container
   • Forward port 8888 to port 8888 in the container

4. We are missing “src” folder so let’s create it among Dockerfile and docker-compose.yml. Our jupyter notebooks will be stored there.

   

5. Run sudo docker-compose up command inside docker_setup directory. Jupyter notebook should be hosted locally on port 8888:

6. Browse 127.0.0.1:8888/?token=your token:

7. You can test if tensorflow is running by creating simple script in new notebook:

In case of any problems, whole solution is hosted on github: https://github.com/sikora507/docker-tensorflow-gpu-jupyter