Docker

NVIDIA Cuda Docker with JupyterLab.

Assuming that Ubuntu is used.

0) Installing Docker and NVIDIA Container Runtime:

Run bash install-docker.sh to install docker (half-automated). Installation instructions are similar to the one in Docker Website.

Run nvidia-container-runtime.sh to install NVIDIA Docker container runtime. Installation instructions are from NVIDIA Docker Github. After the script is executed completely, run docker run --gpus all nvidia/cuda:10.0-base nvidia-smi to test if NVIDIA is working properly with the container.

1) Pulling Cuda Docker Image and running in a container:

Run ts_docker.sh to pull docker image and mount volumes.

Details: The last line specifies the image being pulled, (i.e. tensorflow/tensorflow for Tensorflow Image, nvidia/cuda:10.0-base for Cuda 10.0 Image), can be changed accordingly. If Cuda Image is pulled, run nvidia-smi to test if NVIDIA is working in the container.

Mounted volumes can be changed in the -v command (i.e. Assuming that you have a folder in home directory, the same folder that you want to mount will appear in the docker container, and any changes you make in that folder will also be changed in the Docker container automatically). To be specific, the command is run such that -v <directory in your Ubuntu to link>\<directory in the Docker container>. PycharmProjects directory is also mounted in the script to be used with Pycharm.

--gpus all argument is to enable running the container with GPU.

Ports can be configured in the -p argument.

--name <name> specifies the container name, instead of randomly generated names by Docker.

To link a Docker container with another Docker container, use --link <container_to_link>:<container_to_link>. Details on container linking at Legacy Container Links website.

2) Installing packages in the container

Once in the container, run bash install.sh to install the packages. Packages installed using apt-get will be installed from packages.sh, while packages installed using pip will be installed from requirements.txt. Edit the files accordingly for packages to be installed.

Note: The most straightforward way to do so is by putting install.sh, packages.sh, requirements.txt inside a mounted folder (and changing directory) to be able to install the necessary packages. Running the bash install.sh script will automatically install the packages inside the Docker container. Else, you may create your own script or manually install it yourself.

3) JupyterLab

Once in the container, enter the following command to run JupyterLab jupyter lab --ip=127.0.0.1 --port=8888 --allow-root

Note: Port should first be changed accordingly in ts_docker.sh file to enable running. To enter Jupyter, click on the link in terminal, or enter 127.0.0.1:8888. Enter the token provided in the terminal to start browsing. The IP and port should be changed accordingly based on how the Jupyter notebook command is run.

* Basic Docker commands

• Check (active) containers: docker ps (-a)
• Check images: docker images
• Pull images from Docker repository: docker pull image(/image):image-tags
• Start existing container:
  • docker start <container id>
  • docker attach <container id>
• Start new container: docker exec -it <container-id> /bin/bash - to run as bash for detached containers (-d)
• Stop container: docker stop <container id>
• Delete (Note: You must delete all containers related to an image before deleting the image):
  • Container: docker rm <container id>
  • Image: docker rmi <container id>