Deep Learning Docker Image

Unified docker environment for DL development. Contains Jupyter, TF2.3 and Tensorboard.

Setup on Oracle cloud

To setup the requirements (NVIDIA drivers and docker plugin) see the ORACLE.md.

Setup

The following tutorial uses deepdream project as an example. Replace this with the name of your own project.

NOTE: It is possible that you don't have docker-compose installed. In that case simply install it with sudo pip install docker-compose.

Clone the current project onto your development server

git clone git@github.com:mskl/docker-deep-learning.git

Enter the folder and clone your project. If you want to run multiple projects, clone this dir, rename it to something else and then rename it again for different projects.

cd docker-deep-learning
git clone git@github.com:mskl/deepdream.git

Edit the .env file and change the values to your liking

# Which GPUs are visible to the docker
NVIDIA_VISIBLE_DEVICES=all

# Port running Jupyter Notebook
JUPYTERPORT=43603

# Port running TensorBoard server
TENSORBOARDPORT=43604

# Name of the project directory
PROJECTFOLDER=./deepdream/

# Preferably your nickname. If you are running multiple projects on same server,
# use different project names in each so there are no conflicts between them
COMPOSE_PROJECT_NAME="matyas"

Usage

Once the setup is done, you can use the following commands to interactively interact with the running docker images. Makefile contains paths to GPU image which requires nvidia driver to be installed. In case of running on device without nvidia driver, you can interact with docker by simply using docker-compose up and so on. The GPU commands are:

• make build to build a docker image
• make run to build, run and follow logs of an image
• make stop to stop a running image
• make bash to open bash inside running image

Once the container is running, the services should be by default accessible at:

• tensorboard - http://127.0.0.1/43604
• jupyter-notebook - http://127.0.0.1/43603

When using tensorboard callbacks, the logs directory is mapped inside the docker image as/logs.

Maintenance

When we need to update some of the python requirements, please update them and then create a PR.