Docker container for keras + cntk intended for nvidia-docker
Python Shell
Switch branches/tags
Nothing to show
Clone or download
Latest commit 4efba87 Jun 30, 2018


Docker container for keras + cntk intended for nvidia-docker. Based off of Durgesh Mankekar's minimal keras/jupyter container + corresponding blog post.

This container contains CNTK, TensorFlow, and Keras. When Keras is run in the container, CNTK will be the selected backend by default. Only recommended to run on a GPU instance.

Container on Docker Hub


To set up the CUDA+container drivers from scratch on the GPU instance, run these commands:

curl -O -s
sudo sh

You may need to run the script again after a few days (due to a GPU update on host hardware)

To run a deep learning script in the container:

sudo docker run --runtime=nvidia run -it --rm -v $(pwd)/:/keras --name keras minimaxir/keras-cntk python3 <x>.py

Where <x> is the Python script on the host server.

To run TensorFlow on Keras in the container instead of CNTK, add a -e KERAS_BACKEND='tensorflow' flag:

sudo docker run --runtime=nvidia run -it --rm -v $(pwd)/:/keras --name keras -e KERAS_BACKEND='tensorflow' minimaxir/keras-cntk python3 <x>.py

To run a Jupyter Notebook in the container (where invoking Keras in a notebook will use the CNTK backend):

sudo docker run --runtime=nvidia -it --rm -p 8888:8888 -v $(pwd):/keras --name jupyter minimaxir/keras-cntk jupyter notebook --allow-root

Run TensorBoard and Jupyter

Running both TensorBoard and Jupyter in the same container requires starting up TensorBoard in the background, then running Jupyter in the foreground:

sudo docker run --runtime=nvidia -d --rm -p 8888:8888 -p 6006:6006 -v $(pwd):/keras --name jupyter -e KERAS_BACKEND='tensorflow' minimaxir/keras-cntk tensorboard --logdir=/keras/logs
sudo docker exec -it jupyter bash
jupyter notebook --allow-root

TensorBoard will be available on Port 6006; you'll need to use the TensorBoard callback in Keras to make results visible in TensorBoard.

Note that with this approach, killing the Jupyter console will not kill the container; you have to exit bash and run sudo docker kill jupyter.

Error Note

If you get a docker: Error response from daemon: OCI runtime create failed: error when creating the container, try reinstalling the GPU drivers on the host:

curl -O -s
sudo sh ./ -a --ui=none --no-x-check && rm


Max Woolf (@minimaxir)

Max's open-source projects are supported by his Patreon. If you found this project helpful, any monetary contributions to the Patreon are appreciated and will be put to good creative use.


Install script sourced from Durgesh Mankekar's blog post noted above. (reproduced here for postertity)