TensorFlow Runtime Dockerfiles

Simple Dockerfiles for running TensorFlow, with Jupyter and GPU variants.

Maintainer: @angerson (TensorFlow, SIG Build)

These containers are built by an internal job at Google and published to tensorflow/tensorflow on Docker Hub. Here's a quick way to try out TensorFlow with GPU support and Jupyter:

docker run --gpus=all -it --rm -v $(realpath ~/notebooks):/tf/notebooks -p 8888:8888 tensorflow/tensorflow:nightly-gpu-jupyter

Refer to the tensorflow.org Docker installation instructions for more details.

Building Containers

Builds are straightforward. Here's a sample:

docker build --target=base --build-arg TENSORFLOW_PACKAGE=tf-nightly-cpu -t tensorflow-nightly -f cpu.Dockerfile .

Look at the Dockerfiles for full details.

The builds include very simple import tests to verify that the packages work. You can run the tests like so:

docker build --target=test --build-arg TENSORFLOW_PACKAGE=tf-nightly-cpu -f cpu.Dockerfile .
docker build --target=base --build-arg TENSORFLOW_PACKAGE=tf-nightly-cpu -t tensorflow-nightly -f cpu.Dockerfile .

The test layer starts from the base layer, so the second command will complete instantly.

Contributions

If you would like to contribute a small change, please make a pull request. For large changes such as support for additional platforms, please clone this directory into a new directory and update the README to indicate that you are the new maintainer.