Based on https://github.com/tensorflow/tensorflow/tree/master/tensorflow/tools/docker
Nothing unusual here: docker build -t TAG [-f Dockerfile.gpu]
Once built (you have to do this yourself for now), run the commands in run-cpu or run-gpu (with your modifications, if needed). Both are configured so that you can pass an absolute path to a directory for jupyter notebooks: run-cpu PATH