Deep Learning Frameworks The Easy Way (NVIDIA-Docker)

Install drivers and cuda.

Install Docker from the official site (NOT the Ubuntu repos)
sudo apt-get install docker.io DO NOT DO THIS
(What will happen is nvidia-docker will complain during installation that it can't find docker)
see https://github.com/NVIDIA/nvidia-docker/issues/134

Then install nvidia-docker
https://github.com/NVIDIA/nvidia-docker

Finally go to Docker Hub and find a suitable Docker image. e.g. "tensorflow/tensorflow:latest-gpu"

From what I understand these deep learning + gpu docker images are based on Nvidia's official CUDA images at https://hub.docker.com/r/nvidia/cuda

Notes:
Theano + GTX1060 needs the latest version of cuda, v8.0.
The latest binaries for Tensorflow + GTX1060 only work with CUDA v7.5

That means I need two different docker images, one for Tensorflow and one for Theano.

An alternative would be to build Tensorflow from source so that it can use v8.0 but I'm not experienced with that.