This post documents setting up a Jupyter/TensorFlow Docker stack that performs well on an old Core2Duo CPU and Windows 7. I do a lot of projects with Jupyter notebooks and TensorFlow on this system—I’ve discovered that:

  • TensorFlow is quicker on a Virtual Box Linux guest system.
  • TensorFlow with Eigen is quicker than TensorFlow with MKL, which is the default math library.

Altogether, TensorFlow with Eigen on Linux trains convolutional neural networks 4x faster than default TensorFlow on Windows 7. However, using a full Linux distro in Virtual Box eats a significant chunk of ram. The steps outlined in this post use Docker Toolbox and Jupyter Docker Stacks for a memory efficient solution. Additionally, it allows me to work on files stored locally on the Windows 7 host.

Steps Outline

Setup Docker Toolbox

Docker Toolbox is the only option for running Docker on Windows 7. It uses Virtual Box for the Docker virtual machine.

Besides just running the Docker Toolbox installer, do the following steps:

Change .docker Directory (optional)

The Docker Virtual Box image can get quite big and you may not want to store it in the default install location. The easiest solution I found involves moving the .docker folder to the desired location and then creating a junction1:

$ C:\Users\username>mklink /j .docker D:\.docker
Junction created for .docker <<===>> D:\.docker

The image can also be resized by deleting it and creating a new one with a different size2:

$ docker-machine rm default
$ docker-machine create -d virtualbox --virtualbox-disk-size "20000" default

Setup Port Forwarding

To access the Jupyter Notebook server with a local web browser, you need to use port forwarding for the Docker virtual machine in Virtual Box:

Port Forwarding

Setup Shared Folder

To work on local files, setup a shared folder for the local directory you work from:

Shared Folder

Make Custom Docker Image

The Jupyter Docker Stacks are a collection of pre-made Docker images for working with Jupyter notebooks. For this project, the Deep Learning Stack is close to what I want, so I use it as a starting point.

Dockerfile

Below is an edited Dockerfile from the Deep Learning Stack that does the following:

  • Installs tensorflow-eigen instead of tensorflow.
  • Adds custom.css to the Docker image for Jupyter notebook styling.
  • Installs jupyter_contrib_nbextensions and enables the spell checker by default.

The edited Dockerfile looks like this:

# Copyright (c) Jupyter Development Team.
# Distributed under the terms of the Modified BSD License.
ARG BASE_CONTAINER=jupyter/scipy-notebook
FROM $BASE_CONTAINER

# Install Tensorflow
RUN conda install --quiet --yes \
	'tensorflow-eigen=1.12*' \
	'keras=2.2*' && \
	conda clean -tipsy && \	
	fix-permissions $CONDA_DIR && \
	fix-permissions /home/$NB_USER
	
COPY custom.css /home/$NB_USER/.jupyter/custom/

# Install nbextensions
RUN conda install --quiet --yes -c 'conda-forge' \
	'jupyter_contrib_nbextensions=0.5*' && \
	conda clean -tipsy && \	
	fix-permissions $CONDA_DIR && \
	fix-permissions /home/$NB_USER && \
	jupyter nbextension enable spellchecker/main

Build Image

Once the files are created, place them in a folder like so:

tensorflow-eigen-notebook
├── custom.css
└── Dockerfile

Opening the Docker Quick Start Terminal will start the Docker virtual machine. In the terminal, navigate into the tensorflow-eigen-notebook directory and run:

docker build -t `jupyter/tensorflow-eigen-notebook` .

Usage

To start the notebook server on a local directory, first use the Docker Quick Start Terminal (or start Docker some other way). Navigate to the directory containing the files you want to work on, then run:

docker run --rm -p 8888:8888 -v "$PWD":/home/jovyan/work jupyter/tensorflow-eigen-notebook

Next, start a web browser and go to 127.0.0.1:8888 to access the notebook tree. When you want to shut the notebook server down, just select quit from the web UI.

Useful Commands

When you’re finished working and want to free up resources, remember to stop the Docker virtual machine from the terminal:

docker-machine stop

You may want to adjust how much memory is available to the Docker virtual machine in Virtual Box based on how much memory Docker needs. To see how much memory and other resources Docker is using, run:

docker stats

  1. Change Docker machine location - Windows ↩

  2. Increase Disk Space on Docker Toolbox ↩