Parallel Machine Learning with scikit-learn and IPython
Video recording of this tutorial given at PyCon in 2013. The tutorial material
has been rearranged in part and extended. Look at the title of the of the
notebooks to be able to follow along the presentation.
Browse the static notebooks on nbviewer.ipython.org.
Scope of this tutorial:
-
Learn common machine learning concepts and how they match the scikit-learn
Estimator API. -
Learn about scalable feature extraction for text classification and
clustering -
Learn how to perform parallel cross validation and hyper parameters grid
search in parallel with IPython. -
Learn to analyze the kinds of common errors predictive models are subject to
and how to refine your modeling to take this analysis into account. -
Learn to optimize memory allocation on your computing nodes with numpy memory
mapping features. -
Learn how to run a cheap IPython cluster for interactive predictive modeling on
the Amazon EC2 spot instances using StarCluster.
Target audience
This tutorial targets developers with some experience with scikit-learn and
machine learning concepts in general.
It is recommended to first go through one of the tutorials hosted at
scikit-learn.org if you are new to scikit-learn.
You might might also want to have a look at SciPy Lecture
Notes first if you are new to the NumPy /
SciPy / matplotlib ecosystem.
Setup
Install NumPy, SciPy, matplotlib, IPython, psutil, and scikit-learn in their latest
stable version (e.g. IPython 2.2.0 and scikit-learn 0.15.2 at the time of
writing).
You can find up to date installation instructions on
scikit-learn.org and
ipython.org .
To check your installation, launch the ipython interactive shell in a console
and type the following import statements to check each library:
>>> import numpy
>>> import scipy
>>> import matplotlib
>>> import psutil
>>> import sklearn
If you don't get any message, everything is fine. If you get an error message,
please ask for help on the mailing list of the matching project and don't
forget to mention the version of the library you are trying to install along
with the type of platform and version (e.g. Windows 8.1, Ubuntu 14.04, OSX
10.9...).
You can exit the ipython shell by typing exit.
Fetching the data
It is recommended to fetch the datasets ahead of time before diving into the
tutorial material itself. To do so run the fetch_data.py script in this
folder:
python fetch_data.py
Using the IPython notebook to follow the tutorial
The tutorial material and exercises are hosted in a set of IPython executable
notebook files.
To run them interactively do:
$ cd notebooks
$ ipython notebook
This should automatically open a new browser window listing all the notebooks
of the folder.
You can then execute the cell in order by hitting the "Shift-Enter" keys and
watch the output display directly under the cell and the cursor move on to the
next cell. Go to the "Help" menu for links to the notebook tutorial.
Credits
Some of this material is adapted from the scipy 2013 tutorial:
http://github.com/jakevdp/sklearn_scipy2013
Original authors:
- Gael Varoquaux @GaelVaroquaux, http://gael-varoquaux.info
- Jake VanderPlas @jakevdp, http://jakevdp.github.com
- Olivier Grisel @ogrisel, http://ogrisel.com