Gael Varoquaux GaelVaroquaux

Notes on scikit-learn round table

Q:What possible additions to scikit-learn are important to you?

Xavier Dupré (Microsoft): keep .fit in the API but X can be a stream from Spark for example. Transparent for the user. Gaël: indexable and len is n_samples, is that good enough? Answer: X accessible through sequential iterator.
Jean-François Puget (IBM): IBM is betting on Spark at the scale of the company. Most machine learning applications have small data but some don't. How can the bridge b/w scikit-learn and Spark get better? How to get scikit-learn used in a distributed environment? Not all algorithms can work out-of-core, need distributed algorithm.
Jean-Paul Smet (Nexedi): Nexedi is an example company. Wendelin.core helps us removing the overhead, and enabling out of core computing. Next step of the story in a year.

Introductory quote:

"Machine learning people use hugely complex algorithms on trivially simple datasets. Biology does trivially simple algorithms on hugely complex datasets."

Concepts of reproducible science

Replicability

	# how to use : df should be the dataframe restricted to categorical values to impact,
	# target should be the pd.series of target values.
	# Use fit, transform etc.
	# three types : binary, multiple, continuous.
	# for now m is a param <===== but what should we put here ? I guess some function of total shape.
	# I mean what would be the value of m we want to have for 0.5 ?

	import pandas as pd
	import numpy as np

	"""
	A scikit-learn like transformer to remove a confounding effect on X.
	"""

	from sklearn.base import BaseEstimator, TransformerMixin, clone
	from sklearn.linear_model import LinearRegression
	import numpy as np

	class DeConfounder(BaseEstimator, TransformerMixin):
	""" A transformer removing the effect of y on X.

	"""
	Fast counting of 3-grams for short strings.


	Quick benchmarking seems to show that pure Python code is faster when
	for strings less that 1000 characters, and numpy versions is faster for
	longer strings.

	Very long strings would benefit from probabilistic counting (bloom
	filter, count min sketch) as implemented eg in the "bounter" module.

	import numpy as np
	import pylab as pl
	import matplotlib.transforms as mtransforms

	################################################################################
	# Display correlation matrices

	def fit_axes(ax):
	""" Redimension the given axes to have labels fitting.
	"""


	"""Persistence strategies comparison script.

	This script compute the speed, memory used and disk space used when dumping and
	loading arbitrary data. The data are taken among:
	- scikit-learn Labeled Faces in the Wild dataset (LFW)
	- a fully random numpy array with 10000x10000 shape
	- a dictionary with 1M random keys/values
	- a list containing 10M random value
	The compared persistence strategies are:

	import numpy as np
	import matplotlib.pyplot as plt

	from sklearn.linear_model import Ridge, Lasso
	from sklearn.cross_validation import ShuffleSplit
	from sklearn.grid_search import GridSearchCV
	from sklearn.utils import check_random_state
	from sklearn import datasets

	'''
	Non-parametric computation of entropy and mutual-information

	Adapted by G Varoquaux for code created by R Brette, itself
	from several papers (see in the code).

	This code is maintained at https://github.com/mutualinfo/mutual_info
	Please download the latest code there, to have improvements and
	bug fixes.

	### Keybase proof

	I hereby claim:

	* I am GaelVaroquaux on github.
	* I am gaelvaroquaux (https://keybase.io/gaelvaroquaux) on keybase.
	* I have a public key whose fingerprint is 44B8 B843 6321 47EB 59A9 8992 6C52 6A43 ABE0 36FC

	To claim this, I am signing this object: