Spark acceleration for Scikit-Learn

This project is a major re-write of the spark-sklearn project, which seems to no longer be under development. It focuses specifically on the acceleration of Scikit-Learn's cross validation functionality using PySpark.

Improvements over spark-sklearn

scikit-spark supports scikit-learn versions past 0.19, spark-sklearn have stated that they are probably not going to support newer versions.

The functionality in scikit-spark is based on sklearn.model_selection module rather than the deprecated and soon to be removed sklearn.grid_search. The new model_selection versions contain several nicer features and scikit-spark maintains full compatibility.

Installation

The package can be installed through pip:

pip install scikit-spark

It has so far only been tested with Spark 2.2.0 and up, but may work with older versions.

Supported scikit-learn versions

• 0.18 untested, likely doesn't work
• 0.19 supported
• 0.20 supported
• 0.21 supported (Python 3 only as sklearn dropped support for Python 2)

Usage

The functionality here is meant to as closely resemble using Scikit-Learn as possible. By default (with spark=True) the SparkSession is obtained internally by calling SparkSession.builder.getOrCreate(), so the instantiation and calling of the functions is the same (You will preferably have already created a SparkSession).

This example is adapted from the Scikit-Learn documentation. It instantiates a local SparkSession, and distributes the cross validation folds and iterations using this. In actual use, to get the benefit of this package it should be used distributed across several machines with Spark as running it locally is slower than the Scikit-Learn parallelisation implementation.

from sklearn import svm, datasets
from pyspark.sql import SparkSession

iris = datasets.load_iris()
parameters = {'kernel':('linear', 'rbf'), 'C':[0.01, 0.1, 1, 10, 100]}
svc = svm.SVC()

spark = SparkSession.builder\
    .master("local[*]")\
    .appName("skspark-grid-search-doctests")\
    .getOrCreate()

# How to run grid search
from skspark.model_selection import GridSearchCV

gs = GridSearchCV(svc, parameters)
gs.fit(iris.data, iris.target)

# How to run random search
from skspark.model_selection import RandomizedSearchCV

rs = RandomizedSearchCV(spark, svc, parameters)
rs.fit(iris.data, iris.target)

Current and upcoming functionality

• Current
  • model_selection.RandomizedSearchCV
  • model_selection.GridSearchCV
• Upcoming
  • model_selection.cross_val_predict
  • model_selection.cross_val_score

The docstrings are modifications of the Scikit-Learn ones and are still being converted to specifically refer to this project.

Performance optimisations

Reducing RAM usage

Coming soon