Awesome

A Python Wrapper for PROPACK

This wrapper is incomplete, buggy, unsupported, and does questionable things with memory pointers. Use it at your own risk.

• Author: Jake Vanderplas vanderplas@astro.washington.edu
• License: BSD

About

This is a Python wrapper for the PROPACK package, which implements efficient partial singular value decompositions of large sparse matrices and linear operators. The package is meant to stand alone, but may be included in Scipy at a later date.

Installation & Testing

To install in-place, type

python setup.py build_ext --inplace

To run the few unit tests that currently exist, type

nosetests pypropack/tests

Usage

Currently very straightforward. You can do

>>> from pypropack import svdp
>>> import numpy as np   
>>> A = np.random.random((10, 20))
>>> u, s, vt = svdp(A, k=3)  # compute top k singular values

The svdp documentation has more information.

Benchmarks

use the script plot_benchmarks.py for a comparison of PROPACK, ARPACK, SVDLIBC, and the standard LAPACK approaches to computing the svd.

PROPACK modifications

PROPACK had to be slightly modified to play well with gfortran. The PROPACK code used is version 2.1, downloaded from http://soi.stanford.edu/~rmunk/PROPACK/

For compatibility with gfortran, the files PROPACK/*/second.f needed to be slightly modified. The line

REAL      ETIME
EXTERNAL  ETIME

was changed to

EXTERNAL REAL ETIME