Skip to content

UW-AMO/spgl1

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

67 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

SPGL1-General can solve regularization problems with the following features: 

1) nonlinear forward models
2) general misfit functions (e.g. robust penalties like huber)
3) general regularizers 
------------------------------------------------------

SPGL1-General was forked from SPGL1 by Aleksandr Aravkin, in July 2012. 

References: 

1) Incorporating robust misfit functions, and general regularizers, see 

    A.Y. Aravkin, J.V. Burke, M.P. Friedlander, Variational properties of value functions, 
    SIAM Journal of Optimization (SIOPT), Vol. 23, No. 3, pp. 1689?1717, August 2013.

2) Nonlinear forward models that allow factorized matrix factorization approaches, see 

    A.Y. Aravkin, R. Kumar, H. Mansour, B. Recht, and F.J. Herrmann, 
    Fast methods for denoising matrix completion formulations, with applications to robust 
    seismic data interpolation, submitted to SIAM Journal on Scientific Computing (SISC), 
    April 2013.


A MovieLens example from the SISC paper is in the folder SISC-Scripts:

spgl1/SISC-scripts/Movielens_CleverInit.m 

Please follow the instructions provided in the script to set up fast modeling operators. 
I will include additional examples in the future, including the robust examples from 
the SIOPT paper. 


The original content of the README file for SPGL1 appears below: 


1. Introduction
===============

Thank you for downloading the SPGL1 solver!  SPGL1 is a Matlab solver
for large-scale one-norm regularized least squares.  It is designed to
solve any of the following three problems:

1. Basis pursuit denoise (BPDN):
   minimize  ||x||_1  subject to  ||Ax - b||_2 <= sigma,

2. Basis pursuit (BP):
   minimize   ||x||_1  subject to  Ax = b
 
3. Lasso:
   minimize  ||Ax - b||_2  subject to  ||x||_1 <= tau,

The matrix A can be defined explicily, or as an operator (i.e., a
function) that return both both Ax and A'y.  SPGL1 can solve these
three problems in both the real and complex domains.


Home page: http://www.cs.ubc.ca/labs/scl/spgl1/


2. Quick start
==============

Start Matlab and make sure the working directory is set to the
directory containing the SPGL1 source files. When this is done, run

 >> spgdemo

at the Matlab prompt.  This script illustrates various uses of SPGL1:

- Solve (BPDN) for some sigma > 0
- Solve (Lasso)
- Solve (BP)
- Solve a (BP) problem in complex variables
- Sample the entire Pareto frontier (i.e., ||Ax-b||_2 vs ||x||_1)
  for a small test problem.


3. Installation
===============

3.1  MEX interface
------------------

A vital component of SPGL1 is a routine (oneProjector.m) for
projecting vectors onto the one-norm ball.  The default distribution
includes a pure Matlab version of oneProjector which should work on
all platforms, and also a C-version of this routine that is more
efficient on large problems.  Precompiled MEX interfaces to the C
implementation of oneProjector are included for Windows
(oneProjector.dll), Linux/x86 (oneProjector.mexglx) and MacOSX/Intel
(oneProjector.mexmaci).  If you need to compile the MEX interface on
your own machine, run the following command at the Matlab prompt:

 >> spgsetup

or, equivalently, change to the "private" directory and issue the
command

 >> mex oneProjector.c oneProjector_core.c -output oneProjector -DNDEBUG

If the MEX interface cannot be found, SPGL1 falls back to the slower
Matlab implementation of oneProjector.

3.2  Path
---------

In order to use SPGL1 from any directory other than the one
containing the main spgl1 routine, add the SPGL1 package to your
default path:

 >> addpath <dir-name>

where <dir-name> is the location of spgl1.m.  You can also add this
command to your startup.m file.

4. References
=============

The algorithm implemented by SPGL1 is described in the paper

- E. van den Berg and M. P. Friedlander, "Probing the Pareto frontier
  for basis pursuit solutions", SIAM J. on Scientific Computing,
  31(2):890-912, November 2008

- Sparse optimization with least-squares constraints E. van den Berg
  and M. P. Friedlander, Tech. Rep. TR-2010-02, Dept of Computer
  Science, Univ of British Columbia, January 2010

About

A solver for large-scale sparse reconstruction

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • MATLAB 80.6%
  • C 13.9%
  • CSS 3.3%
  • M 2.2%