SPArse Modeling Software
now open-source!

This software is distributed under GPLv3 licence.

Depending on the part of the toolbox you use

J. Mairal, F. Bach, J. Ponce and G. Sapiro. Online Learning for Matrix Factorization and Sparse Coding. Journal of Machine Learning Research, volume 11, pages 19-60. 2010.

J. Mairal, F. Bach, J. Ponce and G. Sapiro. Online Dictionary Learning for Sparse Coding. International Conference on Machine Learning, Montreal, Canada, 2009

or

J. Mairal, R. Jenatton, G. Obozinski and F. Bach. Network Flow Algorithms for Structured Sparsity. Adv. Neural Information Processing Systems (NIPS). 2010.

R. Jenatton, J. Mairal, G. Obozinski and F. Bach. Proximal Methods for Sparse Hierarchical Dictionary Learning. International Conference on Machine Learning. 2010.

It works on Linux and Mac OS and Windows. We provide 64bits and 32bits versions. A 64bits Windows version will probably be included in a future release.

At the moment, yes. Future releases might include interfaces with other environments such as R, Python, but also the possiblity to call the functions of the toolbox from any C++ program. For the Windows 32bits version, Matlab >= R2009b is eventually required.

SPAMS is designed/optimized to efficiently solve a large number of small/medium size sparse decomposition problems, as they often occur in image processing, not a single large decomposition problem.

It indicates the number of threads that should be used. If set to "-1", it is set automatically to the number of the cores of the machines, which is a good idea in general. However, it is often a bad idea to use more threads than available resources. For instance, on a Quad-core machine, if one core is used 100 percent of the time by a process, then you should consider that only three cores are available and set param.numThreads to three. Another example: On a dual-core machine, if the two cores are available and you run two instances of the program (using two Matlab), you should better set param.numThreads to 1.

Matlab uses its own version of the Intel Math Kernel Library. We could not manage to make it work under Matlab with a dynamic linking. This is why the mex-files of this version are so big.

This happens with some Matlab installation for an unknown reason. However, since it happens only when you exit Matlab, it should not be a big problem.

We have observed a better performance with the Intel/MKL version. This might not be true for every computer.

While this will be the topic of a future major release, this package already contains key building blocks to develop numerous such applications.

This will be the topic of a future major release.