[SciPy-dev] [ANN] pyem 0.4.1, a numpy module for Gaussian Mixture Models

[David Cournapeau]

Hi there,

    a few weeks ago, I submitted a small module for Expectation 
Maximization for Gaussian Mixture Models. I had the time recently to 
improve it significantly, and I've just made available a second public 
release:

    http://www.ar.media.kyoto-u.ac.jp/members/david/pyem-0.4.1.tar.gz

    Example of training:

    http://www.ar.media.kyoto-u.ac.jp/members/david/example2.png

    The major user-visible changes are the use of distutils for 
packaging, and speed improvement for diagonal models. Now, on a pentium 
M @ 1.2 Mhz, 10 iterations of EM takes around 6 seconds for a 
20-dimension, diagonal model of 20 components with 10^4 points for 
training, including the k-mean initialization (pyrex version). This 
should make it usable for applications such as speaker recognition, 
etc... More speed improvements (particularly full covariance matrices) 
are to be expected once I managed to include my separate project which 
implements the whole EM for GMM in C.
   
    Examples are available in the module, and several parts of the 
module can be executed directly to see some basic uses (including plotting).

    I tried several tools which I am not very familiar with yet 
(distutils, pyrex), so there may be some problems depending on the 
configuration; I tested the package on linux x86 (ubuntu dapper), one 
machine with and and one without pyrex.

    As always, comments, suggestion, bugs reports are welcome,

    David