[Python-projects] HMM test cases

[Alexandre Fayolle]

On Sat, Feb 09, 2008 at 03:09:38PM +0200, Orestis Markou wrote:
> Hello,
>
> thanks for having logilab-hmm as open source.
>
> I've recently created my own HMM library (I wanted to learn about HMMs a 
> bit deeper), and it's working fine, but it's slow (it's coded in pure 
> Python). I've tried to use numpy but the results I get are wrong and 
> inaccurate. I've tried to create some unit tests to test my  
> implementation but it's got me nowhere (ie. the tests pass, but the real 
> model still gets wrong results). It seems that for simple models my 
> implementations work but for large ones (~350 states and observations), 
> errors start to accumulate.
>
> I was wondering if you have unit tests or some sample data that I could 
> use. I'm mainly interested in the Viterbi algorithm, so I'm looking for 
> complete HMM models and pairs of observations/states that could be used 
> to test the accuracy of my library. I've searched the net but I can't 
> find anything, and it's impossible to do this by hand.
>
> I would be grateful if you can point me to a direction.

To discuss logilab's hmm implementation, I suggest you join the
python-projects mailing list (cc'ed). 

Every thing wa have for hmm is in the source tarball
(http://ftp.logilab.org/pub/hmm/logilab-hmm-0.4.0.tar.gz) 

You may retreive a slightly more up to date version by installing
mercurial [1] and running:

hg clone http://www.logilab.org/cgi-bin/hgwebdir.cgi/logilab/hmm/rev/tip hmm

[1] http://www.selenic.com/mercurial/wiki/

-- 
Alexandre Fayolle                              LOGILAB, Paris (France)
Formations Python, Zope, Plone, Debian:  http://www.logilab.fr/formations
Développement logiciel sur mesure:       http://www.logilab.fr/services
Informatique scientifique:               http://www.logilab.fr/science
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 481 bytes
Desc: Digital signature
Url : http://lists.logilab.org/pipermail/python-projects/attachments/20080211/d9496281/attachment.pgp