2010/3/18 alex <[email protected]>:
> Hi William,
>
>> I'm curious if anybody reading this uses the Hidden Markov Models
>> (HMM) code in Sage for anything.  If so, send me an email, since I'm
>> working on 100% replacing it by some brand new better quality code.
>
> I have not used HMM code in Sage for anything, but I am interested in HMM 
> algorithms in Python.  I have two use cases for HMMs which I could not coax 
> from GHMM.
>
> 1) Posterior decoding of a very long (not in RAM) sequence using fully 
> general emission distributions (i.e. not finite or mixed Gaussian).

Is this the Viterbi algorithm?

One of my problems with the GHMM library in Sage now is that it is
very memory inefficient when running the Viterbi algorithm, and it
leaks a lot of memory too.    My reimplementation of this for the next
HMM package in Sage is much more memory efficient (it doesn't save
everything when running the forward algorithm for Viterbi).  However,
of course it won't be as powerful as something disk based... unless
one has more RAM than disk (like I do on my 128GB computers).

I didn't implement arbitrary distributions (only finite, Gaussian, and
mixed Gaussian) yet, but I'll add that to my todo list.   I think the
only hard (impossible in general) part in general is Baum-Welch, but
the other algorithms are straightforward in general, so I'll just not
implement Baum-Welch for a general distribution.

> For this problem I wrote this thing last week:
> http://github.com/argriffing/hmmus

Cool -- I just skimmed some of the code.

> 2) Exact or bounded-error evaluation of per-position hidden state 
> distributions for general emission distributions and smallish sequences.  
> This is where I might find a Sage implementation particularly useful.
>

This is on my todo list.    More precisely, I'll have a generic HMM
class, and one could simply work over RealIntervalField, and hence get
output with provable error bounds.   This will of course be slow
compared to working with doubles, but should be useful for certain
applications.

> I see that Bilmes also @ washington.edu has written a graphical models 
> program and that "work is underway to make the toolkit available in full open 
> source form"
> http://ssli.ee.washington.edu/~bilmes/gmtk/
> so maybe this could eventually go into Sage?


-- 
William Stein
Associate Professor of Mathematics
University of Washington
http://wstein.org

-- 
To post to this group, send an email to [email protected]
To unsubscribe from this group, send an email to 
[email protected]
For more options, visit this group at http://groups.google.com/group/sage-devel
URL: http://www.sagemath.org

Reply via email to