[mlpack] GMM result in NaN matrices

Ryan Curtin gth671b at mail.gatech.edu
Wed Apr 10 15:45:20 EDT 2013


On Wed, Apr 10, 2013 at 12:06:12PM +0200, Jaelon wrote:
> Hey,
> 
> I am currently working with GMM and it worked for two models with
> decent results up to now.
> However when attempting to train for a third model I get mean,
> covariance and transmission matrices filled with NaN.

You say transmission matrices, so do you mean that you are using GMMs
inside of HMMs?

> Could you give me an possible reason for this problem and/or point
> me in the direction to solving this, I have been stuck several hours
> on this now).
> 
> PS. I updated from the repository about 6hours ago (version r14885)

I sent an email to the list just a second ago detailing a fix in r14887
which avoids a divide-by-zero condition which is generally caused by
outliers in data.  I think that might fix your problem.  If not, let me
know.

-- 
Ryan Curtin       | "She fell..."
ryan at igglybob.com |   - Ludvig


More information about the mlpack mailing list