I'm trying to implement Gaussian Mixture Model and EM algorithm using a matlab toolbox called Netlab. It uses EM algorithem in order to find the parameters of gaussian components. When I use "full" covariance matrix I get an error that says "covariance matrix must be positive in chol() funcion"(chol() is used to something called "cholesky decomposition"). But when I choose "diag" covariance matrix or "Spherical" covariance matrix. I will appriciate your helpful ideas.
Hold on, Sajjad, I will zip 'Netlab' folder and post it here for you to test - as far as I remember I used to get similar error and did some corrections in the code. =].
I made a lot of customization to it but could not recall if I had any problem along the way.
Here's the code I use and it works for me with 'full' covariance matrix:
http://dl.dropbox.com/u/17356785/Datamining/Netlab_27_files_different.zip
From this zip you basically need 'gmmem.m' and 'gmminit.m' - that's where this error originates from and if you want to review the fix I made, search for 'Replace/fix NaN and Inf'. Hope it will work for you too, inform on the progress and good luck!
Hi Sajad,
Try using "gmdistribution.fit" instead. You can easily find parameters of gaussian components using this function. Matlab Help has an illustrative example for this function.
Yes, Hossein, but for this you need 'Statistics Toolbox' which is not free.
Dear Evaldas, thanks a million. Fortunately it works for me too. I'm happy :)
The only problem is that I got "Some posterior probability are zero" warnings. Although I increase the number of components there are some zero posteriors.
Thank you again
Thanks Hossein for your consideration. I am supposed to work with Netlab but I'm really
intrigued how "gmdistribution.fit" works. I try to use it too.
I'm very happy too, Sajjad, that it fixed your error. Warnings are OK. =].
- Badges
- Science topic