[mlpack] HMM problem with nan

Ryan Curtin gth671b at mail.gatech.edu
Mon Jul 21 11:21:13 EDT 2014


On Mon, Jul 21, 2014 at 12:13:03PM -0300, Thuener Silva wrote:
> I think there is some problem with the train model using the Baum-Welch
> algorithm, with only the given unlabeled observations(void
> HMM<Distribution>::Train(const std::vector<arma::mat>& dataSeq))
> 
> Sometimes the mean and covariance are returned all nan and it takes a long
> time to terminate the train. Another strange think is that I have to
> reinstanciate the HMM to get out of this problem. Even changing the initial
> mean and covariance, I get same erro if I don't resinstanciate the HMM.
> 
> I'm trying to get to the root of the problem. I already know that after
> some time the value of oldLoglik is nan and with that then the Baum-Welch
> will only exit when it get to maximum number of iterations(1000).
> 
> How get the debug messages from mlpack?

Hello Thuener,

When you configure mlpack with CMake, pass the argument -DDEBUG=ON,
i.e.,

(in directory /path/to/mlpack/build/)
$ cmake -DDEBUG=ON ../

and then when you compile against mlpack, you can pass -DDEBUG and -g to
the compiler.

If you can reproduce the error, you could file a bug on Trac and maybe
we can help dig to the bottom of the issue.

Thanks,

Ryan

-- 
Ryan Curtin    | "I can't believe you like money too.  We should
ryan at ratml.org | hang out."  - Frito


More information about the mlpack mailing list