[mlpack-git] [mlpack] How to debug NaN log-likelihoods HMM ? (#570)

[Ryan Curtin]

I wish I had a better answer, but this has always been hard to debug.  Usually I end up trying to step through the program to find the first iteration where anything from stateProb has any NaNs in it.  Another idea might be to take a look in this stack overflow post:

http://stackoverflow.com/questions/3615724/how-to-trace-a-nan-in-c

Basically, the idea is to enable all floating point exceptions, then run inside gdb and catch the first NaN like that.

If you can provide your dataset, I'm happy to help debug, but fair warning---I'm pretty overloaded at the moment because of the Google Summer of Code applications, so it may be some time before I am able to fully debug the issue.  If you can take the time and figure out where the NaN is coming from, maybe we can add some kind of handler or better logic to help prevent these cases in the future.  Handling these types of problems with the GMM and HMM code has always been problematic and difficult...

---
Reply to this email directly or view it on GitHub:
https://github.com/mlpack/mlpack/issues/570#issuecomment-196360753
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.cc.gatech.edu/pipermail/mlpack-git/attachments/20160314/6feccf7e/attachment.html>