One channel full of NA's results in a seqfault

[helske]

One channel full of NA's results in a seqfault

[helske]

Glad you found out the source of the error, we'll have to work out something regarding that kind of case at some point (either a check and stop or fix the EM algorithm so that it is possible to model that kind of data as well).

[helske]

It's hard to debug without minimal working example. Are you altering the model manually after build_hmm? My first guess would be that there is some out of bound reference due to mismatched dimensions in C++ side. Or, it could be some actual bug due to recent changes in the EM algorithm. For tracking down that I would need more info though.

[vrodriguezf]

Finally the problem was in my sequence dataset.

I was introducing some "empty" sequences, full of NA symbols in one of the channels, and that was causing the EM algorithm to crash. I have considered those sequences as useless for my experiments and now it is working well :)

Thanks!

[helske]

As I started to work on this I realized that I had misread your previous comment. It is of course pointless to use model where one channel contains only NA's for each sequence (so informative error message should be provided in model building phase), but a case where some sequences are full of NA's in one or multiple channels should be possible. And I did not find any problems in that kind of cases in my experiments. Could you send a minimal example which causes the seqfault for you either here or by email to me (for example the model object used in call to fit_model)?

[helske]

There is now a check in fit_model which gives and error if one channel is full of NA's.