search

dattalab / pyhsmm-library-models

library models built on top of pyhsmm
0 stars 1 forks source link

'zero probability under model' error happens sometimes #34

Open mattjj opened 10 years ago

mattjj commented 10 years ago

I changed it to be just a warning, since it can just ignore any infringing data. I don't know why it would happen, though. It only seems to happen when running code in parallel.

Maybe the new library model classes aren't passing left_censoring correctly? Or is it something else?

mattjj commented 10 years ago

It only seems to happen on the left, so the (uncensored) durations member must not be generating its left side correctly.

mattjj commented 10 years ago

Okay, because of the difference between

 if any(feasible): error # wrong

and

if not any(feasible): error # right

I'm not sure this bug even exists at the moment...

EDIT: okay yeah it does still exist

mattjj commented 10 years ago

Only happens on the first iteration, so it's just an issue with the initialization from the prior, not from any of the resampling code.

alexbw commented 10 years ago

On new subHMM code, also only seems to happen in the first iteration, but sometimes happens multiple times in the first iteration.

mattjj commented 10 years ago

Multiple times just means it happened to multiple duration distributions on the first iteration.