|
I am using Mark Schmidt's CRF library in MATLAB and trying to build upon the TrainCRF Demo, so I could use continuous input data. But I do not quite understand how to define the features when the input data is continuous. The data I'm using as inputs are basically feeds from several sensors (e.g. 10 channels of continuous real valued feeds), and at certain parts of the feed, for a certain duration, an event occurs, and those are the labels (e.g. I have 5 different labels for 5 different events). Meaning, at each time instance, I have 10 real valued sensor readings (my observations) and 1 appropriate label, which are properly aligned in time. I have a sequence of approx 20 000 time samples including all situations (labels). I would like to model this system using CRFs, so I could use the sample data to learn the parameters (w). I assume that I would have to introduce some kind of a probability density function to model the relation between the observations and the label? Of course, the problem I have is how to define this structure (e.g. in Matlab) or are there maybe some implementation examples? Thanks! |
