Open mikewojnowicz opened 1 year ago
I'm trying to make a toy example of a non-stationary model (if you have one I'd like to see it too). I noticed the same pattern of shapes you noticed. In my case it's related to these two lines:
where each line takes 1 away from the length of the transition probabilities. Maybe the -1
in len(filtered_probs)-1
isn't necessary?
In the "non-stationary" setting (which seems to be operationally defined here as referring to time-dependent parameters), the
hmm_smoother
function applied to a time-series of T elements returns T-2 transition probabilities, rather than T-1 transition probabilities.E.g. if one enters into the unit tests here, and adds the two assertions below, both pass.
Isn't this a mistake?
hmm_expected_states
function in the oldssm
repo, which performed a similar role, returned anexpected_joints
value that had T-1, not T-2, entries.