I have timeseries data that I’d like to model with pm.sample_smc() by first getting the posterior given the data at timepoint 1, then pass that posterior as inits for a next run that has data for timepoints 1-3, then repeat until the model has “seen” all the data. Does anyone have any existing code to facilitate this workflow?
No but I’ve wanted to test that for a while. How fast is it?
Would be good to avoid recompiling the same functions in every call (if you set the data as mutable)
Working on my own code now; is there a straightforward way to extract unconstrained representation of draws from a trace object?
nm, figured out the key was
Sneak peek at progress:
Photoplethsymography data; uncertainty sufficiently low/accurate that the green true-value line largely hides the estimate uncertainty ribbon
That’s really cool!