Initializing larger models using samples from smaller models

While I avidly await the arrival of pathfinder, I’m exploring initializing my larger models using samples from the posterior of smaller models sampled against the same data (inspired by this). It’s straightforward enough to extract draws from a given model and write to csvs for inits of the same model, but when going from smaller to larger one needs to add inits for the parameters of the larger model that are missing from the smaller model.

I gather the route to achieve this is to use the new variable_skeleton() and constrain_variables(), but I thought I’d check here to see if anyone (@andrjohns perhaps?) had done this before I made my own an attempt.


Oh, turns out no actual effort is required here as I just read in the cmdstan docs that it’s permitted to supply inits for just a subset of the parameters, where those outside that subset will then initialize as usual (uniform draw from -2:+2 on the unconstrained scale).

1 Like