What are "sampler diagnostic params"?


#1

Couldn’t figure out what “sampler diagnostic params” are from a cursory look through the C++ code. Perhaps someone else knows the code better can tell me?

  1. “sample params” (get_sample_param_names) include lp__ and accept_stat__
  2. “sampler params” (get_sampler_param_names) include stepsize__, treedepth__, etc.
  3. “sampler diagnostic params” (get_sampler_diagnostic_names) include ?

Nothing important hangs on this. I’m just curious at this point.


#2

You have to specify the diagnostic_file argument to sampling, but then it writes a CSV file with the values of the unconstrained parameters, the gradient with respect to the unconstrained parameters, and the momentums corresponding to each of the unconstrained parameters.


#3

These are the things like energy, stepsize, treedepth, accept stats, etc. The things that come out every iteration but aren’t parameters of the model or derived quantities (like the log density lp__).