Using a fixed step size for HMC (or NUTS) and other sampler options

Algorithms
21

It actually seems like n_leapfrog__ is not tracking anything like the number of log_prob evaluations… For example, if I run for 1 warmup and 1 sample, my C++ instrumentation shows that log_prob is executed 36 times, but I get n_leapfrog__ set to 1 for the one sample that gets output.

Anyone have other ideas for tracking this within Stan or is my instrumented C++ the best way to go?

22

log_prob is evaluated multiple times at startup to ensure valid initialization and estimate run time. n_leapfrog quantifies the number of calls per transition after the initial set up.