I have just started to discover the
metric_file option from cmdstan and I must say it is really awesome. A few questions:
engagedoption be set to 0 if I give Stan the inverse mass matrix and the stepsize? I thought this would be the case, but I got very odd behavior such that I am confused. So when doing this Stan runs super-super fast, but gives me crap. Maybe this is a bug.
Is there some documentation on the mapping from the parameters to the unconstrained parameter space. In particular I would be interested in ordering.
Do we have recommendations or some guidance (likely with big warnings) what can be the minimal warmup when giving the metric file?
Can cmdstan export the inverse masses? Currently I use rstan to parse things out.
Can rstan already use the inverse mass matrix (@bgoodri? )?
This is getting really interesting in the context of MPI. Now you can use all available CPUs you have to get a single chain through warmup ASAP. Then you just start of as many chains as you like using the already calculated inverse weights. Certainly this is dangerous, but likely very handy for model-development.