Include (unconstrained) parameter names along with the inverse metric?

[jgabry]

CmdStan writes the inverse metric to a file, but it doesn't include the names of the unconstrained parameters corresponding to the rows/columns. See stan-dev/cmdstanr#716.

To do that I think it would just involve including the names returned by

  model.unconstrained_param_names(names, false, false);

in the JSON file, right? For example,

  {
    "stepsize": 0.1,
    "metric_type": "diag_e",
    "unconstrained_param_names": ["alpha", "beta.1", "beta.2"],
    "inv_metric": [0.1, 0.2, 0.3]
  }

This would make it much easier for users to figure out which unconstrained parameter elements correspond to which elements of the matrix.

[WardBrian]

I think this is a very reasonable extension. The code that writes these files actually lives at

stan/src/stan/mcmc/hmc/base_hmc.hpp

Lines 70 to 76 in d6e3d03

	void write_sampler_state_struct(callbacks::structured_writer& struct_writer) {
	struct_writer.begin_record();
	struct_writer.write("stepsize", get_nominal_stepsize());
	struct_writer.write("metric_type", z_.metric_type());
	struct_writer.write("inv_metric", z_.inv_e_metric_);
	struct_writer.end_record();
	}

, so I'll move this to stan

[bob-carpenter]

I'm confused about why the step size is included in the metric output. Is this meant to be more general? What's the function name in CmdStanPy or CmdStanR that gets this? What's the name of the saved file?

[WardBrian]

The Stan issue that led to these files was just to write 'tuning parameters' #3181 / #3230

The default filename for them in cmdstan does end in _metric.json, I believe, but these files aren't directly exposed to the user (at least in cmdstanpy), the fields are read out individually

[jgabry]

What's the function name in CmdStanPy or CmdStanR that gets this? What's the name of the saved file?

In cmdstanr if you do

fit <- model$sample(..., save_metric = TRUE)

Then cmdstanr can tell you where CmdStan wrote the metric files if you call fit$metric_files(). You can also directly get the inverse mass matrix as an R object by calling fit$inv_metric(). But the rows and columns have no names, which is why I opened this issue (or really someone opened an issue in CmdStanR a long time that I just came back to because I ran into it myself).