log_prob - allow only constrained or unconstrained inputs; allow multiple constrained inputs

[mitzimorris]

Summary:

• Revise the log_prob implementation to allow only constrained or only unconstrained params, in accord with the documentation

• Allow multiple constrained parameter inputs in the form of a CSV file, similar to standalone_gqs

Description:

The log_prob method exposes a Stan model's log_prob function which takes as an input the vector of parameter values on the unconstrained scale and returns the gradients.

Desiderata

• To amortize the cost of instantiating the model, the log_prob method should be able to accept an array of parameter vectors and then call the model's log_prob method on each.

• Depending on the use case, the input parameters may be available on the constrained or unconstrained scale.

Current implementation

• Constrained parameters are specified as a JSON object with entries for all parameters in the model, structured accordingly. Because the input file can only contain one JSON object, the current implementation doesn't admit of getting radients for more than one set of parameters on the constrained scale.

• The documentation states that the user can only specify one or the other.

• The code, as written, allows for inputs in the form of a file of one or more sets of parameters on the unconstrained scale and/or a file of a single set of parameter values on the constrained scale. If both are present, the logic will first computes the gradients for all unconstrained parameter sets, then compute the gradients for the constrained parameter set. This means that the last row of the output corresponds to the constrained parameters, if constrained parameter input file is present.

cmdstan/src/test/interface/log_prob_test.cpp

Lines 42 to 54 in 6bfcc1c

	TEST_F(CmdStan, log_prob_good_rdump) {
	std::stringstream ss;
	ss << convert_model_path(bern_log_prob_model)
	<< " data file=" << convert_model_path(bern_data)
	<< " output file=" << convert_model_path(dev_null_path)
	<< " method=log_prob unconstrained_params="
	<< convert_model_path(bern_unconstrained_params_rdump)
	<< " constrained_params="
	<< convert_model_path(bern_constrained_params_rdump);
	std::string cmd = ss.str();
	run_command_output out = run_command(cmd);
	ASSERT_FALSE(out.hasError);
	}

Reproducible Steps:

Changed CmdStan code to check for this; unit tests fail.
.

Current Version:

v2.31.0

[mitzimorris]

checking the header - all other methods output lp__ - two underscores - log_prob CSV output has only one - lp_.
technically OK, but inconsistent.