nnpd.nre.inference

Inference algebra independent of the simulator and of the experiment runner.

Functions

ensemble_log_ratio(model, observations, theta, ...)	Sum per-observation logits, without sigmoid clipping or ratio products.
summarize_density(theta, log_density, log_measure, levels)	Joint HLD sets, projected axis bounds, and a density-mode estimate.
normalized_mass(log_density, log_measure)

Module Contents

nnpd.nre.inference.ensemble_log_ratio(model, observations, theta, batch_size, device)

Sum per-observation logits, without sigmoid clipping or ratio products.

Temporary model inputs are bounded by batch_size (or one observation ensemble). Returns one score per candidate. All observations share the same candidate theta.

nnpd.nre.inference.summarize_density(theta, log_density, log_measure, levels)

Joint HLD sets, projected axis bounds, and a density-mode estimate.

log_measure is quadrature volume (continuous) or counting weight (discrete). HLD ranking uses DENSITY, not quadrature mass. Boundary ties are included as a whole, so enclosed mass may exceed the request. This also handles flat densities. These are candidate approximations, not guarantees about the continuous region.

nnpd.nre.inference.normalized_mass(log_density, log_measure)