Log-density of many observations under many parameter values

The cross product n_theta x n_obs, in standardized space. Only sum_iid = FALSE wants that matrix; everything else – log_lik()'s default and every MCMC step – wants its row sums, and gets them from de_iid_evaluator() without ever building it.

Usage

de_log_lik_iid(de, x, theta, max_batch = 1e+05)

Arguments

de

A fitted density estimator.

x

Standardized observations, n_obs x dim_x.

theta

Standardized parameters, n_theta x dim_theta.

max_batch

Largest number of pairs evaluated at once.

Value

An n_theta x n_obs matrix of log-densities.

Details

The default expands the cross product and makes one batched call, which is what a flow needs: its transforms depend on the observation as well as the parameter, so there is nothing to reuse between observations. Estimators whose conditional distribution depends on the parameter alone – the MDN and the linear-Gaussian baseline – override this and compute that distribution once per parameter, which turns the i.i.d. sum from n_theta * n_obs network passes into n_theta of them. With a few thousand observations that is the difference between usable and not.