A summed i.i.d. log-likelihood with the observation held fixed

log_lik() and every MCMC step ask the same question over and over: the summed log-density of one fixed set of observations under a theta that changes. de_iid_evaluator() returns a closure over the observations, so whatever an estimator can settle once settles when the closure is built rather than on every call. For the MDN that is coercing the observations to a tensor, which at a few thousand rows is not a rounding error next to the forward pass.

Usage

de_iid_evaluator(de, x, max_batch = 1e+05)

Arguments

de

A fitted density estimator.

x

Standardized observations, n_obs x dim_x.

max_batch

Largest number of pairs evaluated at once.

Value

function(theta) giving one summed log-density per row of theta, in standardized space.

Details

Reducing inside the closure matters as much as the hoisting. The n_theta x n_obs matrix is the largest object in the loop and none of it is wanted, so the sum happens where the log-densities are produced and only n_theta numbers ever cross back into R.

Estimators need a method here only if they can beat the default, which is de_log_lik_iid() with its row sums taken block by block.