Require every entry to be finite

NA, NaN and Inf all reach the estimators as a chol() failure or a non-finite validation loss, which blames training for a bad input. Naming the argument and the first offending position at the boundary is cheaper to act on.

Usage

check_finite(m, arg, allow_inf = FALSE)

Arguments

m

A numeric vector or matrix.

arg

Name of the argument.

allow_inf

Let Inf/-Inf through and only reject NA/NaN. Use this where Inf already has a well-defined meaning downstream: an MCMC posterior's theta routes through the prior's own density first, which correctly sends an infinite parameter value to zero mass rather than to the estimator.

Value

m, invisibly.