continue: Continues MCMC sampling

In deepgp: Bayesian Deep Gaussian Processes using MCMC

Continues MCMC sampling

Description

Acts on a gp, gpvec, dgp2, dgp2vec, dgp3, or dgp3vec object. Continues MCMC sampling of hyperparameters and hidden layers using settings from the original object. Appends new samples to existing samples. When vecchia = TRUE, this function provides the option to update Vecchia ordering/conditioning sets based on latent layer warpings through the specification of re_approx = TRUE.

Usage

continue(object, new_mcmc, verb, re_approx, ...)

## S3 method for class 'gp'
continue(object, new_mcmc = 1000, verb = TRUE, ...)

## S3 method for class 'dgp2'
continue(object, new_mcmc = 1000, verb = TRUE, ...)

## S3 method for class 'dgp3'
continue(object, new_mcmc = 1000, verb = TRUE, ...)

## S3 method for class 'gpvec'
continue(object, new_mcmc = 1000, verb = TRUE, re_approx = FALSE, ...)

## S3 method for class 'dgp2vec'
continue(object, new_mcmc = 1000, verb = TRUE, re_approx = FALSE, ...)

## S3 method for class 'dgp3vec'
continue(object, new_mcmc = 1000, verb = TRUE, re_approx = FALSE, ...)

Arguments

object	object from fit_one_layer, fit_two_layer, or fit_three_layer
new_mcmc	number of new MCMC iterations to conduct and append
verb	logical indicating whether to print iteration progress
re_approx	logical indicating whether to re-randomize the ordering and update Vecchia nearest-neighbor conditioning sets (only for fits with vecchia = TRUE)
...	N/A

Details

See fit_one_layer, fit_two_layer, or fit_three_layer for details on MCMC. The resulting object will have nmcmc equal to the previous nmcmc plus new_mcmc. It is recommended to start an MCMC fit then investigate trace plots to assess burn-in. The primary use of this function is to gather more MCMC iterations in order to obtain burned-in samples.

Specifying re_approx = TRUE updates random orderings and nearest-neighbor conditioning sets (only for vecchia = TRUE fits). In one-layer, there is no latent warping but the Vecchia approximation is still re-randomized and nearest-neighbors are adjusted accordingly. In two- and three-layers, the latest samples of hidden layers are used to update nearest-neighbors. If you update the Vecchia approximation, you should later remove previous samples (updating the approximation effectively starts a new chain). When re_approx = FALSE the previous orderings and conditioning sets are used (maintaining the continuity of the previous chain).

Value

object of the same class with the new iterations appended

Examples

# See ?fit_two_layer for an example

deepgp documentation built on Sept. 11, 2024, 8:30 p.m.