deepspat.Rd

% Generated by roxygen2: do not edit by hand
% Please edit documentation in R/deepspat_main.R
\name{deepspat}
\alias{deepspat}
\title{Deep compositional spatial model}
\usage{
deepspat(
  f,
  data,
  layers,
  method = c("VB", "ML"),
  par_init = initvars(),
  learn_rates = init_learn_rates(),
  MC = 10L,
  nsteps
)
}
\arguments{
\item{f}{formula identifying the dependent variable and the spatial inputs (RHS can only have one or two variables)}

\item{data}{data frame containing the required data}

\item{layers}{list containing the warping layers}

\item{method}{either 'ML' (for the SIWGP) or 'VB' (for the SDSP)}

\item{par_init}{list of initial parameter values. Call the function \code{initvars()} to see the structure of the list}

\item{learn_rates}{learning rates for the various quantities in the model. Call the function \code{init_learn_rates()} to see the structure of the list}

\item{MC}{number of MC samples when doing stochastic variational inference}

\item{nsteps}{number of steps when doing gradient descent times three (first time the weights are optimised, then the covariance-function parameters, then everything together)}
}
\value{
\code{deepspat} returns an object of class \code{deepspat} with the following items
\describe{
 \item{"Cost"}{The final value of the cost (NMLL for the SIWGP and the lower bound for the SDSP, plus a constant)}
 \item{"mupost_tf"}{Posterior means of the weights in the top layer as a \code{TensorFlow} object}
 \item{"Qpost_tf"}{Posterior precision of the weights in the top layer as a \code{TensorFlow} object}
 \item{"eta_tf"}{Estimated or posterior means of the weights in the warping layers as a list of \code{TensorFlow} objects}
 \item{"precy_tf"}{Precision of measurement error, as a \code{TensorFlow} object}
 \item{"sigma2eta_tf"}{Variance of the weights in the top layer, as a \code{TensorFlow} object}
 \item{"l_tf"}{Length scale used to construct the covariance matrix of the weights in the top layer, as a \code{TensorFlow} object}
 \item{"scalings"}{Minima and maxima used to scale the unscaled unit outputs for each layer, as a list of \code{TensorFlow} objects}
 \item{"method"}{Either 'ML' or 'VB'}
 \item{"nlayers"}{Number of layers in the model (including the top layer)}
 \item{"MC"}{Number of MC samples when doing stochastic variational inference}
 \item{"run"}{\code{TensorFlow} session for evaluating the \code{TensorFlow} objects}
 \item{"f"}{The formula used to construct the deepspat model}
 \item{"data"}{The data used to construct the deepspat model}
 \item{"negcost"}{Vector of costs after each gradient-descent evaluation}
 \item{"data_scale_mean"}{Empirical mean of the original data}
 \item{"data_scale_mean_tf"}{Empirical mean of the original data as a \code{TensorFlow} object}
 }
}
\description{
Constructs a deep compositional spatial model
}
\examples{
df <- data.frame(s = rnorm(100), z = rnorm(100))
layers <- c(AWU(r = 50, dim = 1L, grad = 200, lims = c(-0.5, 0.5)),
            bisquares1D(r = 50))
\dontrun{d <- deepspat(f = z ~ s - 1, data = df, layers = layers, method = "ML", nsteps = 100L)}
}