R/layer-creation.R

In davharris/mistnet: Stochastic Neural Networks

#' Define a layer's properties
#' 
#' This function produces a description of a \code{layer} in a \code{network}
#' 
#' @param nonlinearity a \code{nonlinearity} object describing the activation
#'  function used in this layer
#' @param size an integer describing the number of output neurons in this layer.
#' @param prior \code{prior} object for constraining the layer's weights.
#' @param ... additional arguments to be passed to or from other methods 
#'  (currently not used).
#' 
#' @seealso \code{\link{layer}}
#' @include prior.R
#' @include nonlinearity.R
#' @export
defineLayer = function(nonlinearity, size, prior, ...){
  assert_that(inherits(nonlinearity, "nonlinearity"))
  assert_that(inherits(prior, "prior"))
  assert_that(length(size) == 1L)
  list(
    nonlinearity = nonlinearity, 
    size = safe.as.integer(size), 
    prior = prior
  )
}

createLayer = function(
  n.inputs,
  n.outputs,
  prior,
  nonlinearity,
  updater,
  n.minibatch,
  n.importance.samples
){
  coef.updater = updater$copy()
  coef.updater$initialize(delta = matrix(0, nrow = n.inputs, ncol = n.outputs))
  
  bias.updater = updater$copy()
  bias.updater$initialize(delta = matrix(0, nrow = 1, ncol = n.outputs))
  
  out = layer$new(
    coef.dim = c(n.inputs, n.outputs),
    weights = matrix(0, nrow = n.inputs, ncol = n.outputs),
    biases = matrix(rep(0, n.outputs), nrow = 1),
    coef.updater = coef.updater,
    bias.updater = bias.updater,
    nonlinearity = nonlinearity,
    prior = prior
  )
  out$resetState(n.minibatch, n.importance.samples)
  
  out
}