R/2-layers-stereo.R

In tpq/caress: Caress, a Gentler Introduction to Keras

#' Split Model into Two Parallel Layers
#'
#' This function splits the input into two parallel layers.
#'  Each layer has the same architecture, but is influenced
#'  by different weights.
#'
#' @param object A \code{keras} model.
#' @param name A string. The prefix label for all layers.
#' @param units,activation,... Arguments passed to \code{keras::layer_dense}.
#' @return A list of two layers: Channel 1 and Channel 2.
#' @export
layer_to_dense_stereo <- function(object, units, activation = NULL, name = NULL, ...){

  # Name layer based on incoming data
  if(is.null(name)){
    name <- get_incoming_layer_name(object)
  }

  channel1 <- object %>% keras::layer_dense(units = units, activation = activation,
                                            name = paste0(name, "_channel", 1), ...)
  channel2 <- object %>% keras::layer_dense(units = units, activation = activation,
                                            name = paste0(name, "_channel", 2), ...)

  return(list("channel1" = channel1,
              "channel2" = channel2))
}

#' Split Model into Two Parallel Layers
#'
#' This function splits the input into three parallel layers.
#'  The first two layers have the same architecture, but are influenced
#'  by different weights. The third layer is the concatenation of
#'  the other two layers.
#'
#' @inheritParams layer_to_dense_stereo
#' @return A list of three layers: Channel 1, Channel 2, and
#'  Channel 1 & Channel 2.
#' @export
layer_to_dense_stereo_and_cat <- function(object, units, activation = NULL, name = NULL, ...){

  # Name layer based on incoming data
  if(is.null(name)){
    name <- get_incoming_layer_name(object)
  }

  channel1 <- object %>% keras::layer_dense(units = units, activation = activation,
                                            name = paste0(name, "_channel", 1), ...)
  channel2 <- object %>% keras::layer_dense(units = units, activation = activation,
                                            name = paste0(name, "_channel", 2), ...)
  channel12 <- keras::layer_concatenate(list(channel1, channel2),
                                        name = paste0(name, "_channel", 12))

  return(list("channel1" = channel1, "channel2" = channel2,
              "channel12" = channel12))
}

#' Split Model into Two Parallel Layers
#'
#' This function splits the input into three parallel layers.
#'  The first two layers have the same architecture, but are influenced
#'  by different weights. The third layer is the sum of
#'  the other two layers.
#'
#' @inheritParams layer_to_dense_stereo
#' @return A list of three layers: Channel 1, Channel 2, and
#'  Channel 1 & Channel 2.
#' @export
layer_to_dense_stereo_and_add <- function(object, units, activation = NULL, name = NULL, ...){

  # Name layer based on incoming data
  if(is.null(name)){
    name <- get_incoming_layer_name(object)
  }

  channel1 <- object %>% keras::layer_dense(units = units, activation = activation,
                                            name = paste0(name, "_channel", 1), ...)
  channel2 <- object %>% keras::layer_dense(units = units, activation = activation,
                                            name = paste0(name, "_channel", 2), ...)
  channel12 <- keras::layer_add(list(channel1, channel2),
                                name = paste0(name, "_channel", 12))

  return(list("channel1" = channel1, "channel2" = channel2,
              "channel12" = channel12))
}

#' Split Model into Two Parallel Layers
#'
#' This function splits the input into three parallel layers.
#'  The first two layers have the same architecture, but are influenced
#'  by different weights. The third layer is the difference of
#'  the other two layers.
#'
#' @inheritParams layer_to_dense_stereo
#' @return A list of three layers: Channel 1, Channel 2, and
#'  Channel 1 & Channel 2.
#' @export
layer_to_dense_stereo_and_diff <- function(object, units, activation = NULL, name = NULL, ...){

  # Name layer based on incoming data
  if(is.null(name)){
    name <- get_incoming_layer_name(object)
  }

  channel1 <- object %>% keras::layer_dense(units = units, activation = activation,
                                            name = paste0(name, "_channel", 1), ...)
  channel2 <- object %>% keras::layer_dense(units = units, activation = activation,
                                            name = paste0(name, "_channel", 2), ...)
  channel12 <- keras::layer_subtract(list(channel1, channel2),
                                     name = paste0(name, "_channel", 12))

  return(list("channel1" = channel1, "channel2" = channel2,
              "channel12" = channel12))
}