R/define_last_layer.R
In neuroimaginador/dl4ni: Deep Learning Flows for Neuroimaging
#' @title Define the Last Layer of a Model according to Problem Type
#'
#' @description This function uses the information about a problem to define
#' a possible last layer.
#'
#' @param info (environment) Problem information as returned by \code{\link{get_problem_info}}.
#' @param units (integer) Number of output units
#' @param only_convolutionals (logical) If \code{TRUE}, just returns convolutional layers, otherwise dense layers, Default: FALSE
#' @param force_categorical (logical) When only 2 classes in a classification problem, one can use a \code{softsign} regression or a categorical layer, Default: TRUE
#' @param multioutput (logical) Split the output units, each one with a loss function, Default: TRUE
#' @param ... arguments passed to other functions.
#'
#' @return A list with:
#' - \code{last_layer}, The definition of the last layer,
#' - \code{output_activation}, The output activation function,
#' - \code{loss}, The loss function(s) to use for each output,
#' - \code{remap_classes}, A mapping from the class numbers in the original images and 1:num_classes,
#' - \code{binarise}, A logical indicating if there are only 2 classes and not using a categorical output,
#' - \code{num_classes}, The number of classes, in a classification problem,
#' - \code{y_label}, The set of labels in the original output image in a classification problem,
#' - \code{multioutput}, A logical to indicate if the output is split.
#'
#' @export
#'
define_last_layer <- function(info,
units,
only_convolutionals = FALSE,
force_categorical = TRUE,
multioutput = TRUE, ...) {
# Basic input class checking
stopifnot(inherits(info, "DLproblem"))
# Initialize variables
binarise <- FALSE
num_classes <- 0
extra_args <- list(...)
last_layer_info <- switch(info$type,
"image_labelling" = {
# If it's a classification problem
# Get the number of classes and their labels
num_classes <- length(info$values)
loss_function <- keras::loss_categorical_crossentropy
if (num_classes == 1)
loss_function <- keras::loss_binary_crossentropy
if ("loss_function" %in% names(extra_args)) {
loss_function <- extra_args$loss_function
}
if (is.null(info$subsetted)) {
y_label <- info$values
remap_classes <- info$remap_classes
} else {
if (!info$subsetted) {
y_label <- info$values
remap_classes <- list(source = info$values,
target = seq_along(info$values))
} else {
y_label <- info$original_values
remap_classes <- info$remap_classes
}
}
if (only_convolutionals) {
output_activation <- "softmax"
last_layer <- conv3d(filters = num_classes + 1,
kernel_size = c(3, 3, 3),
padding = "same",
activation = output_activation,
batch_normalization = FALSE)
multioutput <- FALSE
loss <- loss_function
} else {
# If more than 2 classes, only option is a categorical layer as output.
if (num_classes > 1) {
# If multioutput (split the output in different units)
if (multioutput) {
# Categorical layer with categorical_crossentropy as loss
last_layer <- categorical(num_classes = num_classes + 1,
units = units, ...)
loss <- list()
for (i in seq(units)) {
loss[[i]] <- loss_function
}
} else {
# Categorical layer with mean_squared_error as loss
last_layer <- categorical(num_classes = num_classes + 1,
units = units,
concatenate = TRUE, ...)
loss <- keras::loss_mean_squared_error
}
output_activation <- "sigmoid"
} else {
# Only two classes. We can force a categorical layer or a
# simple dense layer with softsign activation.
if (force_categorical) {
# As before, but with binary_crossentropy
if (multioutput) {
last_layer <- categorical(num_classes = num_classes + 1,
units = units, ...)
loss <- list()
for (i in seq(units)) {
loss[[i]] <- loss_function
}
} else {
last_layer <- categorical(num_classes = num_classes + 1,
units = units,
concatenate = TRUE, ...)
loss <- keras::loss_mean_squared_error
}
output_activation <- "sigmoid"
} else {
# 2 classes mapped to -1, +1. Use softsign.
last_layer <- dense(units = units)
output_activation <- "softsign"
binarise <- TRUE
loss <- keras::loss_mean_squared_error
}
}
}
list(last_layer = last_layer,
output_activation = output_activation,
loss = loss,
remap_classes = remap_classes,
binarise = binarise,
num_classes = num_classes,
num_volumes = info$num_volumes,
y_label = y_label,
multioutput = multioutput,
categorize_output = TRUE)
},
"image_regression" = {
# If the problem is not classification, it is regression.
# Thus, output activation should be linear.
# No classes, like in communism.
output_activation <- "relu"
multioutput <- FALSE
remap_classes <- NULL
y_label <- NULL
binarise <- FALSE
# Regression layer.
if (only_convolutionals) {
last_layer <- conv3d(filters = 1,
kernel_size = c(3, 3, 3),
padding = "same",
activation = output_activation,
batch_normalization = FALSE)
} else {
last_layer <- regression(units = units,
output_activation = output_activation,
...)
}
loss <- keras::loss_mean_squared_error
list(last_layer = last_layer,
output_activation = output_activation,
loss = loss,
remap_classes = remap_classes,
binarise = binarise,
num_classes = num_classes,
num_volumes = info$num_volumes,
y_label = y_label,
multioutput = multioutput,
categorize_output = FALSE)
},
"subject_classification" = {
},
"subject_regression" = {
})
return(last_layer_info)
}
neuroimaginador/dl4ni documentation built on May 3, 2019, 5:47 p.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
#' @title Define the Last Layer of a Model according to Problem Type
#'
#' @description This function uses the information about a problem to define
#' a possible last layer.
#'
#' @param info (environment) Problem information as returned by \code{\link{get_problem_info}}.
#' @param units (integer) Number of output units
#' @param only_convolutionals (logical) If \code{TRUE}, just returns convolutional layers, otherwise dense layers, Default: FALSE
#' @param force_categorical (logical) When only 2 classes in a classification problem, one can use a \code{softsign} regression or a categorical layer, Default: TRUE
#' @param multioutput (logical) Split the output units, each one with a loss function, Default: TRUE
#' @param ... arguments passed to other functions.
#'
#' @return A list with:
#' - \code{last_layer}, The definition of the last layer,
#' - \code{output_activation}, The output activation function,
#' - \code{loss}, The loss function(s) to use for each output,
#' - \code{remap_classes}, A mapping from the class numbers in the original images and 1:num_classes,
#' - \code{binarise}, A logical indicating if there are only 2 classes and not using a categorical output,
#' - \code{num_classes}, The number of classes, in a classification problem,
#' - \code{y_label}, The set of labels in the original output image in a classification problem,
#' - \code{multioutput}, A logical to indicate if the output is split.
#'
#' @export
#'
define_last_layer <- function(info,
units,
only_convolutionals = FALSE,
force_categorical = TRUE,
multioutput = TRUE, ...) {
# Basic input class checking
stopifnot(inherits(info, "DLproblem"))
# Initialize variables
binarise <- FALSE
num_classes <- 0
extra_args <- list(...)
last_layer_info <- switch(info$type,
"image_labelling" = {
# If it's a classification problem
# Get the number of classes and their labels
num_classes <- length(info$values)
loss_function <- keras::loss_categorical_crossentropy
if (num_classes == 1)
loss_function <- keras::loss_binary_crossentropy
if ("loss_function" %in% names(extra_args)) {
loss_function <- extra_args$loss_function
}
if (is.null(info$subsetted)) {
y_label <- info$values
remap_classes <- info$remap_classes
} else {
if (!info$subsetted) {
y_label <- info$values
remap_classes <- list(source = info$values,
target = seq_along(info$values))
} else {
y_label <- info$original_values
remap_classes <- info$remap_classes
}
}
if (only_convolutionals) {
output_activation <- "softmax"
last_layer <- conv3d(filters = num_classes + 1,
kernel_size = c(3, 3, 3),
padding = "same",
activation = output_activation,
batch_normalization = FALSE)
multioutput <- FALSE
loss <- loss_function
} else {
# If more than 2 classes, only option is a categorical layer as output.
if (num_classes > 1) {
# If multioutput (split the output in different units)
if (multioutput) {
# Categorical layer with categorical_crossentropy as loss
last_layer <- categorical(num_classes = num_classes + 1,
units = units, ...)
loss <- list()
for (i in seq(units)) {
loss[[i]] <- loss_function
}
} else {
# Categorical layer with mean_squared_error as loss
last_layer <- categorical(num_classes = num_classes + 1,
units = units,
concatenate = TRUE, ...)
loss <- keras::loss_mean_squared_error
}
output_activation <- "sigmoid"
} else {
# Only two classes. We can force a categorical layer or a
# simple dense layer with softsign activation.
if (force_categorical) {
# As before, but with binary_crossentropy
if (multioutput) {
last_layer <- categorical(num_classes = num_classes + 1,
units = units, ...)
loss <- list()
for (i in seq(units)) {
loss[[i]] <- loss_function
}
} else {
last_layer <- categorical(num_classes = num_classes + 1,
units = units,
concatenate = TRUE, ...)
loss <- keras::loss_mean_squared_error
}
output_activation <- "sigmoid"
} else {
# 2 classes mapped to -1, +1. Use softsign.
last_layer <- dense(units = units)
output_activation <- "softsign"
binarise <- TRUE
loss <- keras::loss_mean_squared_error
}
}
}
list(last_layer = last_layer,
output_activation = output_activation,
loss = loss,
remap_classes = remap_classes,
binarise = binarise,
num_classes = num_classes,
num_volumes = info$num_volumes,
y_label = y_label,
multioutput = multioutput,
categorize_output = TRUE)
},
"image_regression" = {
# If the problem is not classification, it is regression.
# Thus, output activation should be linear.
# No classes, like in communism.
output_activation <- "relu"
multioutput <- FALSE
remap_classes <- NULL
y_label <- NULL
binarise <- FALSE
# Regression layer.
if (only_convolutionals) {
last_layer <- conv3d(filters = 1,
kernel_size = c(3, 3, 3),
padding = "same",
activation = output_activation,
batch_normalization = FALSE)
} else {
last_layer <- regression(units = units,
output_activation = output_activation,
...)
}
loss <- keras::loss_mean_squared_error
list(last_layer = last_layer,
output_activation = output_activation,
loss = loss,
remap_classes = remap_classes,
binarise = binarise,
num_classes = num_classes,
num_volumes = info$num_volumes,
y_label = y_label,
multioutput = multioutput,
categorize_output = FALSE)
},
"subject_classification" = {
},
"subject_regression" = {
})
return(last_layer_info)
}
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.