Refactor plot_n_important_coeffs, unexport others (#68)

* rewrite plot_n_important_coefficients as nn2poly method * add dontrun also to luz example * unexport plot_ functions * rename source vignettes to get proper syntax highlighting from editors * reorganize dependencies by type * rename vignette sources to prevent pkgdown from compiling them * Added example to plot.nn2poly and fixed bug when n=NULL * Added test for plot.nn2poly when n=NULL * Added details. * Reordered checks for keep_layers and vector in plot function as they needed to be sequential. Also fixed previous bug with n=NULL as dim instead of length had to be used. Added tests for those situations. * Removed the need to use tidytext::reorder_within * Removed the need to suggest tidytext by substituting scale_x_reordered. * Removed faceting when teher is a single polynomial output --------- Co-authored-by: moralapablo <[email protected]>
IBiDat · Jan 17, 2024 · 6a0584a · 6a0584a
1 parent 45ca38d
commit 6a0584a
Show file tree

Hide file tree

Showing 33 changed files with 680 additions and 561 deletions.
diff --git a/DESCRIPTION b/DESCRIPTION
@@ -23,25 +23,16 @@ Encoding: UTF-8
 LazyData: true
 Depends: R (>= 3.5.0)
 Imports:
+    Rcpp,
     generics,
     matrixStats,
-    pracma,
-    Rcpp
+    pracma
 Suggests:
-    cowplot,
-    ggplot2,
-    keras,
-    knitr,
-    luz,
-    patchwork,
-    reticulate,
-    rmarkdown,
-    R6,
-    tensorflow,
-    testthat (>= 3.0.0),
-    tidytext,
-    torch,
-    vdiffr
+    keras, tensorflow, reticulate,
+    luz, torch,
+    cowplot, ggplot2, patchwork,
+    testthat (>= 3.0.0), vdiffr,
+    knitr, rmarkdown
 LinkingTo:
     Rcpp,
     RcppArmadillo

diff --git a/NAMESPACE b/NAMESPACE
@@ -5,16 +5,12 @@ S3method(add_constraints,luz_module_generator)
 S3method(fit,nn2poly)
 S3method(nn2poly,default)
 S3method(nn2poly,list)
-S3method(plot_taylor_and_activation_potentials,default)
-S3method(plot_taylor_and_activation_potentials,list)
+S3method(plot,nn2poly)
 S3method(predict,nn2poly)
 export(add_constraints)
 export(fit)
 export(luz_model_sequential)
 export(nn2poly)
-export(plot_diagonal)
-export(plot_n_important_coeffs)
-export(plot_taylor_and_activation_potentials)
 importFrom(Rcpp,sourceCpp)
 importFrom(generics,fit)
 importFrom(stats,density)

diff --git a/R/helpers.R b/R/helpers.R
@@ -15,6 +15,7 @@
 #' @seealso [add_constraints()]
 #'
 #' @examples
+#' \dontrun{
 #' if (requireNamespace("luz", quietly=TRUE)) {
 #' # Create a NN using luz/torch as a sequential model
 #' # with 3 fully connected linear layers,
@@ -35,7 +36,7 @@
 #' # Check that the nn is of class nn_squential
 #' class(nn)
 #' }
-#'
+#' }
 #'
 #' @export
 luz_model_sequential <- function(...) {

diff --git a/R/nn2poly.R b/R/nn2poly.R
@@ -176,155 +176,3 @@ nn2poly.default <- function(object, ...) {
 
   nn2poly(object, ...)
 }
-
-#' Predict method for \code{nn2poly} objects.
-#'
-#' Predicted values obtained with a nn2poly object on given data.
-#'
-#' @inherit eval_poly
-#' @param object Object of class inheriting from 'nn2poly'.
-#' @param layers Vector containing the chosen layers from \code{object} to be
-#' evaluated. If set to \code{NULL}, all layers are computed. Default is set
-#' to \code{NULL}.
-#' @param ... 	Further arguments passed to or from other methods.
-#'
-#' @details
-#' Internally uses `eval_poly()` to obtain the predictions. However, this only
-#' works with a objects of class \code{nn2poly} while `eval_poly()` can be used
-#' with a manually created polynomial in list form.
-#'
-#' When \code{object} contains all the internal polynomials also, as given by
-#' \code{nn2poly(object, keep_layers = TRUE)}, it is important to note that there
-#' are two polynomial items per layer (input/output). These polynomial items will
-#' also contain several polynomials of the same structure, one per neuron in the
-#' layer, stored as matrix rows in \code{$values}. Please see the NN2Poly
-#' original paper for more details.
-#'
-#' Note also that "linear" layers will contain the same input and output results
-#' as Taylor expansion is not used and thus the polynomials are also the same.
-#' Because of this, in the situation of evaluating multiple layers we provide
-#' the final layer with "input" and "output" even if they are the same, for
-#' consistency.
-#'
-#' @seealso [nn2poly()]: function that obtains the \code{nn2poly} polynomial
-#' object, [eval_poly()]: function that can evaluate polynomials in general,
-#' [stats::predict()]: generic predict function.
-#'
-#' @return Returns a matrix or list of matrices with the evaluation of each
-#' polynomial at each layer as given by the provided \code{object} of class
-#' \code{nn2poly}.
-#'
-#' If \code{object} contains the polynomials of the last layer, as given by
-#' \code{nn2poly(object, keep_layers = FALSE)}, then the output is a matrix with
-#' the evaluation of each data point on each polynomial. In this matrix, each
-#' column represents the evaluation of a polynomial and each column corresponds
-#' to each point in the new data to be evaluated.
-#'
-#' If \code{object} contains all the internal polynomials also, as given by
-#' \code{nn2poly(object, keep_layers = TRUE)}, then the output is a list of
-#' layers (represented by \code{layer_i}), where each one is another list with
-#' \code{input} and \code{output} elements, where each one contains a matrix
-#' with the evaluation of the "input" or "output" polynomial at the given layer,
-#' as explained in the case without internal polynomials.
-#'
-#'
-#' @examples
-#' # Build a NN structure with random weights, with 2 (+ bias) inputs,
-#' # 4 (+bias) neurons in the first hidden layer with "tanh" activation
-#' # function, 4 (+bias) neurons in the second hidden layer with "softplus",
-#' # and 1 "linear" output unit
-#'
-#' weights_layer_1 <- matrix(rnorm(12), nrow = 3, ncol = 4)
-#' weights_layer_2 <- matrix(rnorm(20), nrow = 5, ncol = 4)
-#' weights_layer_3 <- matrix(rnorm(5), nrow = 5, ncol = 1)
-#'
-#' # Set it as a list with activation functions as names
-#' nn_object = list("tanh" = weights_layer_1,
-#'                  "softplus" = weights_layer_2,
-#'                  "linear" = weights_layer_3)
-#'
-#' # Obtain the polynomial representation (order = 3) of that neural network
-#' final_poly <- nn2poly(nn_object, max_order = 3)
-#'
-#' # Define some new data, it can be vector, matrix or dataframe
-#' newdata <- matrix(rnorm(10), ncol = 2, nrow = 5)
-#'
-#' # Predict using the obtained polynomial
-#' predict(object = final_poly, newdata = newdata)
-#'
-#' # Change the last layer to have 3 outputs (as in a multiclass classification)
-#' # problem
-#' weights_layer_4 <- matrix(rnorm(20), nrow = 5, ncol = 4)
-#'
-#' # Set it as a list with activation functions as names
-#' nn_object = list("tanh" = weights_layer_1,
-#'                  "softplus" = weights_layer_2,
-#'                  "linear" = weights_layer_4)
-#'
-#' # Obtain the polynomial representation of that neural network
-#' # Polynomial representation of each hidden neuron is given by
-#' final_poly <- nn2poly(nn_object, max_order = 3, keep_layers = TRUE)
-#'
-#' # Define some new data, it can be vector, matrix or dataframe
-#' newdata <- matrix(rnorm(10), ncol = 2, nrow = 5)
-#'
-#' # Predict using the obtained polynomials (for all layers)
-#' predict(object = final_poly, newdata = newdata)
-#'
-#' # Predict using the obtained polynomials (for chosen layers)
-#' predict(object = final_poly, newdata = newdata, layers = c(2,3))
-#'
-#' @export
-predict.nn2poly <- function(object, newdata, layers = NULL, ...) {
-  if (length(class(object)) > 1)
-    return(NextMethod())
-
-  # Check if object is a single polynomial or a list of polynomials.
-  # If we get only the output layer, then it has to be a list with 2 elements,
-  # values and labels. We check one of them:
-  if (!is.null(object$labels)){
-    # If we have a final polynomial, directly evaluate the results:
-    result <- eval_poly(poly = object, newdata = newdata)
-  } else {
-    # Multiple layer case:
-
-    # If layer = NULL, set all layers to be used
-    if (is.null(layers)){
-      layers <- 1:length(object)
-    }
-
-    # Check if a vector or number is given
-    if (!(is.atomic(layers) & is.numeric(layers))){
-      stop("Argument layers is neither a numeric vector nor NULL.",
-           call. = FALSE
-      )
-    }
-    # Check that selected layers are within object dimension
-    # To do so, we need to check if "layer_maxvalue" exists:
-    final_layer <- paste0("layer_", max(layers))
-    if (is.null(object[[final_layer]])){
-      stop("Argument layers contains elements that exceed number of layers in nn2poly object.",
-           call. = FALSE
-      )
-    }
-
-    # Make sure layers are ordered, just for consistent output
-    layers <- sort(layers)
-
-    # Compute results for the given layers.
-    result <- list()
-    for (i in layers){
-      layer_name <- paste0("layer_", i)
-      result[[layer_name]] <- list()
-      result[[layer_name]][["input"]] <- eval_poly(
-        poly = object[[layer_name]][["input"]],
-        newdata = newdata)
-      result[[layer_name]][["output"]] <- eval_poly(
-        poly = object[[layer_name]][["output"]],
-        newdata = newdata)
-    }
-  }
-
-return(result)
-
-}