R/SoftMax.R

In frhl/neuralnetr: Neural Nets in R

#' @title SoftMax
#' @description Classic softmax that can be interpreted as a probability
#' distribution over n items
#'
#' For `SoftMax$class_fun()`, given the column
#' vector of class probabilities for each point (computed by Softmax), return
#' a vector of the classes (integers) with the highest probability for each point.
#'
#' @family activation
#' @export


SoftMax <- R6Class("SoftMax", inherit = ClassModule, list(

  #' @param Z a vector of pre-activations.
  #' @return probabilities of the given class
  forward = function(Z) {

    sums = apply(Z, 2, function(x) {
      sigma = sum(exp(x))
      return(exp(x)/sigma)
    })

    return(sums)

  },

  #' @param dLdZ a vector of gradients dLdZ.
  #' @return dLdZ
  backward = function(dLdZ) {
    return(dLdZ)
  },

  #' @description a classify function.
  #' @param Ypred a vector of predictions.
  #' @return the index for the vector of the
  #' predicted label.
  class_fun = function(Ypred){
    return(argmax(Ypred, 2)-1)
  }
  )
)