R/basic-probs.R

In bmihaljevic/bnclassify: Learning Discrete Bayesian Network Classifiers from Data

# Adds a list of matrices. 
# 
# @return A numeric matrix. Multiplied in log space.
sum_matrices <- function(matrices) {
  # Must have at least on member 
  stopifnot(length(matrices) > 0)
  # Check all are numeric matrices of same size with same colnames
  n <- nrow(matrices[[1]])
  nobs <- ncol(matrices[[1]])
  values <- colnames(matrices[[1]])
  valid <- vapply(matrices, valid_factor, n, nobs, values, 
                  FUN.VALUE = logical(1L))
  stopifnot(all(valid))
  sum <- Reduce('+', matrices)
  stopifnot(identical(colnames(sum), values))
  sum
}
valid_factor <- function(x, nrow, ncol, colnames) {
  identical(dim(x), c(nrow, ncol)) && is.numeric(x) && identical(colnames, colnames(x))
}  
#' Normalize log probabilities. 
#' 
#' Uses the log-sum-exp trick.
#' 
#' @references Murphy KP (2012). \emph{Machine learning: a probabilistic
#'   perspective}. The MIT Press. pp. 86-87.
#' @keywords internal
log_normalize <- function(lp) {
  stopifnot(is.matrix(lp))
  # Check p is matrix of log probs?(<= 0)
  # Normalize with log sum exp 
  log_probs <- lp - matrixStats::rowLogSumExps(lp)
  log_probs 
}
exponentiate_probs <- function(p) {
  p <- exp_sideeffect(p)
  # rowAnys Does not distinguish between NA and NaN. 
  nans <- which(matrixStats::rowAnys(p, value = NaN))
  if (length(nans) > 0) {
    p[nans, ] <- 1 / ncol(p)
  } 
  p
}
#' Returns \code{TRUE} is \code{x} is a valid probability distribution.
#' 
#'@keywords internal
are_pdists <- function(x) {
  stopifnot(is.matrix(x))  
  are_probs(x) && all(fast_equal(rowSums(x), 1))  
}
are_probs <- function(x) {
  !anyNA(x) && all(x >= 0) && all(x <= 1)
}