Updated.

UofUEpiBio/PHS7045-advanced-programming#13 @gvegayon @chipmanj

DESCRIPTION

@@ -21,7 +21,8 @@ Imports:
     stats,
     Rmpfr,
     nnet,
-    pROC
+    pROC,
+    bench
 Suggests: 
     knitr,
     rmarkdown,

NAMESPACE

@@ -11,6 +11,7 @@ export(feed_forward)
 export(map)
 export(sigmoid)
 importFrom(Rmpfr,pmax)
+importFrom(bench,mark)
 importFrom(nnet,nnet)
 importFrom(pROC,roc)
 importFrom(stats,predict)

R/BackPropNN.R

@@ -10,11 +10,22 @@ NULL
 #' @importFrom nnet nnet
 NULL
 
+#' @importFrom bench mark
+NULL
+
 #' BackPropNN
 #'
 #' A collection of functions
 #'
-#' @description We add stuff up...
+#' @description This R-package will contain functions that will implement NN
+#'  training (via back-propagation) from scratch (using basic R packages).
+#'  The function will take data (X and Y) as an input and will ask user to specify
+#'  no. of input nodes, no. of hidden nodes, no. of output nodes, learning rate, and an
+#'  activation function to be used. The function will produce the NN model in terms
+#'  of matrices containing weights for each nodes. The user will be able to choose
+#'  activation function between ReLU and Sigmoid. The user will be able to compare
+#'  the performance of this R-package in comparison to existing R NN packages in
+#'  terms of accuracy and computational time.
 #'
 #' @docType package
 #' @name BackPropNN

R/back_propagation_training.R

@@ -13,21 +13,17 @@
 #' @return A list of class \code{BackPropNN_back_propagation_training}:
 #' @examples
 #' set.seed(100)
-#' i <- 2 # number of input nodes
-#' h <- 2 # number of hidden nodes
-#' o <- 1 # number of output nodes
-#' learning_rate <- 0.1 # The learning rate of the algorithm
-#' activation_func <- "sigmoid" # the activation function
 #' data <- data.frame(X1 = 1:100, X2 = 2:101, Y = sample(c(0,1), 100, replace=TRUE))
-#' nn_model <- back_propagation_training(i, h, o, learning_rate, activation_func, data)
+#' nn_model <- back_propagation_training(i=2, h=2, o=1, learning_rate=0.01,
+#' activation_func="sigmoid", data=data)
 #'
 #' @export
 back_propagation_training <- function(i, h, o, learning_rate, activation_func, data){
 
-  W_IH = matrix(0.01,h,i)
-  W_HO = matrix(0.01,o,h)
-  B_H = matrix(0.01,h,1)
-  B_O = matrix(0.01,o,1)
+  W_IH = matrix(0.01,nrow=h,ncol=i)
+  W_HO = matrix(0.01,nrow=o,ncol=h)
+  B_H = matrix(0.01,nrow=h,ncol=1)
+  B_O = matrix(0.01,nrow=o,ncol=1)
 
   X = as.matrix(data[1:ncol(data)-1])
   Y = as.matrix(data[,ncol(data)])
@@ -86,7 +82,7 @@ plot.BackPropNN_back_propagation_training <- function(x) {
   Y = as.matrix(data[,ncol(data)])
 
   nn_R <- nnet::nnet(X,Y,size=x$num_nodes[2], trace=FALSE)
-  nn_R_pred <- as.numeric(stats::predict(nn_R,X))
+  nn_R_pred <- as.numeric(stats::predict(nn_R,X, type="raw"))
 
   pROC::roc(data[,ncol(data)],feed_forward(data,x)$pred,
             plot=TRUE, print.auc=TRUE, main="ROC curve by BackPropNN")
@@ -114,7 +110,7 @@ print.BackPropNN_back_propagation_training <- function(x) {
   Y = as.matrix(data[,ncol(data)])
 
   nn_R <- nnet::nnet(X,Y,size=x$num_nodes[2], trace=FALSE)
-  nn_R_pred <- as.numeric(stats::predict(nn_R,X))
+  nn_R_pred <- as.numeric(stats::predict(nn_R,X, type="raw"))
   nn_R_mse <- mean((Y - nn_R_pred)^2)
 
   my_nn_mse <- mean((Y - feed_forward(x$input_data,x)$pred)^2)

vignettes/BackPropNN.Rmd

@@ -18,13 +18,13 @@ knitr::opts_chunk$set(
 library(BackPropNN)
 ```
 
-# Simulated data - OR data.
+# Simulated data - AND data.
 
 ```{r}
-num_obs <- 10000 # Number of observations
+num_obs <- 100000 # Number of observations
 X1 <- sample(c(0,1),num_obs, replace = TRUE)
 X2 <- sample(c(0,1),num_obs, replace = TRUE)
-Y <- ifelse(X1==0 & X2==0, 0, 1)
+Y <- ifelse(X1==1 & X2==1, 1, 0)
 data <- data.frame(X1,X2,Y)
 ```