Pi And More 11 - R Script For Neuronal Net And Training

This is the R script used to create and train the neuronal net for the Python project. Please consider that the lines for the export of the neuron weights are disabled by the comment marker #.

##################################
#
#  COMMON SAMPLE DATA HANDLING
#
##################################
load_samples = function(filename, attrib)
{
 data = read.csv(filename, sep="\t")
 # Remove result column
 data=data[,-5]
 # Calculate all combination with count of measure points
 n = aggregate(x ~ run + subset, data = data, FUN=length)
 # Keep only measurement with 80 points
 n = n[n[,3] == 80,]
 # Create the training data matrix for non-matches
 dataset = matrix(ncol=80, nrow=0)
 # Fill up the training data matrix with the valid cases (n) of data and transpose it
 for (i in 1:dim(n)[1])
  dataset = rbind(dataset, t(data$x[(data$run == n[i, 1]) & (data$subset == n[i, 2])]))
 # add class column
 dataset = cbind(c(attrib), dataset)
 # tidy up
 remove(data)
 remove(n)
 
 return(dataset);
}

# Sigmoid function like in Python3 example
sigmoid = function(x)
{
  1 / (1 + exp(-x))
}

#test = load_samples("C://...//test.csv", -1);

neg = load_samples("C://...//neg.csv", -1);
pos = load_samples("C://...//pos.csv", 1);

alldata = as.data.frame(rbind(pos, neg))
colnames(alldata) = c("event", paste("n", rep(1:80), sep=""))

library(caTools)
set.seed(101)

# Create Split (any column is fine)
split = sample.split(alldata$event, SplitRatio = 0.90)

# Split based off of split Boolean Vector
train = subset(alldata, split == TRUE)
test = subset(alldata, split == FALSE)

feats <- names(alldata[,-1])

# Concatenate strings
f = paste(feats, collapse=' + ')
f = paste('event ~',f)
# Convert to formula
f <- as.formula(f)

library(neuralnet)
# Works good # nn <- neuralnet(f, train, hidden = c(20, 10, 4), linear.output = TRUE)
#nn <- neuralnet(f, train, hidden = c(20, 10, 4), act.fct = 'logistic')
nn <- neuralnet(f, train, hidden = c(40, 20, 4), act.fct = 'logistic')

# Compute Predictions off Test Set
pred <- compute(nn, test[2:81])
#pred <- compute(nn, test[1, 2:81])

pred$net.result = (pred$net.result > apply(pred$net.result, 2, mean))*1
pred$net.result[pred$net.result == 0] = -1

#pred$net.result <- sapply(pred$net.result, round, digits = 0)

# Check of accuracy by the confusion matrix.
table(test$event, pred$net.result)

# Try to plot the neuronal net ... try. ;)
# plot(nn, show.weights=FALSE, radius=0.05, x.entry=0.15, intercept=FALSE)

# neural net weights export
#write.csv(nn$weights[[1]][[1]][1:81,], file = "C://...//nnlayer1.csv", row.names = FALSE)
#write.csv(nn$weights[[1]][[2]][1:41,], file = "C://...//nnlayer2.csv", row.names = FALSE)
#write.csv(nn$weights[[1]][[3]][1:21,], file = "C://...//nnlayer3.csv", row.names = FALSE)
#write.csv(nn$weights[[1]][[4]][1:5,], file = "C://...//nnlayer4.csv", row.names = FALSE)

#nn$weights[[1]][[1]][1,]