library(class)
library(e1071)

source("data")
x <- subset(data, select = -Class)
y <- as.factor(data$Class)

pred <- rep(0, length(y))
ng <- length(unique(y))
np <- dim(x)
ncb <-ng * round((0.4*ng*(ng-1 + np[2]/2))/ng,0) #modified version of original suggestion from Kohonen, but here with$

numclasses = length(unique(y)) 
ncp = rep(1/numclasses, numclasses); 

rand <- sample(10, length(t(y)), replace = T)
pred <- rep(0,length(t(y)))
for(iter in sort(unique(rand)))
{
         set.seed(0) # note that LVQ in R is only reproducible given a fixed seed
         mod0 <- lvqinit(x[rand != iter,], y[rand != iter], ncb, prior=ncp)
         mod1 <- olvq1(x[rand != iter,], y[rand != iter], mod0)
         mod1 <- lvq3(x[rand != iter,], y[rand != iter], mod1)
         pred[rand == iter] <- lvqtest(mod1, x[rand == iter,])
}
cM <- table(true = data$Class, predicted = pred)
cMstats <- classAgreement(cM)


write(cM, 'confusionMatrix', ncolumns=1)
write(cMstats$diag, 'accuracy')
write(cMstats$kappa, 'kappa')
write(pred, 'Pred', ncolumns=1)