library(RSNNS)
library(data.table)

set.seed(1)

#CARGA DE LOS DATOS
trainSet <- data.table(read.csv("DatosTrain.csv",dec=",",sep=";",header = T))
validSet <- data.table(read.csv("DatosValid.csv",dec=",",sep=";",header = T))
testSet  <- data.table(read.csv( "DatosTest.csv",dec=",",sep=";",header = T))

#SELECCION DE LA SALIDA
target <- "cloudType"

#TRANSFORMAR LA SALIDA DISCRETA A NUMÉRICA (Matriz con 3 columnas, una por etiqueta única, hay un 1 por cada fila en la columna que pertenece a la clase)
trainTarget <- decodeClassLabels(unlist(trainSet[,target,with=F]))
validTarget <- decodeClassLabels(unlist(validSet[,target,with=F]))
testTarget <-  decodeClassLabels(unlist(testSet[,target,with=F]))

#SEPARAR ENTRADA DE LA SALIDA
trainInput <- trainSet[,-target,with=F]
validInput <- validSet[,-target,with=F]
testInput <-  testSet[,-target,with=F]

#SELECCION DE LOS PARAMETROS
topologia        <- c(50)
razonAprendizaje <- 0.01
ciclosMaximos    <- 1000

#EJECUCION DEL APRENDIZAJE Y GENERACION DEL MODELO
model <- mlp(x= trainInput,
             y= trainTarget,
             inputsTest= validInput,
             targetsTest= validInput,
             size= topologia,
             maxit=ciclosMaximos,
             learnFuncParams=c(razonAprendizaje),
             shufflePatterns = F
)

#GRAFICO DE LA EVOLUCION DEL ERROR
plotIterativeError(model)

#GENERAR LAS PREDICCIONES
trainPred <- predict(model,trainInput)
validPred <- predict(model,validInput)
testPred  <- predict(model,testInput)

#CALCULO DE LAS MATRICES DE CONFUSION
trainCm <- confusionMatrix(as.matrix(trainTarget),trainPred)
validCm <-  confusionMatrix(as.matrix(validTarget),validPred)
testCm <- confusionMatrix(as.matrix(testTarget), testPred)

#METRICAS DE ERROR
accuracy <- function (cm) sum(diag(cm))/sum(cm)
macroAvg <- function (cm) mean(diag(cm)/rowSums(cm))

#VECTOR DE PRECISIONES
accuracies <- c(TrainAccuracy= accuracy(trainCm), TrainMacroAvg= macroAvg(trainCm),
                ValidAccuracy= accuracy(validCm), ValidMacroAvg= macroAvg(validCm),
                TestAccuracy=  accuracy(testCm) , TestMacroAvg= macroAvg(testCm))

#TABLA CON LOS ERRORES POR CICLO
iterativeErrors <- data.table(RMSETrain= (model$IterativeFitError/nrow(trainSet)) ^(1/2),
                              RMSEValid= (model$IterativeTestError/nrow(validSet))^(1/2))

#UTILIDADES PARA GENERAR LAS TABLAS DE SALIDAS
targets <- unlist(unique(trainSet[,target,with=F]))
padMatrix <- function(m) rbind(m,matrix(rep(NA,(nrow(trainSet)-nrow(m))*ncol(trainPred)),ncol=ncol(trainPred)))
predictClass <- function(predMat,targets) as.character(unlist(data.table(predMat)[,targets[apply(.SD,1,which.max)]]))

#SALIDA DE LA RED EN BRUTO, SIN ELEGIR LA CLASE
rawOutputs <- data.table(train= trainPred,
                      valid= padMatrix(validPred),
                      test=  padMatrix(testPred))
names <- data.table(expand.grid(c("train","valid","test"),unlist(unique(trainSet[,target,with=F]))))
colnames(rawOutputs) <- apply(rbind(names[Var1=="train"],names[Var1=="valid"],names[Var1=="test"]),1,paste,collapse=".")

#SALIDA DE LA RED ETIQUETADA (Eleccion del maximo por fila y asignar etiqueta)
outputs <- data.table(train= predictClass(trainPred,targets),
                      valid= c(predictClass(validPred,targets),rep(NA,nrow(trainSet)-nrow(validSet))),
                      test=  c(predictClass(testPred,targets),rep(NA,nrow(trainSet)-nrow(testSet))))
colnames(outputs) <- c("train","validation","test")

#GUARDADO DE RESULTADOS
saveRDS(model,"nnet.rds")
write.csv2(accuracies,"finalAccuracies.csv")
write.csv2(iterativeErrors,"iterativeErrors.csv")
write.csv2(outputs,"netOutputs.csv")
write.csv2(rawOutputs,"netRawOutputs.csv")
write.csv2(trainCm,"trainCm.csv")
write.csv2(validCm,"validCm.csv")
write.csv2(testCm,"testCm.csv")