boston nn-model

https://onesixx.com/global/

predict a continuous value (instead of a discrete label)

source(file.path(getwd(),"00.global.R"))

##1# DATA Source : INPUT LAYER -------------------------------------------------
#데이터 량이 적다.
#feature간 척도가 다르다.
# y : 주택가격 median (단위 1,000$)
dataset <- dataset_boston_housing()
c(c(trnData, trnTarget), c(tstData, tstTarget)) %<-% dataset 

melt_trnX <- data.table(dataset$train$x) %>% melt.data.table(measure=str_c("V",1:13))
melt_trnX %>% ggplot(aes(variable, value)) + geom_boxplot()

##2# Preprocess  ---------------------------------------------------------------
trnData %>% str()
tstData %>% str()
trnTarget %>% summary()

# *Normalize ----
trn_mean   <- apply(trnData, 2, mean)
trn_stdev  <- apply(trnData, 2, sd)
trnData <- scale(trnData, center=trn_mean, scale=trn_stdev)
tstData <- scale(tstData, center=trn_mean, scale=trn_stdev)   # test data!!!

melt_trnData <- data.table(trnData) %>% melt.data.table(measure=str_c("V",1:13))
melt_trnData %>% ggplot(aes(variable, value)) + geom_boxplot()

##3# Build the model -----------------------------------------------------------
uF_buildModel <- function(){
\tmodel <- keras_model_sequential() %>% 
\t\tlayer_dense(units=64, activation="relu", input_shape=dim(trnData)[[2]]) %>% 
\t\tlayer_dense(units=64, activation="relu") %>% 
\t\tlayer_dense(units=1)                              # pure linear
\t
\tmodel %>% compile(
\t\toptimizer = "rmsprop",
\t\tloss="mse",
\t\tmetrics=c("mae")
\t)
}

##4# Train the model -----------------------------------------------------------
history <- model %>% fit(trnData, trnTarget, 
\t\t\t\t\t\t\t\t\t\t\t\t batch_size=10, epoches=30, validation_split=.2)
#history %>% uF_histPlot()
#history %>% plot()

##5# Evaluate accuracy ---------------------------------------------------------
eval <- model %>% evaluate(tstData, tstTarget)

##6# Make predictions ----------------------------------------------------------
rlt <- model %>% predict(tstData)
#data.table(cbind(tstTarget, rlt)) %>% ggplot(aes(tstTarget, rlt)) + geom_point() + geom_abline()


# K-fold Cross Validation ------------------------------------------------------
k <- 4
indices <- sample(1:nrow(trnData))
folds <- cut(1:length(indices), breaks=k, labels=F)

num_epochs <- 100
scores <- c()
for (i in 1:k){
\t#i=1
\tcat("processing fold #", i, "\
")
\t
\t# k 부분 데이터 : 검증 데이터 준비 
\tval_indices <- which(folds==i, arr.ind=T)
  val_trnData   <-   trnData[val_indices, ] 
\tval_trnTarget <- trnTarget[val_indices]
\t
\t# 나머지 데이터 : 훈련 데이터 준비 
\tpartial_trnData   <-   trnData[-val_indices, ]
\tpartial_trnTarget <- trnTarget[-val_indices]
\t
\t# Training
\tmodel <- uF_buildModel()\t
\tmodel %>% fit(partial_trnData, partial_trnTarget, 
\t\t\t\t\t\t\t\tepochs=num_epochs, batch_size=1, verbose=0)
\t# evaluate
\teval <- model %>% evaluate(val_trnData, val_trnTarget, verbose=0)
\tscores <- c(scores, eval$mean_absolute_error)
}

scores
scores %>% mean()