### Keras ###
required.packages <- c('keras', 'caret', 'ggplot2')
new.packages <- required.packages[!(required.packages %in% installed.packages()[,"Package"])]
if(length(new.packages)) install.packages(new.packages, repos='http://cran.us.r-project.org')

library(keras, quietly = TRUE) 
library(caret, quietly = TRUE)
library(ggplot2, quietly = TRUE)

neural.init <- function() {
  set.seed(503)
  Models <<- vector("list")
}

neural.train <- function(model, XY) {
  XY <- as.matrix(XY, ncol = ncol(XY)) 
  X <- XY[,-ncol(XY)]
  Y <- XY[,ncol(XY)]
  
  # convert target to binary
  Y <- ifelse(Y > 0, 1, 0)
  
  # build network architecture
  mod <- keras_model_sequential() 
  mod %>% 
    layer_dense(units = 20, activation = 'relu', input_shape = c(ncol(X))) %>% 
    layer_dropout(rate = 0.2) %>% 
    layer_dense(units = 20, activation = 'relu') %>%
    layer_dropout(rate = 0.2) %>%
    layer_dense(units = 20, activation = 'relu') %>%
    layer_dropout(rate = 0.2) %>%
    layer_dense(units = 1, activation = 'sigmoid')
  
  mod %>% compile(
    loss = 'binary_crossentropy',
    optimizer = optimizer_rmsprop(),
    metrics = c('accuracy')
  )
  
  history <- mod %>% fit(
    X, Y, 
    epochs = 150, batch_size = 1000, 
    validation_split = 0, shuffle = FALSE
  )
  
  Models[[model]] <<- serialize_model(mod, include_optimizer = TRUE)
  
}

neural.predict <- function(model, X) {
  X <- as.matrix(X, nrow=1)
  m <- Models[[model]]
  #pred <- m %>% predict_classes(t(X))
  pred <- m %>% predict_proba(t(X))
  return(pred)
}

neural.save <- function(name)
{
  save(Models,file=name)  
}

neural.load <- function(name)
{
  load(name)
  for(i in c(1:length(Models))) {
    Models[[i]] <<- unserialize_model(Models[[i]], custom_objects = NULL, compile = NULL)
  }
}

hp.tune <- function(dropout, num_epochs, filename) {
  neural.init()
  # load
  path <- "C:/Zorro/Data/"
  XY <<- read.csv(paste0(path, 'adv_69_export_dataEURUSD_L.csv'), header = F)
  XY <- data.frame(XY) # convert to data frame. some models will need a matrix, but df OK for caret
  
  # convert target to factor
  XY[, ncol(XY)] <- ifelse(XY[, ncol(XY)] > 0, 1, 0)
  XY[, ncol(XY)] <- as.factor(XY[, ncol(XY)])
  
  # split into training and test sets
  training.samples <- createDataPartition(XY[, ncol(XY)], p=0.75, list=FALSE)
  XY.train <- XY[training.samples, ]
  XY.test <- XY[-training.samples, ]
  
  X_train <- as.matrix(XY.train[,-ncol(XY)])
  Y_train <- as.matrix(XY.train[, ncol(XY)])
  X_test <- as.matrix(XY.test[, -ncol(XY)])
  Y_test <- as.matrix(XY.test[, ncol(XY)])
  
  # scale features
  scaler <- preProcess(X_train, method = c("center", "scale"))
  X_train <- predict(scaler, X_train)
  X_test <- predict(scaler, X_test)
  
  # build network architecture
  model <- keras_model_sequential() 
  model %>% 
    layer_dense(units = 20, activation = 'relu', input_shape = c(ncol(X_train))) %>% 
    layer_dropout(rate = dropout) %>% 
    layer_dense(units = 20, activation = 'relu') %>%
    layer_dropout(rate = dropout) %>%
    layer_dense(units = 20, activation = 'relu') %>%
    layer_dropout(rate = dropout) %>%
    layer_dense(units = 1, activation = 'sigmoid')
  
  model %>% compile(
    loss = 'binary_crossentropy',
    optimizer = optimizer_rmsprop(),
    metrics = c('accuracy')
  )
  
  history <- model %>% fit(
    X_train, Y_train, 
    epochs = num_epochs, batch_size = 1000, 
    validation_split = 0.33, shuffle = FALSE
  )
  
  train_val <- plot(history)
  png(paste0(path, filename, ".png"))
  print(train_val)
  dev.off()
  
}