R/fit_model.R

In LDNN: Longitudinal Data Neural Network

#' Fit the pre-defined Neural Network for Longitudinal Data 
#' 
#' @param model The model object produced by create_model().
#' @param ver ver=0 to show nothing, ver=1 to show animated progress bar, ver=2 to just mention the number of epoch during training. 
#' @param n_epoch The number of epochs to train the model.
#' @param bsize The batch size.
#' @param X1 Features as inputs of 1st LSTM.
#' @param X2 Features as inputs of 2nd LSTM.
#' @param X3 Features as inputs of 3rd LSTM.
#' @param X4 Features as inputs of 4th LSTM.
#' @param X5 Features as inputs of 5th LSTM.
#' @param X6 Features as inputs of 6th LSTM.
#' @param X7 Features as inputs of 7th LSTM.
#' @param X8 Features as inputs of 8th LSTM.
#' @param X9 Features as inputs of 9th LSTM.
#' @param X10 Features as inputs of 10th LSTM.
#' @param Xif The features to be concatenated with the outputs of the LSTMs.
#' @param y The target variable.
#' @return The fitted model.
#' @description Fit the created Neural Network model (Keras).
#' @examples
#' X1 <- matrix(runif(500*20), nrow=500, ncol=20)
#' X2 <- matrix(runif(500*24), nrow=500, ncol=24)
#' X3 <- matrix(runif(500*24), nrow=500, ncol=24)
#' X4 <- matrix(runif(500*24), nrow=500, ncol=24)
#' X5 <- matrix(runif(500*16), nrow=500, ncol=16)
#' X6 <- matrix(runif(500*16), nrow=500, ncol=16)
#' X7 <- matrix(runif(500*16), nrow=500, ncol=16)
#' X8 <- matrix(runif(500*16), nrow=500, ncol=16)
#' X9 <- matrix(runif(500*16), nrow=500, ncol=16)
#' X10 <- matrix(runif(500*15), nrow=500, ncol=15)
#' Xif <- matrix(runif(500*232), nrow=500, ncol=232)
#' y <- matrix(runif(500), nrow=500, ncol=1)
#' \dontrun{
#' fitted_model = fit_model(model,0,1,32,X1,X2,X3,X4,X5,X6,X7,X8,X9,X10,Xif,y)
#' }
#' # The functions require to have python installed
#' # As well as tensorflow, keras and reticulate package.
#' @import keras
#' @export fit_model
fit_model<-function(model, ver, n_epoch, bsize, X1, X2, X3, X4, X5, X6, X7, X8, X9, X10, Xif, y){
  checkpoint_path <- "checkpoints/cp.ckpt"
  
  # Create checkpoint callback
  cp_callback <- callback_model_checkpoint(
    monitor = "val_loss", 
    filepath = checkpoint_path,
    save_weights_only = TRUE,
    save_best_only = TRUE,
    verbose = 0 # not printing
  )
  
  trained_model <- model %>% fit(
    x = list(inp1 = X1,
             inp2 = X2,
             inp3 = X3,
             inp4 = X4,
             inp5 = X5,
             inp6 = X6,
             inp7 = X7,
             inp8 = X8,
             inp9 = X9,
             inp10 = X10,
             inpif = Xif), # sequence we're using for prediction 
    y = y, # sequence we're predicting
    batch_size = bsize, # how many samples to pass to our model at a time
    epochs = n_epoch, # how many times we'll look at the whole dataset
    validation_split = 0.1, # how much data to hold out for testing as we go along
    callbacks = list(cp_callback),  # pass callback to training
    verbose = ver) # printing during training
  
  fitted = model %>% load_model_weights_tf(filepath = checkpoint_path)
  
  return(fitted)
}