R/fit.ppred_model.R

In processpredictR: Process Prediction

#' @title Default fit function for ProcessTransformer model
#' @param object Object of class [ppred_model]: A transformer model
#' @param train_data A training dataset
#' @param batch_size A batch size
#' @param num_epochs  A number of epochs
#' @param verbose A verbose
#' @param callbacks [`list`]: A list of callbacks. `keras` default is `NULL`, but can be adjusted (ex. keras::callback_csv_logger(filename = paste("log_", object$task)), #or NULL
#' keras::callback_tensorboard())
#' @param shuffle [`logical`] (default [`TRUE`]): If [`TRUE`] shuffles the data
#' @param validation_split A ratio to split on
#' @param ... Additional Arguments
#'
#' @seealso See [keras::fit()] for documentation of parameters
#'
#' @importFrom keras fit
#' @export
keras::fit

#' @export
fit.ppred_model <- function(object,
                            train_data,
                            batch_size = 10,
                            epochs = 10,
                            verbose = 1,
                            callbacks = NULL,
                            shuffle = TRUE,
                            validation_split = 0.2,
                            ...) {

  if (object$task == "remaining_trace_s2s") {

    train_data_list <- train_data %>% prep_remaining_trace2()
    keras::fit(object$model, list(train_data_list$current_tokens, train_data_list$remaining_tokens),
               train_data_list$remaining_tokens_shifted,
               batch_size = batch_size,
               epochs = epochs,
               verbose = verbose,
               callbacks = callbacks,
               shuffle = shuffle,
               validation_split = validation_split,
               ...)
  }

  else {

    # preparation -------------------------------------------------------------
    # parameters
    maxlen <- object$max_case_length %>% as.integer()
    num_epochs <- epochs %>% as.integer()
    batch_size <- batch_size %>% as.integer()

    # tokenizing
    tokens_train <- tokenize(train_data)

    # preparring x_train, y_train for keras::fit()
    x_token_train <- tokens_train$token_x %>% keras::pad_sequences(maxlen = maxlen, value = 0)
    #x_token_train <- x_token_train #%>% reticulate::np_array(dtype = "float32")
    x_numeric_features <- tokens_train$numeric_features
    x_categorical_features <- tokens_train$categorical_features

    x_train_list <- list(x_token_train)
    if (!is.null(x_numeric_features)) x_train_list <- x_train_list %>% append(list(x_numeric_features))
    if (!is.null(x_categorical_features)) x_train_list <- x_train_list %>% append(list(x_categorical_features))
    y_token_train <- tokens_train$token_y #%>% reticulate::np_array(dtype = "float32")

    if (object$task %in% c("next_time", "remaining_time")) {
      #######################    #######################    #######################    #######################
      # y_token_train <- object$y_normalize_layer(y_token_train)
      y_token_train <- y_token_train / object$sd_time
      #######################    #######################    #######################    #######################

      # normalize_y <- keras::layer_normalization(mean = as.double(object$y_normalize_layer$mean),
      #                                           variance = as.double(object$y_normalize_layer$variance))
      #normalize_y %>% adapt(y_token_train)
      #y_token_train <- normalize_y(y_token_train)
    }
    # keras::fit --------------------------------------------------------------
    #if (object$number_features > 0) {
    keras::fit(object$model,
               #x = list(x_token_train, x_features_train),
               x = x_train_list,
               y = y_token_train,
               batch_size = batch_size,
               epochs = epochs,
               verbose = verbose,
               callbacks = callbacks,
               shuffle = shuffle,
               validation_split = validation_split,
               ...
    )
  }
}