R/dglm-util.r

In kaneplusplus/dlrm: Deep Versions of Generalized Linear Models (including Proportional Hazard)

#' @importFrom crayon yellow
make_model_matrix <- function(form, x) {
  ret <- model.matrix(form, x)
  ts <- terms(form, data = x)
  if (attributes(ts)$intercept == 0) {
    stop(yellow("No intercept specified. This should be done using the", 
                "use_bias parameter."))
  } else {
    ret <- ret[,-1]
  }
  ret
}

create_input_and_hidden_layers <- 
  function(x_train, hidden_layers, hidden_layers_activation, 
           hidden_layer_names,
           use_bias, 
           model_name) {
  
  model <- keras_model_sequential(name = model_name)

  for (i in seq_along(hidden_layers)) {
    if (i == 1) {
      model %>% layer_dense(input_shape = ncol(x_train),
                            units = hidden_layers[i],
                            activation = hidden_layers_activation[i],
                            name = paste("hidden_layer", i, sep = "_"),
                            use_bias = use_bias[i],
                            trainable = TRUE,
                            kernel_initializer = "random_normal")
    } else {
      model %>% layer_dense(units = hidden_layers[i],
                            activation = hidden_layers_activation[i],
                            name = paste("hidden_layer", i, sep = "_"),
                            use_bias = use_bias[i],
                            trainable = TRUE,
                            kernel_initializer = "random_normal")
    }
  }
  model
}