Enable factor or logical features for tabnet about mlr3keras HOT 6 CLOSED

def get_columns():
  """Get the representations for all input columns."""

  columns = []
  if float_columns:
    columns += [tf.feature_column.numeric_column(ci) for ci in float_columns]
  if int_columns:
    columns += [tf.feature_column.numeric_column(ci) for ci in int_columns]
  if str_columns:
    # pylint: disable=g-complex-comprehension
    columns += [
        tf.feature_column.embedding_column(
            tf.feature_column.categorical_column_with_hash_bucket(
                ci, hash_bucket_size=int(3 * num)),
            dimension=1) for ci, num in zip(str_columns, str_nuniques)
    ]
  if bool_columns:
    # pylint: disable=g-complex-comprehension
    columns += [
        tf.feature_column.embedding_column(
            tf.feature_column.categorical_column_with_hash_bucket(
                ci, hash_bucket_size=3),
            dimension=1) for ci in bool_columns
    ]
  return columns

  library("reticulate")
  library("tensorflow")
  library("keras")
  # keras::install_keras(extra_packages = c("tensorflow-hub", "tabnet==0.1.4.1"))
  
  use_implementation("tensorflow")
  
  tabnet <- import("tabnet") #0.1.4.1
  
  col_names = c('Sepal.Length', 'Sepal.Width', 'Petal.Length', 'Species')
  
  feature_columns <- lapply(col_names, function(x) { 
    if(x == "Species") {
      tf$feature_column$indicator_column(
        tf$feature_column$categorical_column_with_vocabulary_list(
          x, levels(iris$Species))
        ) 
    } else {
      tf$feature_column$numeric_column(x)  
    }
  })
  
  # The trick is that the one hot categorical counts as multiple columns for num_features
  model = tabnet$TabNetRegressor(feature_columns, num_regressors=1, num_features = 3 + length(levels(iris$Species)),
                                 feature_dim=4, output_dim=4,
                                 num_decision_steps=2, relaxation_factor=1.0,
                                 sparsity_coefficient=1e-5, batch_momentum=0.98,
                                 virtual_batch_size=NULL, norm_type='group',
                                 num_groups=1)
  model %>% compile(
    loss='mean_squared_error',
    optimizer=optimizer_adam()
  )

  x <- lapply(col_names, function(x) { as.matrix(iris[x])})
  names(x) <- col_names
  
  y <- model.matrix(~ 0 + iris$Petal.Width)
  
  model %>%
    fit(x, y, epochs=100, verbose=2)