def _prepare_sparse_dense_dropout_layers(self, name: Text,
drop_rate: float) -> None:
self._tf_layers[f"sparse_input_dropout.{name}"] = layers.SparseDropout(
rate=drop_rate)
self._tf_layers[
f"dense_input_dropout.{name}"] = tf.keras.layers.Dropout(
rate=drop_rate)
def __init__(self, dense_dim: List[int], model_dim: int, reg_lambda: float,
drop_rate: float):
super(InputLayer, self).__init__()
self.dense_layers = [
tf.keras.layers.Dense(i, activation='relu') for i in dense_dim
]
self.sparse_dropout_layer = layers.SparseDropout(drop_rate)
self.sparse_to_dense_layer = layers.DenseForSparse(
units=dense_dim[0], reg_lambda=reg_lambda)
self.output_layer = tf.keras.layers.Dense(model_dim, activation='relu')
def _prepare_layers_for_sparse_tensors(self, attribute: Text,
feature_type: Text,
config: Dict[Text, Any]) -> None:
"""Sets up sparse tensor pre-processing before combining with dense ones."""
# For optionally applying dropout to sparse tensors
if config[SPARSE_INPUT_DROPOUT]:
self._tf_layers[self.SPARSE_DROPOUT] = layers.SparseDropout(
rate=config[DROP_RATE])
# For converting sparse tensors to dense
self._tf_layers[self.SPARSE_TO_DENSE] = layers.DenseForSparse(
name=f"sparse_to_dense.{attribute}_{feature_type}",
units=config[DENSE_DIMENSION][attribute],
reg_lambda=config[REGULARIZATION_CONSTANT],
)
# For optionally apply dropout to sparse tensors after they're converted to
# dense tensors.
if config[DENSE_INPUT_DROPOUT]:
self._tf_layers[self.DENSE_DROPOUT] = tf.keras.layers.Dropout(
rate=config[DROP_RATE])