Function _is_keras_rnn_cell

tensorflow/python/ops/rnn.py:145–165

Check whether the cell is a Keras RNN cell. The Keras RNN cell accept the state as a list even the state is a single tensor, whereas the TF RNN cell does not wrap single state tensor in list. This behavior difference should be unified in future version. Args: rnn_cell: An RNN cell inst

(rnn_cell)

def _is_keras_rnn_cell(rnn_cell):
  """Check whether the cell is a Keras RNN cell.

  The Keras RNN cell accept the state as a list even the state is a single
  tensor, whereas the TF RNN cell does not wrap single state tensor in list.
  This behavior difference should be unified in future version.

  Args:
    rnn_cell: An RNN cell instance that either follow the Keras interface or TF
      RNN interface.

  Returns:
    Boolean, whether the cell is an Keras RNN cell.
  """
  # Cell type check is not strict enough since there are cells created by other
  # library like Deepmind that didn't inherit tf.nn.rnn_cell.RNNCell.
  # Keras cells never had zero_state method, which was from the original
  # interface from TF RNN cell.
  return (not isinstance(rnn_cell, rnn_cell_impl.RNNCell) and
          isinstance(rnn_cell, base_layer.Layer) and
          getattr(rnn_cell, "zero_state", None) is None)


# pylint: disable=unused-argument