Class MultiRNNCell

caffe2/python/rnn_cell.py:906–1101

Multilayer RNN via the composition of RNNCell instance. It is the responsibility of calling code to ensure the compatibility of the successive layers in terms of input/output dimensiality, etc., and to ensure that their blobs do not have name conflicts, typically by creating th

class MultiRNNCell(RNNCell):
    '''
    Multilayer RNN via the composition of RNNCell instance.

    It is the responsibility of calling code to ensure the compatibility
    of the successive layers in terms of input/output dimensiality, etc.,
    and to ensure that their blobs do not have name conflicts, typically by
    creating the cells with names that specify layer number.

    Assumes first state (recurrent output) for each layer should be the input
    to the next layer.
    '''

    def __init__(self, cells, residual_output_layers=None, **kwargs):
        '''
        cells: list of RNNCell instances, from input to output side.

        name: string designating network component (for scoping)

        residual_output_layers: list of indices of layers whose input will
        be added elementwise to their output elementwise. (It is the
        responsibility of the client code to ensure shape compatibility.)
        Note that layer 0 (zero) cannot have residual output because of the
        timing of prepare_input().

        forward_only: used to construct inference-only network.
        '''
        super().__init__(**kwargs)
        self.cells = cells

        if residual_output_layers is None:
            self.residual_output_layers = []
        else:
            self.residual_output_layers = residual_output_layers

        output_index_per_layer = []
        base_index = 0
        for cell in self.cells:
            output_index_per_layer.append(
                base_index + cell.get_output_state_index(),
            )
            base_index += len(cell.get_state_names())

        self.output_connected_layers = []
        self.output_indices = []
        for i in range(len(self.cells) - 1):
            if (i + 1) in self.residual_output_layers:
                self.output_connected_layers.append(i)
                self.output_indices.append(output_index_per_layer[i])
            else:
                self.output_connected_layers = []
                self.output_indices = []
        self.output_connected_layers.append(len(self.cells) - 1)
        self.output_indices.append(output_index_per_layer[-1])

        self.state_names = []
        for i, cell in enumerate(self.cells):
            self.state_names.extend(