layers.py 文件源码

def __init__(self, states, inner_size, trainable=True, scope=None):
        """Initiailzes a new instance of the BahdanauAttention class.

        The attention mechanism implemented in this class is the one
        described by Bahdanau et al. here: https://arxiv.org/abs/1409.0473.
        The attention states and the query are projected to the attention
        inner size, then summed together and processed with a tanh and
        finally dot producted with an attention vector. All the operations
        are performed on a reference size, named as the attention size, which
        must be set during the initialization phase (with the `size` argument).

        Arguments:
          states: 3-D Tensor of shape [batch, timesteps, state] representing the
            states on which the attention scores will be computed; the third dimension
            of the tensor must be statically determined.
          inner_size: int representing the inner attention size;
          trainable: if True, variables will be trainable;
          scope: None, str or tf.VariableScope representing the variable scope
            of the layer which will be used to create all the needed variables.

        Raises:
          ValueError: if the last dimension of the `state` argument is not
            statically determined.
        """
        super(BahdanauAttention, self).__init__(trainable=trainable, scope=scope)
        self._states = states
        self._size = inner_size
        self._memory = None
        self._vector = None
        self._var_op_names = set()

        # check that the last dimension of the `states`
        # variable is fully defined.
        state_size = states.get_shape()[-1].value
        if state_size is None:
            raise ValueError('Last dimension of `states` must be defined, found %s'
                             % str(tf.shape(states)))
        self._state_size = state_size