model_components.py 文件源码

def task_specific_attention(inputs, output_size,
                            initializer=layers.xavier_initializer(),
                            activation_fn=tf.tanh, scope=None):
    """
    Performs task-specific attention reduction, using learned
    attention context vector (constant within task of interest).

    Args:
        inputs: Tensor of shape [batch_size, units, input_size]
            `input_size` must be static (known)
            `units` axis will be attended over (reduced from output)
            `batch_size` will be preserved
        output_size: Size of output's inner (feature) dimension

    Returns:
        outputs: Tensor of shape [batch_size, output_dim].
    """
    assert len(inputs.get_shape()) == 3 and inputs.get_shape(
    )[-1].value is not None

    with tf.variable_scope(scope or 'attention') as scope:
        attention_context_vector = tf.get_variable(name='attention_context_vector',
                                                   shape=[output_size],
                                                   initializer=initializer,
                                                   dtype=tf.float32)
        input_projection = layers.fully_connected(inputs, output_size,
                                                  activation_fn=activation_fn,
                                                  scope=scope)
        attention_weights = tf.nn.softmax(
            tf.multiply(input_projection, attention_context_vector)
        )

        weighted_projection = tf.multiply(input_projection, attention_weights)

        outputs = tf.reduce_sum(weighted_projection, axis=1)

        return outputs