ops.py 文件源码

def attention(decoder_output, seq_outputs, output_size, time_steps,
              name="attention"):

    with tf.variable_scope(name):
        ui = []
        w_1 = tf.get_variable("w1", [output_size, output_size],
                              tf.float32,
                              tf.contrib.layers.xavier_initializer())
        w_2 = tf.get_variable("w2", [output_size, output_size],
                              tf.float32,
                              tf.contrib.layers.xavier_initializer())
        v = tf.get_variable("v", [output_size, 1],
                              tf.float32,
                              tf.contrib.layers.xavier_initializer())
        for seq_out in seq_outputs:
            ui.append(tf.matmul(tf.nn.tanh(tf.matmul(seq_out, w_1) +
                                    tf.matmul(decoder_output, w_2)), v))

        return ui