Attention.py 文件源码

def __init__(self, attention_units, memory, time_major=True):
        self.attention_units    = attention_units
        self.enc_units          = memory.get_shape()[-1].value

        if time_major:
            memory = tf.transpose(memory, perm=(1,0,2))

        self.enc_length = tf.shape(memory)[1]
        self.batch_size = tf.shape(memory)[0]

        self.memory = tf.reshape(memory, (tf.shape(memory)[0], self.enc_length, 1, self.enc_units))
        # pre-compute Uahj to minimize the computational cost
        with tf.variable_scope('attention'):
            Ua = tf.get_variable(name='Ua', shape=(1, 1, self.enc_units, self.attention_units),
                                 initializer=gaussian_initializer(mean=0.0, std=0.001))
        self.hidden_feats = tf.nn.conv2d(self.memory, Ua, [1,1,1,1], "SAME")