p1_HierarchicalAttention_model_transformer.py 文件源码

def gru_single_step_sentence_level(self, Xt,
                                       h_t_minus_1):  # Xt:[batch_size, hidden_size*2]; h_t:[batch_size, hidden_size*2]
        """
        single step of gru for sentence level
        :param Xt:[batch_size, hidden_size*2]
        :param h_t_minus_1:[batch_size, hidden_size*2]
        :return:h_t:[batch_size,hidden_size]
        """
        # update gate: decides how much past information is kept and how much new information is added.
        z_t = tf.nn.sigmoid(tf.matmul(Xt, self.W_z_sentence) + tf.matmul(h_t_minus_1,
                                                                         self.U_z_sentence) + self.b_z_sentence)  # z_t:[batch_size,self.hidden_size]
        # reset gate: controls how much the past state contributes to the candidate state.
        r_t = tf.nn.sigmoid(tf.matmul(Xt, self.W_r_sentence) + tf.matmul(h_t_minus_1,
                                                                         self.U_r_sentence) + self.b_r_sentence)  # r_t:[batch_size,self.hidden_size]
        # candiate state h_t~
        h_t_candiate = tf.nn.tanh(tf.matmul(Xt, self.W_h_sentence) + r_t * (
            tf.matmul(h_t_minus_1, self.U_h_sentence)) + self.b_h_sentence)  # h_t_candiate:[batch_size,self.hidden_size]
        # new state: a linear combine of pervious hidden state and the current new state h_t~
        h_t = (1 - z_t) * h_t_minus_1 + z_t * h_t_candiate
        return h_t

    # forward gru for first level: word levels