p1_HierarchicalAttention_model_transformer.py 文件源码

def gru_single_step_word_level(self, Xt, h_t_minus_1):
        """
        single step of gru for word level
        :param Xt: Xt:[batch_size*num_sentences,embed_size]
        :param h_t_minus_1:[batch_size*num_sentences,embed_size]
        :return:
        """
        # 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) + tf.matmul(h_t_minus_1,
                                                                self.U_z) + self.b_z)  # z_t:[batch_size*num_sentences,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) + tf.matmul(h_t_minus_1,
                                                                self.U_r) + self.b_r)  # r_t:[batch_size*num_sentences,self.hidden_size]
        # candiate state h_t~
        h_t_candiate = tf.nn.tanh(tf.matmul(Xt, self.W_h) +r_t * (tf.matmul(h_t_minus_1, self.U_h)) + self.b_h)  # h_t_candiate:[batch_size*num_sentences,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  # h_t:[batch_size*num_sentences,hidden_size]
        return h_t