a1_seq2seq_attention_model.py 文件源码

def gru_cell_decoder(self, Xt, h_t_minus_1,context_vector):
        """
        single step of gru for word level
        :param Xt: Xt:[batch_size,embed_size]
        :param h_t_minus_1:[batch_size,embed_size]
        :param context_vector. [batch_size,embed_size].this represent the result from attention( weighted sum of input during current decoding step)
        :return:
        """
        # 1.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_decoder) + tf.matmul(h_t_minus_1,self.U_z_decoder) +tf.matmul(context_vector,self.C_z_decoder)+self.b_z_decoder)  # z_t:[batch_size,self.hidden_size]
        # 2.reset gate: controls how much the past state contributes to the candidate state.
        r_t = tf.nn.sigmoid(tf.matmul(Xt, self.W_r_decoder) + tf.matmul(h_t_minus_1,self.U_r_decoder) +tf.matmul(context_vector,self.C_r_decoder)+self.b_r_decoder)  # r_t:[batch_size,self.hidden_size]
        # candiate state h_t~
        h_t_candiate = tf.nn.tanh(tf.matmul(Xt, self.W_h_decoder) +r_t * (tf.matmul(h_t_minus_1, self.U_h_decoder)) +tf.matmul(context_vector, self.C_h_decoder)+ self.b_h_decoder)  # 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  # h_t:[batch_size*num_sentences,hidden_size]
        return h_t,h_t

    # forward gru for first level: word levels