Bidirectional LSTM，由兩個LSTMs上下疊加在 一起組成。輸出由這兩個LSTMs的隱藏層的狀態決定。

def bilstm(self,x):
        
        # 輸入的資料格式轉換
        # x.shape [batch_size, time_steps, input_size]
        x=tf.transpose(x,[1,0,2])
        fw_x = tf.reshape(x, [-1, self.n_input_text]) # step*batch, feature
        fw_x = tf.split(0, self.n_step_text, fw_x)
        
    
        with tf.variable_scope('bilstm_lt'):
            
            #定義Cell,單層LSTM
            lstm_fw_cell = rnn_cell.BasicLSTMCell(self.n_hidden_text, forget_bias=1.0, state_is_tuple=True)#前向的lstm cell
            lstm_bw_cell = rnn_cell.BasicLSTMCell(self.n_hidden_text, forget_bias=1.0, state_is_tuple=True)#反向的rnn cell
            
            #dropout
            lstm_fw_cell = rnn_cell.DropoutWrapper(cell=lstm_fw_cell, input_keep_prob=1.0, output_keep_prob=keep_prob)
            lstm_bw_cell = rnn_cell.DropoutWrapper(cell=lstm_bw_cell, input_keep_prob=1.0, output_keep_prob=keep_prob)
            
            #構建雙向的RNN網路
            with tf.variable_scope('fw_lt'):
                 (output_fw, state_fw) = rnn.rnn(lstm_fw_cell,fw_x,dtype=tf.float32)
                 t=tf.convert_to_tensor(output_fw)
                 print (t.get_shape().as_list())
            with tf.variable_scope('bw_lt'):
                bw_x = tf.reverse(x, [True,False,False])# reverse time dim
                bw_x = tf.reshape(bw_x, [-1, self.n_input_text])  # step*batch, feature
                bw_x = tf.split(0, self.n_step_text, bw_x)
                (output_bw, state_bw) = rnn.rnn(lstm_bw_cell,bw_x,dtype=tf.float32)
                
            # output_bw.shape = [timestep_size, batch_size, hidden_size]
            output_bw = tf.reverse(output_bw, [True,False,False])
            output = tf.concat(2,[output_fw, output_bw])#在第2個維度上，將output_fw, output_bw拼接
        return output#返回值：(outputs, output_states：最後一層隱藏層)