两层LSTM的使用

一层的lstm效果不是很好，使用两层的lstm，代码如下。

 with graph.as_default():
     inputs_ = tf.placeholder(tf.int32, [None, seq_len], name='inputs')
     labels_ = tf.placeholder(tf.int32, [None, 4], name='labels')
     keep_prob = tf.placeholder(tf.float32, name='keep_prob')

     embedding = tf.Variable(tf.random_uniform((n_words + 1, embed_size), -1, 1))
     embed = tf.nn.embedding_lookup(embedding, inputs_)
     ################################# R N N #################################
     def LSTM_with_drop(lstm_size, keep_prob):
         # Your basic LSTM cell
         lstm = tf.nn.rnn_cell.BasicLSTMCell(lstm_size, state_is_tuple=True)

         # Add dropout to the cell
         drop = tf.nn.rnn_cell.DropoutWrapper(lstm, output_keep_prob=keep_prob)
         return drop

     fw_cell = tf.nn.rnn_cell.MultiRNNCell( [LSTM_with_drop(lstm_size, keep_prob) for _ in range(lstm_layers)] )
     bw_cell = tf.nn.rnn_cell.MultiRNNCell( [LSTM_with_drop(lstm_size, keep_prob) for _ in range(lstm_layers)] )
     # Getting an initial state of all zeros
     fw_initial_state = fw_cell.zero_state(batch_size, tf.float32)
     bw_initial_state = bw_cell.zero_state(batch_size, tf.float32)

     outputs, final_state = tf.nn.bidirectional_dynamic_rnn(inputs=embed,
                                                            cell_fw=fw_cell,
                                                            cell_bw=fw_cell,
                                                            initial_state_fw=fw_initial_state,
                                                            initial_state_bw=bw_initial_state)
     state = tf.concat([outputs[0][:,-1], outputs[1][:,-1]], 1)
     ################################# R N N #################################