一层的lstm效果不是很好，使用两层的lstm，代码如下。
1 with graph.as_default(): 2 inputs_ = tf.placeholder(tf.int32, [None, seq_len], name='inputs') 3 labels_ = tf.placeholder(tf.int32, [None, 4], name='labels') 4 keep_prob = tf.placeholder(tf.float32, name='keep_prob') 5 6 embedding = tf.Variable(tf.random_uniform((n_words + 1, embed_size), -1, 1)) 7 embed = tf.nn.embedding_lookup(embedding, inputs_) 8 ################################# R N N ################################# 9 def LSTM_with_drop(lstm_size, keep_prob): 10 # Your basic LSTM cell 11 lstm = tf.nn.rnn_cell.BasicLSTMCell(lstm_size, state_is_tuple=True) 12 13 # Add dropout to the cell 14 drop = tf.nn.rnn_cell.DropoutWrapper(lstm, output_keep_prob=keep_prob) 15 return drop 16 17 fw_cell = tf.nn.rnn_cell.MultiRNNCell( [LSTM_with_drop(lstm_size, keep_prob) for _ in range(lstm_layers)] ) 18 bw_cell = tf.nn.rnn_cell.MultiRNNCell( [LSTM_with_drop(lstm_size, keep_prob) for _ in range(lstm_layers)] ) 19 # Getting an initial state of all zeros 20 fw_initial_state = fw_cell.zero_state(batch_size, tf.float32) 21 bw_initial_state = bw_cell.zero_state(batch_size, tf.float32) 22 23 outputs, final_state = tf.nn.bidirectional_dynamic_rnn(inputs=embed, 24 cell_fw=fw_cell, 25 cell_bw=fw_cell, 26 initial_state_fw=fw_initial_state, 27 initial_state_bw=bw_initial_state) 28 state = tf.concat([outputs[0][:,-1], outputs[1][:,-1]], 1) 29 ################################# R N N #################################