tensorflow学习笔记(三十九):双向rnn

如何在tensorflow中实现双向rnn

单层双向rnn

 

 

单层双向rnn (cs224d)

tensorflow中已经提供了双向rnn的接口,它就是tf.nn.bidirectional_dynamic_rnn(). 我们先来看一下这个接口怎么用.

 

 1 bidirectional_dynamic_rnn(
 2     cell_fw, #前向 rnn cell
 3     cell_bw, #反向 rnn cell
 4     inputs, #输入序列.
 5     sequence_length=None,# 序列长度
 6     initial_state_fw=None,#前向rnn_cell的初始状态
 7     initial_state_bw=None,#反向rnn_cell的初始状态
 8     dtype=None,#数据类型
 9     parallel_iterations=None,
10     swap_memory=False,
11     time_major=False,
12     scope=None
13 )

 

返回值:一个tuple(outputs, outputs_states), 其中,outputs是一个tuple(outputs_fw, outputs_bw). 关于outputs_fw和outputs_bw,如果time_major=True则它俩也是time_major的,vice versa. 如果想要concatenate的话,直接使用tf.concat(outputs, 2)即可.

如何使用: 
bidirectional_dynamic_rnn 在使用上和 dynamic_rn

n是非常相似的.

定义前向和反向rnn_cell
定义前向和反向rnn_cell的初始状态
准备好序列
调用bidirectional_dynamic_rnn
import tensorflow as tf
from tensorflow.contrib import rnn
cell_fw = rnn.LSTMCell(10)
cell_bw = rnn.LSTMCell(10)
initial_state_fw = cell_fw.zero_state(batch_size)
initial_state_bw = cell_bw.zero_state(batch_size)
seq = ...
seq_length = ...
(outputs, states)=tf.nn.bidirectional_dynamic_rnn(cell_fw, cell_bw, seq,
 seq_length, initial_state_fw,initial_state_bw)
out = tf.concat(outputs, 2)

View Code