tensorflow.nn.bidirectional_dynamic_rnn()函数的用法
开门见山来两张比较蛋疼的图，它们确实很流行。直奔主题。def bidirectional_dynamic_rnn(
cell_fw, # 前向RNN
cell_bw, # 后向RNN
inputs, # 输入
sequence_length=None,# 输入序列的实际长度（可选，默认为输入序列的最大长度）
initial_state_fw=None,# 前向的初始化状态（可选）
initial_state_bw=None,# 后向的初始化状态（可选）
dtype=None, # 初始化和输出的数据类型（可选）
parallel_iterations=None,
swap_memory=False,
time_major=False,
# 决定了输入输出tensor的格式：如果为true, 向量的形状必须为 ``.
# 如果为false, tensor的形状必须为``.
scope=None
)
outputs为(output_fw, output_bw)，是一个包含前向cell输出tensor和后向cell输出tensor组成的二元组。假设 time_major=false, 而且tensor的shape为。实验中使用tf.concat(outputs, 2)将其拼接。
output_states为(output_state_fw, output_state_bw)，包含了前向和后向最后的隐藏状态的组成的二元组。
output_state_fw和output_state_bw的类型为LSTMStateTuple。
LSTMStateTuple由（c，h）组成，分别代表memory cell和hidden state。
LSTM应用到双向RNN中
而cell_fw和cell_bw的定义是完全一样的。如果这两个cell选LSTM cell整个结构就是双向LSTM了。
# lstm模型正方向传播的RNN
lstm_fw_cell = tf.nn.rnn_cell.BasicLSTMCell(embedding_size, forget_bias=1.0)
# 反方向传播的RNN
lstm_bw_cell = tf.nn.rnn_cell.BasicLSTMCell(embedding_size, forget_bias=1.0)
但是看来看去，输入两个cell都是相同的啊？
其实在bidirectional_dynamic_rnn函数的内部，会把反向传播的cell使用array_ops.reverse_sequence的函数将输入的序列逆序排列，使其可以达到反向传播的效果。
在实现的时候，我们是需要传入两个cell作为参数就可以了：
(outputs, output_states) = tf.nn.bidirectional_dynamic_rnn(lstm_fw_cell, lstm_bw_cell, embedded_chars,dtype=tf.float32)
embedded_chars为输入的tensor，。batch_size为模型当中batch的大小，应用在文本中时，max_time可以为句子的长度（一般以最长的句子为准，短句需要做padding），depth为输入句子词向量的维度。
代码实践：
import tensorflow as tf
import numpy as np
X = np.random.randn(2, 10, 8)
# The second example is of length 6
X = 0
X_lengths =
cell = tf.nn.rnn_cell.LSTMCell(num_units=5, state_is_tuple=True)
outputs, states = tf.nn.bidirectional_dynamic_rnn(
cell_fw=cell, cell_bw=cell, dtype=tf.float64, sequence_length=X_lengths, inputs=X
)
output_fw, output_bw = outputs
states_fw, states_bw = states
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
states_shape = tf.shape(states)
print(states_shape.eval())
c_f, h_f = states_fw
o_f = output_fw
c_b, h_b= states_bw
o_b = output_bw
print('c_f\n', sess.run(c_f))
print('h_f\n', sess.run(h_f))
print('o_f\n', sess.run(o_f))
print('c_b\n', sess.run(c_b))
print('h_b\n', sess.run(h_b))
print('o_b\n', sess.run(o_b))
输出结果：
c_f
[[-0.43276965 -0.34707254 -0.091809970.268278320.27571178]
[ 0.275752240.151569460.12256522 -0.1233779-0.09387333]]
h_f
[[-0.29557532 -0.19821126 -0.025424680.1287899 0.10906331]
[ 0.139096570.074858120.0607246-0.06372124 -0.04719312]]
o_f
[[[-0.086079120.19634355 -0.04141379 -0.09648713 -0.29296226]
[ 0.0920274 0.122123180.06549744 -0.41358432 -0.02210931]
[ 0.39993605 -0.036047450.38421408 -0.170964210.07381075]
[ 0.17104686 -0.085318270.042495910.053659380.1784615 ]
[ 0.00792906 -0.16713683 -0.021031820.075155170.06772459]
[-0.20100924 -0.355764890.161943110.194469140.25483659]
[-0.18140209 -0.08311345 -0.128168810.070987060.427926]
[-0.17574083 -0.14505373 -0.234014550.156315830.39293472]
[-0.29557532 -0.19821126 -0.025424680.1287899 0.10906331]
[ 0. 0. 0. 0. 0. ]]
[[ 0.09741206 -0.097792950.189188360.032787530.2577792 ]
[ 0.022673910.06850602 -0.0155975-0.23521581 -0.03577484]
[ 0.194296190.062763820.10905737 -0.15550532 -0.01645063]
[ 0.102875250.20157 -0.02434073 -0.114224280.00976497]
[-0.052279360.32488201 -0.06576368 -0.11532339 -0.13688021]
[ 0.225187370.105163090.12899814 -0.1449693-0.00556297]
[ 0.203237790.111705670.10008328 -0.08086347 -0.03259825]
[ 0.139096570.074858120.0607246-0.06372124 -0.04719312]
[ 0. 0. 0. 0. 0. ]
[ 0. 0. 0. 0. 0. ]]]
c_b
[[ 0.097388360.19225204 -0.09284249 -0.473824260.00350991]
[ 0.449975050.194477850.49119047 -0.442520460.31626763]]
h_b
[[ 0.040208060.07441591 -0.03619023 -0.0777202 0.00210421]
[ 0.138884610.098105570.15060079 -0.289644120.16514088]]
o_b
[[[ 0.040208060.07441591 -0.03619023 -0.0777202 0.00210421]
[ 0.30703784 -0.265129750.0314823 0.109289370.28692156]
[ 0.10526645 -0.238501170.076822610.282632130.21087581]
[-0.434327 -0.22145861 -0.215429020.3141704 0.31082225]
[-0.28739246 -0.20374412 -0.020411210.152770310.22083064]
[-0.39453516 -0.178251760.0045626 0.163922250.35356923]
[-0.16832858 -0.00360075 -0.180953530.044360010.35192945]
[-0.09912457 -0.126655070.006391660.12355956 -0.0580625 ]
[-0.11977808 -0.089575230.07406649 -0.00428107 -0.11181204]
[ 0. 0. 0. 0. 0. ]]
[[ 0.138884610.098105570.15060079 -0.289644120.16514088]
[ 0.079459910.16753371 -0.09477983 -0.27083062 -0.16861312]
[ 0.177035910.106701110.054833770.000543340.03132806]
[ 0.008024940.20236404 -0.12328111 -0.07817032 -0.00155747]
[ 0.078957850.134870850.03472546 -0.04419926 -0.03887194]
[ 0.26463148 -0.057146320.169547210.039670120.10644822]
[ 0. 0. 0. 0. 0. ]
[ 0. 0. 0. 0. 0. ]
[ 0. 0. 0. 0. 0. ]
[ 0. 0. 0. 0. 0. ]]]
页:
[1]