tf.contrib.rnn.static_rnn与tf.nn.dynamic_rnn区别

https://blog.csdn.net/u014365862/article/details/78238807

MachineLP的Github（欢迎follow）：https://github.com/MachineLP

我的GitHub：https://github.com/MachineLP/train_cnn-rnn-attention 自己搭建的一个框架，包含模型有：vgg(vgg16,vgg19), resnet(resnet_v2_50,resnet_v2_101,resnet_v2_152), inception_v4, inception_resnet_v2等。

1. chunk_size = 256

2.  

   chunk_n = 160

3.  

   rnn_size = 256

4.  

   num_layers = 2

5.  

   n_output_layer = MAX_CAPTCHA*CHAR_SET_LEN # 输出层

单层rnn：

tf.contrib.rnn.static_rnn：

输入：[步长,batch,input]

输出：[n_steps,batch,n_hidden]

还有rnn中加dropout

1. def recurrent_neural_network(data):

2.  

3.  

   data = tf.reshape(data, [-1, chunk_n, chunk_size])

4.  

   data = tf.transpose(data, [1,0,2])

5.  

   data = tf.reshape(data, [-1, chunk_size])

6.  

   data = tf.split(data,chunk_n)

7.  

8.  

   # 只用RNN

9.  

   layer = {‘w_‘:tf.Variable(tf.random_normal([rnn_size, n_output_layer])), ‘b_‘:tf.Variable(tf.random_normal([n_output_layer]))}

10.  

    lstm_cell = tf.contrib.rnn.BasicLSTMCell(rnn_size)

11.  

    outputs, status = tf.contrib.rnn.static_rnn(lstm_cell, data, dtype=tf.float32)

12.  

    # outputs = tf.transpose(outputs, [1,0,2])

13.  

    # outputs = tf.reshape(outputs, [-1, chunk_n*rnn_size])

14.  

    ouput = tf.add(tf.matmul(outputs[-1], layer[‘w_‘]), layer[‘b_‘])

15.  

16.  

    return ouput

多层rnn：

tf.nn.dynamic_rnn：

输入：[batch,步长,input] 
输出：[batch,n_steps,n_hidden] 
所以我们需要tf.transpose(outputs, [1, 0, 2])，这样就可以取到最后一步的output

1. def recurrent_neural_network(data):

2.  

   # [batch,chunk_n,input]

3.  

   data = tf.reshape(data, [-1, chunk_n, chunk_size])

4.  

   #data = tf.transpose(data, [1,0,2])

5.  

   #data = tf.reshape(data, [-1, chunk_size])

6.  

   #data = tf.split(data,chunk_n)

7.  

8.  

   # 只用RNN

9.  

   layer = {‘w_‘:tf.Variable(tf.random_normal([rnn_size, n_output_layer])), ‘b_‘:tf.Variable(tf.random_normal([n_output_layer]))}

10.  

    #1

11.  

    # lstm_cell1 = tf.contrib.rnn.BasicLSTMCell(rnn_size)

12.  

    # outputs1, status1 = tf.contrib.rnn.static_rnn(lstm_cell1, data, dtype=tf.float32)

13.  

14.  

    def lstm_cell():

15.  

    return tf.contrib.rnn.LSTMCell(rnn_size)

16.  

    def attn_cell():

17.  

    return tf.contrib.rnn.DropoutWrapper(lstm_cell(), output_keep_prob=keep_prob)

18.  

    # stack = tf.contrib.rnn.MultiRNNCell([attn_cell() for _ in range(0, num_layers)], state_is_tuple=True)

19.  

    stack = tf.contrib.rnn.MultiRNNCell([lstm_cell() for _ in range(0, num_layers)], state_is_tuple=True)

20.  

    # outputs, _ = tf.nn.dynamic_rnn(stack, data, seq_len, dtype=tf.float32)

21.  

    outputs, _ = tf.nn.dynamic_rnn(stack, data, dtype=tf.float32)

22.  

    # [batch,chunk_n,rnn_size] -> [chunk_n,batch,rnn_size]

23.  

    outputs = tf.transpose(outputs, (1, 0, 2))

24.  

25.  

    ouput = tf.add(tf.matmul(outputs[-1], layer[‘w_‘]), layer[‘b_‘])

26.  

27.  

    return ouput

原文地址：https://www.cnblogs.com/DjangoBlog/p/9542571.html