RNN之bi-LSTM双向递归神经网络

Halcom

RNN之bi-LSTM递归神经网络
百度网盘代码分享：http://pan.baidu.com/s/1eRQcCuu
电脑：Win7旗舰版+64Bit+AMD Athlon(tm)X2 DualCore QL-64 2.10GHz   RAM2.75GB
Anaconda3-4.2.0-Windows-x86_64

1. import time
   
2. #from tensorflow.examples.tutorials.mnist import input_data
   
3. import tensorflow as tf
   
4. import Get_Mnist_Data
   
5. 
   
6. start=time.clock()
   
7. #mnist = input_data.read_data_sets('/temp/', one_hot=True)
   
8. mnist = Get_Mnist_Data.read_data_sets('Get_Mnist_Data', one_hot=True)
   
9. end=time.clock()  
   
10. print('Runing time = %s Seconds'%(end-start))
    
11. 
    
12. def compute_accuracy(v_x, v_y):
    
13.     global pred
    
14.     #input v_x to nn and get the result with y_pre
    
15.     y_pre = sess.run(pred, feed_dict={x:v_x})
    
16.     #find how many right
    
17.     correct_prediction = tf.equal(tf.argmax(y_pre,1), tf.argmax(v_y,1))
    
18.     #calculate average
    
19.     accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
    
20.     #get input content
    
21.     result = sess.run(accuracy,feed_dict={x: v_x, y: v_y})
    
22.     return result
    
23. 
    
24. def Bi_lstm(X):
    
25.     lstm_f_cell = tf.contrib.rnn.BasicLSTMCell(n_hidden_units, forget_bias=1.0, state_is_tuple=True)
    
26.     lstm_b_cell = tf.contrib.rnn.BasicLSTMCell(n_hidden_units, forget_bias=1.0, state_is_tuple=True)
    
27.     return tf.contrib.rnn.static_bidirectional_rnn(lstm_f_cell, lstm_b_cell, X, dtype=tf.float32)
    
28. 
    
29. def RNN(X,weights,biases):
    
30.     # hidden layer for input
    
31.     X = tf.reshape(X, [-1, n_inputs])
    
32.     X_in = tf.matmul(X, weights['in']) + biases['in']
    
33. 
    
34.     #reshape data put into bi-lstm cell
    
35.     X_in = tf.reshape(X_in, [-1,n_steps, n_hidden_units])
    
36.     X_in = tf.transpose(X_in, [1,0,2])
    
37.     X_in = tf.reshape(X_in, [-1, n_hidden_units])
    
38.     X_in = tf.split(X_in, n_steps)
    
39.     outputs, _, _ = Bi_lstm(X_in)
    
40.    
    
41.     #hidden layer for output as the final results
    
42.     results = tf.matmul(outputs[-1], weights['out']) + biases['out']
    
43. 
    
44.     return results
    
45.    
    
46. 
    
47. #load mnist data
    
48. #mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
    
49. 
    
50. # parameters init
    
51. l_r = 0.001
    
52. training_iters = 100
    
53. batch_size = 128
    
54. 
    
55. n_inputs = 28
    
56. n_steps = 28
    
57. n_hidden_units = 128
    
58. n_classes = 10
    
59. 
    
60. #define placeholder for input
    
61. x = tf.placeholder(tf.float32, [None, n_steps, n_inputs])
    
62. y = tf.placeholder(tf.float32, [None, n_classes])
    
63. 
    
64. # define w and b
    
65. weights = {
    
66.     'in': tf.Variable(tf.random_normal([n_inputs,n_hidden_units])),
    
67.     'out': tf.Variable(tf.random_normal([2*n_hidden_units,n_classes]))
    
68. }
    
69. biases = {
    
70.     'in': tf.Variable(tf.constant(0.1,shape=[n_hidden_units,])),
    
71.     'out': tf.Variable(tf.constant(0.1,shape=[n_classes,]))
    
72. }
    
73. 
    
74. pred = RNN(x, weights, biases)
    
75. cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred,labels=y))
    
76. train_op = tf.train.AdamOptimizer(l_r).minimize(cost)
    
77. 
    
78. correct_pred = tf.equal(tf.argmax(pred,1),tf.argmax(y,1))
    
79. accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
    
80. 
    
81. #init session
    
82. sess = tf.Session()
    
83. #init all variables
    
84. sess.run(tf.global_variables_initializer())
    
85. #start training
    
86. 
    
87. # x_image,x_label = mnist.test.next_batch(500)
    
88. # x_image = x_image.reshape([500, n_steps, n_inputs])
    
89. 
    
90. for i in range(training_iters):
    
91.     #get batch to learn easily
    
92.     batch_x, batch_y = mnist.train.next_batch(batch_size)
    
93.     batch_x = batch_x.reshape([batch_size, n_steps, n_inputs])
    
94.     sess.run(train_op,feed_dict={x: batch_x, y: batch_y})
    
95.     if i % 20 == 0:
    
96.         print(sess.run(accuracy,feed_dict={x: batch_x, y: batch_y,}))
    
97. test_data = mnist.test.images.reshape([-1, n_steps, n_inputs])
    
98. test_label = mnist.test.labels
    
99. #print("Testing Accuracy:", sess.run(accuracy, feed_dict={x: test_data, y: test_label}))
    
100. print("Testing Accuracy: ", compute_accuracy(test_data, test_label))
     
101. sess.close()

复制代码

原理分析：参考： RNN之LSTM递归神经网络

   以arrive、Wuhan、on三个单词为例，分别设为x1、x2、x3。

        LSTM网络的做法1是：y1由x1决定，y2由a1和x2决定，a1由x1得到，y3由a2和x3决定，a2由x1和x2得到。

   针对arrive、Wuhan、on三个单词例子，我们也可以反过来考虑。

        LSTM网络的做法2是：y3由x3决定，y2由a3和x2决定，a3由x3得到，y1由a2和x1决定，a2由x2和x3得到。

        LSTM网络的做法1和LSTM网络的做法2好像都有道理，那么将LSTM网络的做法1和LSTM网络的做法2结合起来，就引入了Bi-LSTM双向LSTM网络。具体如图所示。

Bi-LSTM神经网络