LSTM Implementation in Python Keras

是 否 +2 论坛币

k人 参与回答

经管之家送您一份

应届毕业生专属福利!

求职就业群

赵安豆老师微信：zhaoandou666

经管之家联合CDA

送您一个全额奖学金名额~ !

立即领取

感谢您参与论坛问题回答

经管之家送您两个论坛币！

+2 论坛币

LSTM Implementation in Keras

David Hill

本帖隐藏的内容

LSTM Implementation in Keras.pdf (827.06 KB)

2016-7-25 00:20:31 上传

扫码加我 拉你入群

请注明：姓名-公司-职位

以便审核进群资格，未注明则拒绝

分享0 收藏4 回帖

关键词：Implementa implement python ATION ATI

1. from keras.models import Sequential
   
2. from keras.layers.core import TimeDistributedDense, Activation
   
3. from keras.layers.recurrent import LSTM
   
4. from keras.optimizers import RMSprop
   
5. import numpy as np
   
6. import random
   
7. #if you need help installing keras, cuda or running this code let me know:
   
8. #djhill715@gmail.com
   
9. 
   
10. # Hope is the thing with feathers, Emily Dickenson
    
11. print 'Processing Poem'
    
12. poem = 'Hope is the thing with feathers // That perches in the soul, // And sings the tune without the words, // And never stops at all, // // And sweetest in the gale is heard; // And sore must be the storm // That could abash the little bird // That kept so many warm. // // I\'ve heard it in the chillest land, // And on the strangest sea; // Yet, never, in extremity, // It asked a crumb of me.'
    
13. #break poem into words
    
14. poem = poem.split(' ')
    
15. 
    
16. #build a dictionary
    
17. wordList = []
    
18. for word in poem:
    
19.   wordList.append(word)
    
20. #list(set()) is slow for large datasets, use bloom filters for those
    
21. wordList = list(set(wordList))
    
22. #insert system start token into dictionary
    
23. wordList.insert(0, '#START#')
    
24. 
    
25. #convert words into one-hot tokens
    
26. wordtoix = {}
    
27. ixtoword = {}
    
28. for i, word in enumerate(wordList):
    
29.   wordtoix[word] = i
    
30.   ixtoword[i] = word
    
31. #construct the input numpy arrays for the model
    
32. print 'Building Training Data'
    
33. #array has shape of num_seq X num_time_steps X dimensionality of features
    
34. ins = np.zeros( (1, len(poem)+2, len(wordList)) )
    
35. gts = np.zeros_like(ins)
    
36. 
    
37. #encode the poem into the training sequences
    
38. #ins begins with start
    
39. ins[0, 0, wordtoix['#START#']] = 1
    
40. for t, word in enumerate(poem):
    
41.   #the ground truth at time t is the next word
    
42.   gts[0, t, wordtoix[word]] = 1
    
43.   #the input at time t+1 is the previous ground truth
    
44.   ins[0, t+1, wordtoix[word]] = 1
    
45. #ground truth ends with the end token (or reuse start token)
    
46. gts[:,-1,wordtoix['#START#']] = 1
    
47. 
    
48. #the complete model definition
    
49. #We include the first dense layer because we have one-hot tokens
    
50. #this improves the performance by allowing the model to project the one-hot tokens into a dense encoding
    
51. #where distances between encodings matter
    
52. print 'Building Model'
    
53. model = Sequential()
    
54. #time distributed dense is a fully-connected layer with weights shared across time. updates are averaged across time
    
55. model.add(TimeDistributedDense(input_dim = len(wordList), output_dim = 64))
    
56. #LSTM is a recurrent keras layer that unrolls itself to the number of timesteps
    
57. model.add(LSTM(input_dim = 64, output_dim = 64, return_sequences = True, forget_bias_init='one'))
    
58. #another time distributed dense to map the projection space to the output space
    
59. model.add(TimeDistributedDense(input_dim = 64, output_dim = len(wordList)))
    
60. #softmax to produce probabilities across our dictionary at each timestep
    
61. model.add(Activation('softmax'))
    
62. 
    
63. #keras models must be compiled into cuda C code
    
64. print 'Compiling Model'
    
65. #rms prop provides many benefits to LSTMs, including making the gradient norm more robust through time
    
66. rms = RMSprop(lr=.003)
    
67. model.compile(loss='categorical_crossentropy', optimizer=rms)
    
68. 
    
69. #the default training function is sufficient here
    
70. #keras also has a train on batch function for more control
    
71. model.fit(ins, gts, batch_size=1, nb_epoch=150, verbose=1)
    
72. 
    
73. #print the input for validation
    
74. line = []
    
75. print ''
    
76. print ' =======INPUT======='
    
77. print ''
    
78. for x in xrange(ins.shape[1]):
    
79.   #get the word using the dictionary
    
80.   word = ixtoword[np.argmax(ins[0,x])]
    
81.   if word == '//':
    
82.     print ' '.join(line)
    
83.     line = []
    
84.   else:
    
85.     line.append(word)
    
86. print ' '.join(line)
    
87. print ' '
    
88. 
    
89. #prepare a new array to store predictions, initialize with start token
    
90. ins = np.zeros( (1, len(poem)+2, len(wordList)) )
    
91. ins[:, 0, wordtoix['#START#']] = 1
    
92. #list to hold the output, also initialized with start token
    
93. line = ['#START#']
    
94. print ''
    
95. print '=====OUTPUT====='
    
96. print ''
    
97. for x in xrange(ins.shape[1]-1):
    
98.   #dereference the first and only batch of the predictions
    
99.   predictions = model.predict(ins, batch_size=1, verbose=0)[0]
    
100.   #compute the predicted dictionary index
     
101.   pred_ix = np.argmax(predictions[x])
     
102.   #get the word using the dictionary
     
103.   word = ixtoword[pred_ix]
     
104.   #also set the input for the next timestep to the predicted index
     
105.   ins[:, x+1, pred_ix] = 1
     
106.   #the remaining lines are just for pretty printing
     
107.   if word == '//':
     
108.     print ' '.join(line)
     
109.     line = []
     
110.   else:
     
111.     line.append(word)
     
112. print ' '.join(line)

复制代码

谢谢分享。

好资料

看看啊