VGG16的10折交叉验证实现cifar10的分类（keras实现）

怼烎@ 2022-10-29 05:16 180阅读 0赞

1.  `#!/usr/bin/env python3`
2.  `# -*- coding: utf-8 -*-`
3.  `"""`
4.  `Created on Sun Sep 30 17:12:12 2018`
5.  
6.  
7.  `这是用keras搭建的vgg16网络`
8.  `这是很经典的cnn,在图像和时间序列分析方面有很多的应用`
9.  `@author: lg`
10. `"""`
11. `#################`
12. 
13. `import keras`
14. `from keras import regularizers`
15. `from keras.datasets import cifar10`
16. `from keras.models import Sequential`
17. `from keras.layers import Conv2D, MaxPooling2D, Dense, Dropout, Flatten, BatchNormalization`
18. `from keras.optimizers import SGD`
19. `import os`
20. `import argparse`
21. `import random`
22. `import numpy as np`
23. `from scipy.misc import imread, imresize, imsave`
24. `import pickle`
25. 
26. `from sklearn.model_selection import StratifiedKFold`
27. 
28. `parser = argparse.ArgumentParser()`
29. `parser.add_argument('--train_dir', default='./train/')`
30. `parser.add_argument('--test_dir', default='./test/')`
31. `parser.add_argument('--log_dir', default='./')`
32. `parser.add_argument('--batch_size', default=16)`
33. `parser.add_argument('--gpu', type=int, default=0)`
34. `args = parser.parse_args()`
35. 
36. `os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu)`
37. `type_list = ['cat', 'dog']`
38. 
39. 
40. `def vgg_16_net():`
41. `model = Sequential()`
42. `model.add(Conv2D(64, (3, 3), input_shape=(32, 32, 3), padding='same', activation='relu', name='conv1_block'))`
43. `model.add(Conv2D(64, (3, 3), activation='relu', padding='same', name='conv2_block'))`
44. `model.add(BatchNormalization())`
45. `model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))`
46. `model.add(Conv2D(128, (3, 3), activation='relu', padding='same', name='conv3_block'))`
47. `model.add(Conv2D(128, (3, 3), activation='relu', padding='same', name='conv4_block'))`
48. `model.add(BatchNormalization())`
49. `model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))`
50. `model.add(Conv2D(256, (3, 3), activation='relu', padding='same', name='conv5_block'))`
51. `model.add(Conv2D(256, (3, 3), activation='relu', padding='same', name='conv6_block'))`
52. `model.add(Conv2D(256, (1, 1), activation='relu', padding='same', name='conv7_block'))`
53. `model.add(BatchNormalization())`
54. `model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))`
55. `model.add(Conv2D(512, (3, 3), activation='relu', padding='same', name='conv8_block'))`
56. `model.add(Conv2D(512, (3, 3), activation='relu', padding='same', name='conv9_block'))`
57. `model.add(Conv2D(512, (1, 1), activation='relu', padding='same', name='conv10_block'))`
58. `model.add(BatchNormalization())`
59. `model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))`
60. `model.add(Conv2D(512, (3, 3), activation='relu', padding='same', name='conv11_block'))`
61. `model.add(Conv2D(512, (3, 3), activation='relu', padding='same', name='conv12_block'))`
62. `model.add(Conv2D(512, (1, 1), activation='relu', padding='same', name='conv13_block'))`
63. `model.add(BatchNormalization())`
64. `model.add(MaxPooling2D(pool_size=(2, 2), strides=(2, 2)))`
65. `model.add(Flatten())`
66. `model.add(Dense(2048, activation='relu'))`
67. `model.add(Dropout(0.5))`
68. `model.add(Dense(4096, activation='relu'))`
69. `model.add(Dropout(0.5))`
70. `model.add(Dense(10, activation='softmax'))`
71. `#model.add(Dense(1, activation='sigmoid'))`
72. `return model`
73. 
74. 
75. `def prepare_data():`
76. `file_dict1 = unpickle('F:/cifar10/cifar-10-batches-py/data_batch_1')`
77. `label = file_dict1[b'labels']`
78. `image = file_dict1[b'data']`
79. `print(type(image))`
80. `file_dict2 = unpickle('F:/cifar10/cifar-10-batches-py/data_batch_2')`
81. `label = label + file_dict2[b'labels']`
82. `image = np.vstack((image, file_dict2[b'data']))`
83. `file_dict3 = unpickle('F:/cifar10/cifar-10-batches-py/data_batch_3')`
84. `label = label + file_dict3[b'labels']`
85. `image = np.vstack((image, file_dict3[b'data']))`
86. `file_dict4 = unpickle('F:/cifar10/cifar-10-batches-py/data_batch_4')`
87. `label = label + file_dict4[b'labels']`
88. `image = np.vstack((image, file_dict4[b'data']))`
89. `file_dict5 = unpickle('F:/cifar10/cifar-10-batches-py/data_batch_5')`
90. `label = label + file_dict5[b'labels']`
91. `image = np.vstack((image, file_dict5[b'data']))`
92. `image = np.reshape(image/255, (-1, 32, 32, 3))`
93. `label = keras.utils.to_categorical(label, 10)`
94. `#seed = 7`
95. `#np.random.seed(seed)`
96. `#train_data, test_data, train_label, test_label = train_test_split(image, label, test_size=0.2, random_state=0)`
97. `#train_num = int(len(label) * 0.8 )`
98. `#train_data, train_label, test_data, test_label = image[0:train_num], label[0:train_num], image[train_num:], label[train_num:]`
99. `# (X_train, y_train), (X_test, y_test) = cifar10.load_data()`
100. `# train_data = np.reshape(X_train/255, (-1, 32, 32, 3))`
101. `# train_label = keras.utils.to_categorical(y_train, 10)`
102. `# test_data = np.reshape(X_test/255, (-1, 32, 32, 3))`
103. `# test_label = keras.utils.to_categorical(y_test, 10)`
104. `# return train_data, train_label, test_data, test_label`
105. `return image, label`
106. 
107. `def unpickle(file):`
108. `with open(file, 'rb') as fo:`
109. `file_dict = pickle.load(fo, encoding='bytes')`
110. `return file_dict`
111. 
112. 
113. `def train():`
114. `kfold = StratifiedKFold(n_splits=10, shuffle=True, random_state=7)`
115. `data, label = prepare_data()`
116. `index = 1`
117. `for train, test in kfold.split(data, label.argmax(1)):`
118. `model = vgg_16_net()`
119. `sgd = SGD(lr=0.001, decay=1e-8, momentum=0.9, nesterov=True)`
120. `model.compile(optimizer=sgd, loss='categorical_crossentropy', metrics=['accuracy'])`
121. `model.fit(data[train], label[train], validation_data=(data[test], label[test]), epochs=20, batch_size=32, shuffle=True)`
122. `#train_data, train_label = prepare_data()`
123. `#model.fit(train_data, train_label, batch_size=64, epochs=20, shuffle=True, validation_split=0.2)`
124. `model.save_weights(args.log_dir + 'model_' + str(index) + '.h5')`
125. `index = index + 1`
126. 
127. 
128. 
129. `if __name__ == '__main__':`
130. `try:`
131. `train()`
132. `#model.load_weights(args.log_dir + '/model.h5')`
133. `#predict(model)`
134. `except Exception as err:`
135. `print(err)`
136.

最开始没有使用交叉验证，但是测试集与验证集的准确率一直维持在50%~60%，基本属于盲猜系列。原因大概是数据量太多，进行随机划分时，测试数据的分类不是很均匀。所以采用了交叉验证的方式，最终测试集与训练集的准确率能够达到99%，应该是有点过拟合了，结果还是非常满意的。