机器学习进阶笔记之四 | 深入理解GoogLeNet

GoogLeNet on Tensorflow

　　GoogLeNet为了实现方便，我用tflearn来重写了下，代码中和caffe model里面不一样的就是一些padding的位置，因为改的比较麻烦，必须保持inception部分的concat时要一致，我这里也不知道怎么修改pad的值（caffe prototxt），所以统一padding设定为same，具体代码如下：

1.  # -*- coding: utf-8 -*-
2. 
   
3.  """ GoogLeNet.
4.  Applying 'GoogLeNet' to Oxford's 17 Category Flower Dataset classification task.
5.  References:
6.      - Szegedy, Christian, et al.
7.      Going deeper with convolutions.
8.      - 17 Category Flower Dataset. Maria-Elena Nilsback and Andrew Zisserman.
9.  Links:
10.      - [GoogLeNet Paper](http://www.cv-foundation.org/openaccess/content_cvpr_2015/papers/Szegedy_Going_Deeper_With_2015_CVPR_paper.pdf)
11.      - [Flower Dataset (17)](http://www.robots.ox.ac.uk/~vgg/data/flowers/17/)
12.  """
13. 
    
14.  from __future__ import division, print_function, absolute_import
15. 
    
16.  import tflearn
17.  from tflearn.layers.core import input_data, dropout, fully_connected
18.  from tflearn.layers.conv import conv_2d, max_pool_2d, avg_pool_2d
19.  from tflearn.layers.normalization import local_response_normalization
20.  from tflearn.layers.merge_ops import merge
21.  from tflearn.layers.estimator import regression
22. 
    
23.  import tflearn.datasets.oxflower17 as oxflower17
24.  X, Y = oxflower17.load_data(one_hot=True, resize_pics=(227, 227))
25. 
    
26. 
    
27.  network = input_data(shape=[None, 227, 227, 3])
28.  conv1_7_7 = conv_2d(network, 64, 7, strides=2, activation='relu', name = 'conv1_7_7_s2')
29.  pool1_3_3 = max_pool_2d(conv1_7_7, 3,strides=2)
30.  pool1_3_3 = local_response_normalization(pool1_3_3)
31.  conv2_3_3_reduce = conv_2d(pool1_3_3, 64,1, activation='relu',name = 'conv2_3_3_reduce')
32.  conv2_3_3 = conv_2d(conv2_3_3_reduce, 192,3, activation='relu', name='conv2_3_3')
33.  conv2_3_3 = local_response_normalization(conv2_3_3)
34.  pool2_3_3 = max_pool_2d(conv2_3_3, kernel_size=3, strides=2, name='pool2_3_3_s2')
35.  inception_3a_1_1 = conv_2d(pool2_3_3, 64, 1, activation='relu', name='inception_3a_1_1')
36.  inception_3a_3_3_reduce = conv_2d(pool2_3_3, 96,1, activation='relu', name='inception_3a_3_3_reduce')
37.  inception_3a_3_3 = conv_2d(inception_3a_3_3_reduce, 128,filter_size=3,  activation='relu', name = 'inception_3a_3_3')
38.  inception_3a_5_5_reduce = conv_2d(pool2_3_3,16, filter_size=1,activation='relu', name ='inception_3a_5_5_reduce' )
39.  inception_3a_5_5 = conv_2d(inception_3a_5_5_reduce, 32, filter_size=5, activation='relu', name= 'inception_3a_5_5')
40.  inception_3a_pool = max_pool_2d(pool2_3_3, kernel_size=3, strides=1, )
41.  inception_3a_pool_1_1 = conv_2d(inception_3a_pool, 32, filter_size=1, activation='relu', name='inception_3a_pool_1_1')
42. 
    
43.  # merge the inception_3a__
44.  inception_3a_output = merge([inception_3a_1_1, inception_3a_3_3, inception_3a_5_5, inception_3a_pool_1_1], mode='concat', axis=3)
45. 
    
46.  inception_3b_1_1 = conv_2d(inception_3a_output, 128,filter_size=1,activation='relu', name= 'inception_3b_1_1' )
47.  inception_3b_3_3_reduce = conv_2d(inception_3a_output, 128, filter_size=1, activation='relu', name='inception_3b_3_3_reduce')
48.  inception_3b_3_3 = conv_2d(inception_3b_3_3_reduce, 192, filter_size=3,  activation='relu',name='inception_3b_3_3')
49.  inception_3b_5_5_reduce = conv_2d(inception_3a_output, 32, filter_size=1, activation='relu', name = 'inception_3b_5_5_reduce')
50.  inception_3b_5_5 = conv_2d(inception_3b_5_5_reduce, 96, filter_size=5,  name = 'inception_3b_5_5')
51.  inception_3b_pool = max_pool_2d(inception_3a_output, kernel_size=3, strides=1,  name='inception_3b_pool')
52.  inception_3b_pool_1_1 = conv_2d(inception_3b_pool, 64, filter_size=1,activation='relu', name='inception_3b_pool_1_1')
53. 
    
54.  #merge the inception_3b_*
55.  inception_3b_output = merge([inception_3b_1_1, inception_3b_3_3, inception_3b_5_5, inception_3b_pool_1_1], mode='concat',axis=3,name='inception_3b_output')
56. 
    
57.  pool3_3_3 = max_pool_2d(inception_3b_output, kernel_size=3, strides=2, name='pool3_3_3')
58.  inception_4a_1_1 = conv_2d(pool3_3_3, 192, filter_size=1, activation='relu', name='inception_4a_1_1')
59.  inception_4a_3_3_reduce = conv_2d(pool3_3_3, 96, filter_size=1, activation='relu', name='inception_4a_3_3_reduce')
60.  inception_4a_3_3 = conv_2d(inception_4a_3_3_reduce, 208, filter_size=3,  activation='relu', name='inception_4a_3_3')
61.  inception_4a_5_5_reduce = conv_2d(pool3_3_3, 16, filter_size=1, activation='relu', name='inception_4a_5_5_reduce')
62.  inception_4a_5_5 = conv_2d(inception_4a_5_5_reduce, 48, filter_size=5,  activation='relu', name='inception_4a_5_5')
63.  inception_4a_pool = max_pool_2d(pool3_3_3, kernel_size=3, strides=1,  name='inception_4a_pool')
64.  inception_4a_pool_1_1 = conv_2d(inception_4a_pool, 64, filter_size=1, activation='relu', name='inception_4a_pool_1_1')
65. 
    
66.  inception_4a_output = merge([inception_4a_1_1, inception_4a_3_3, inception_4a_5_5, inception_4a_pool_1_1], mode='concat', axis=3, name='inception_4a_output')
67. 
    
68. 
    
69.  inception_4b_1_1 = conv_2d(inception_4a_output, 160, filter_size=1, activation='relu', name='inception_4a_1_1')
70.  inception_4b_3_3_reduce = conv_2d(inception_4a_output, 112, filter_size=1, activation='relu', name='inception_4b_3_3_reduce')
71.  inception_4b_3_3 = conv_2d(inception_4b_3_3_reduce, 224, filter_size=3, activation='relu', name='inception_4b_3_3')
72.  inception_4b_5_5_reduce = conv_2d(inception_4a_output, 24, filter_size=1, activation='relu', name='inception_4b_5_5_reduce')
73.  inception_4b_5_5 = conv_2d(inception_4b_5_5_reduce, 64, filter_size=5,  activation='relu', name='inception_4b_5_5')
74. 
    
75.  inception_4b_pool = max_pool_2d(inception_4a_output, kernel_size=3, strides=1,  name='inception_4b_pool')
76.  inception_4b_pool_1_1 = conv_2d(inception_4b_pool, 64, filter_size=1, activation='relu', name='inception_4b_pool_1_1')
77. 
    
78.  inception_4b_output = merge([inception_4b_1_1, inception_4b_3_3, inception_4b_5_5, inception_4b_pool_1_1], mode='concat', axis=3, name='inception_4b_output')
79. 
    
80. 
    
81.  inception_4c_1_1 = conv_2d(inception_4b_output, 128, filter_size=1, activation='relu',name='inception_4c_1_1')
82.  inception_4c_3_3_reduce = conv_2d(inception_4b_output, 128, filter_size=1, activation='relu', name='inception_4c_3_3_reduce')
83.  inception_4c_3_3 = conv_2d(inception_4c_3_3_reduce, 256,  filter_size=3, activation='relu', name='inception_4c_3_3')
84.  inception_4c_5_5_reduce = conv_2d(inception_4b_output, 24, filter_size=1, activation='relu', name='inception_4c_5_5_reduce')
85.  inception_4c_5_5 = conv_2d(inception_4c_5_5_reduce, 64,  filter_size=5, activation='relu', name='inception_4c_5_5')
86. 
    
87.  inception_4c_pool = max_pool_2d(inception_4b_output, kernel_size=3, strides=1)
88.  inception_4c_pool_1_1 = conv_2d(inception_4c_pool, 64, filter_size=1, activation='relu', name='inception_4c_pool_1_1')
89. 
    
90.  inception_4c_output = merge([inception_4c_1_1, inception_4c_3_3, inception_4c_5_5, inception_4c_pool_1_1], mode='concat', axis=3,name='inception_4c_output')
91. 
    
92.  inception_4d_1_1 = conv_2d(inception_4c_output, 112, filter_size=1, activation='relu', name='inception_4d_1_1')
93.  inception_4d_3_3_reduce = conv_2d(inception_4c_output, 144, filter_size=1, activation='relu', name='inception_4d_3_3_reduce')
94.  inception_4d_3_3 = conv_2d(inception_4d_3_3_reduce, 288, filter_size=3, activation='relu', name='inception_4d_3_3')
95.  inception_4d_5_5_reduce = conv_2d(inception_4c_output, 32, filter_size=1, activation='relu', name='inception_4d_5_5_reduce')
96.  inception_4d_5_5 = conv_2d(inception_4d_5_5_reduce, 64, filter_size=5,  activation='relu', name='inception_4d_5_5')
97.  inception_4d_pool = max_pool_2d(inception_4c_output, kernel_size=3, strides=1,  name='inception_4d_pool')
98.  inception_4d_pool_1_1 = conv_2d(inception_4d_pool, 64, filter_size=1, activation='relu', name='inception_4d_pool_1_1')
99. 
    
100.  inception_4d_output = merge([inception_4d_1_1, inception_4d_3_3, inception_4d_5_5, inception_4d_pool_1_1], mode='concat', axis=3, name='inception_4d_output')
101. 
     
102.  inception_4e_1_1 = conv_2d(inception_4d_output, 256, filter_size=1, activation='relu', name='inception_4e_1_1')
103.  inception_4e_3_3_reduce = conv_2d(inception_4d_output, 160, filter_size=1, activation='relu', name='inception_4e_3_3_reduce')
104.  inception_4e_3_3 = conv_2d(inception_4e_3_3_reduce, 320, filter_size=3, activation='relu', name='inception_4e_3_3')
105.  inception_4e_5_5_reduce = conv_2d(inception_4d_output, 32, filter_size=1, activation='relu', name='inception_4e_5_5_reduce')
106.  inception_4e_5_5 = conv_2d(inception_4e_5_5_reduce, 128,  filter_size=5, activation='relu', name='inception_4e_5_5')
107.  inception_4e_pool = max_pool_2d(inception_4d_output, kernel_size=3, strides=1,  name='inception_4e_pool')
108.  inception_4e_pool_1_1 = conv_2d(inception_4e_pool, 128, filter_size=1, activation='relu', name='inception_4e_pool_1_1')
109. 
     
110. 
     
111.  inception_4e_output = merge([inception_4e_1_1, inception_4e_3_3, inception_4e_5_5,inception_4e_pool_1_1],axis=3, mode='concat')
112. 
     
113.  pool4_3_3 = max_pool_2d(inception_4e_output, kernel_size=3, strides=2, name='pool_3_3')
114. 
     
115. 
     
116.  inception_5a_1_1 = conv_2d(pool4_3_3, 256, filter_size=1, activation='relu', name='inception_5a_1_1')
117.  inception_5a_3_3_reduce = conv_2d(pool4_3_3, 160, filter_size=1, activation='relu', name='inception_5a_3_3_reduce')
118.  inception_5a_3_3 = conv_2d(inception_5a_3_3_reduce, 320, filter_size=3, activation='relu', name='inception_5a_3_3')
119.  inception_5a_5_5_reduce = conv_2d(pool4_3_3, 32, filter_size=1, activation='relu', name='inception_5a_5_5_reduce')
120.  inception_5a_5_5 = conv_2d(inception_5a_5_5_reduce, 128, filter_size=5,  activation='relu', name='inception_5a_5_5')
121.  inception_5a_pool = max_pool_2d(pool4_3_3, kernel_size=3, strides=1,  name='inception_5a_pool')
122.  inception_5a_pool_1_1 = conv_2d(inception_5a_pool, 128, filter_size=1,activation='relu', name='inception_5a_pool_1_1')
123. 
     
124.  inception_5a_output = merge([inception_5a_1_1, inception_5a_3_3, inception_5a_5_5, inception_5a_pool_1_1], axis=3,mode='concat')
125. 
     
126. 
     
127.  inception_5b_1_1 = conv_2d(inception_5a_output, 384, filter_size=1,activation='relu', name='inception_5b_1_1')
128.  inception_5b_3_3_reduce = conv_2d(inception_5a_output, 192, filter_size=1, activation='relu', name='inception_5b_3_3_reduce')
129.  inception_5b_3_3 = conv_2d(inception_5b_3_3_reduce, 384,  filter_size=3,activation='relu', name='inception_5b_3_3')
130.  inception_5b_5_5_reduce = conv_2d(inception_5a_output, 48, filter_size=1, activation='relu', name='inception_5b_5_5_reduce')
131.  inception_5b_5_5 = conv_2d(inception_5b_5_5_reduce,128, filter_size=5,  activation='relu', name='inception_5b_5_5' )
132.  inception_5b_pool = max_pool_2d(inception_5a_output, kernel_size=3, strides=1,  name='inception_5b_pool')
133.  inception_5b_pool_1_1 = conv_2d(inception_5b_pool, 128, filter_size=1, activation='relu', name='inception_5b_pool_1_1')
134.  inception_5b_output = merge([inception_5b_1_1, inception_5b_3_3, inception_5b_5_5, inception_5b_pool_1_1], axis=3, mode='concat')
135. 
     
136.  pool5_7_7 = avg_pool_2d(inception_5b_output, kernel_size=7, strides=1)
137.  pool5_7_7 = dropout(pool5_7_7, 0.4)
138.  loss = fully_connected(pool5_7_7, 17,activation='softmax')
139.  network = regression(loss, optimizer='momentum',
140.                       loss='categorical_crossentropy',
141.                       learning_rate=0.001)
142.  model = tflearn.DNN(network, checkpoint_path='model_googlenet',
143.                      max_checkpoints=1, tensorboard_verbose=2)
144.  model.fit(X, Y, n_epoch=1000, validation_set=0.1, shuffle=True,
145.            show_metric=True, batch_size=64, snapshot_step=200,
146.            snapshot_epoch=False, run_id='googlenet_oxflowers17')