forked from Ecohnoch/Tiny-Face-Recognition
-
Notifications
You must be signed in to change notification settings - Fork 1
/
resnet101.py
150 lines (121 loc) · 10.1 KB
/
resnet101.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
import tensorflow as tf
import numpy as np
def identity_block2d(input_tensor, kernel_size, filters, stage, block, is_training,
reuse): # x1, 3, [64, 64, 256], stage=2, block='1b', is_training=is_training, reuse=reuse
filters1, filters2, filters3 = filters
conv_name_1 = 'conv' + str(stage) + '_' + str(block) + '_1x1_reduce'
bn_name_1 = 'bn' + str(stage) + '_' + str(block) + '_1x1_reduce'
x = tf.layers.conv2d(input_tensor, filters1, (1, 1), use_bias=False, name=conv_name_1,
reuse=reuse) # 1b(?, 56, 56, 64) 2a(?,28,28,128)
x = tf.layers.batch_normalization(x, training=is_training, name=bn_name_1, reuse=reuse)
x = tf.nn.relu(x)
conv_name_2 = 'conv' + str(stage) + '_' + str(block) + '_3x3'
bn_name_2 = 'bn' + str(stage) + '_' + str(block) + '_3x3'
x = tf.layers.conv2d(x, filters2, kernel_size, padding='SAME', use_bias=False, name=conv_name_2,
reuse=reuse) # 1b(?, 56, 56, 64) 2a(?,28,28,128)
x = tf.layers.batch_normalization(x, training=is_training, name=bn_name_2, reuse=reuse)
x = tf.nn.relu(x)
conv_name_3 = 'conv' + str(stage) + '_' + str(block) + '_1x1_increase'
bn_name_3 = 'bn' + str(stage) + '_' + str(block) + '_1x1_increase'
x = tf.layers.conv2d(x, filters3, (1, 1), name=conv_name_3, use_bias=False,
reuse=reuse) # 1b(?, 56, 56, 256) 2a(?,28,28,512)
x = tf.layers.batch_normalization(x, training=is_training, name=bn_name_3, reuse=reuse)
x = tf.add(input_tensor, x)
x = tf.nn.relu(x)
return x
def conv_block_2d(input_tensor, kernel_size, filters, stage, block, is_training, reuse, strides=(
2, 2)): # x, 3, [64, 64, 256], stage=2, block='1a', strides=(2, 2), is_training=is_training, reuse=reuse
filters1, filters2, filters3 = filters # ResNet利用了1×1卷积，并且是在3×3卷积层的前后都使用了，不仅进行了降维，还进行了升维，使得卷积层的输入和输出的通道数都减小，参数数量进一步减少
conv_name_1 = 'conv' + str(stage) + '_' + str(block) + '_1x1_reduce'
bn_name_1 = 'bn' + str(stage) + '_' + str(block) + '_1x1_reduce'
x = tf.layers.conv2d(input_tensor, filters1, (1, 1), use_bias=False, strides=strides, name=conv_name_1,
reuse=reuse) # 1a(?,56,56,64) 2a(?,28,28,128)
x = tf.layers.batch_normalization(x, training=is_training, name=bn_name_1, reuse=reuse) # 1a(?,56,56,64)
x = tf.nn.relu(x)
conv_name_2 = 'conv' + str(stage) + '_' + str(block) + '_3x3'
bn_name_2 = 'bn' + str(stage) + '_' + str(block) + '_3x3'
x = tf.layers.conv2d(x, filters2, kernel_size, padding='SAME', use_bias=False, name=conv_name_2,
reuse=reuse) # 1a(?,56,56,64) 2a(?,28,28,128)
x = tf.layers.batch_normalization(x, training=is_training, name=bn_name_2, reuse=reuse) # 1a(?,56,56,64)
x = tf.nn.relu(x)
conv_name_3 = 'conv' + str(stage) + '_' + str(block) + '_1x1_increase'
bn_name_3 = 'bn' + str(stage) + '_' + str(block) + '_1x1_increase'
x = tf.layers.conv2d(x, filters3, (1, 1), name=conv_name_3, use_bias=False,
reuse=reuse) # 1a(?,56,56,256) 2a(?,28,28,512)
x = tf.layers.batch_normalization(x, training=is_training, name=bn_name_3,
reuse=reuse) # 1a(?,56,56,256) 2a(?,28,28,512)
conv_name_4 = 'conv' + str(stage) + '_' + str(block) + '_1x1_shortcut'
bn_name_4 = 'bn' + str(stage) + '_' + str(block) + '_1x1_shortcut'
shortcut = tf.layers.conv2d(input_tensor, filters3, (1, 1), use_bias=False, strides=strides, name=conv_name_4,
reuse=reuse) # 1a(?,56,56,256) 2a(?,28,28,512)
shortcut = tf.layers.batch_normalization(shortcut, training=is_training, name=bn_name_4,
reuse=reuse) # 1a(?,56,56,256) 2a(?,28,28,512)
x = tf.add(shortcut, x) # 对应元素相加，f(x) + x，# 1a(?,56,56,256) 2a(?,28,28,512)
x = tf.nn.relu(x)
return x
def resnet50(input_tensor, is_training=True, pooling_and_fc=True, reuse=False):
x = tf.layers.conv2d(input_tensor, 64, (7, 7), strides=(1, 1), padding='SAME', use_bias=False,
name='conv1_1/3x3_s1', reuse=reuse) # (?,112,112,64) 第一层卷积
x = tf.layers.batch_normalization(x, training=is_training, name='bn1_1/3x3_s1',
reuse=reuse) # (?,112,112,64) bn 输入batch标准化
x = tf.nn.relu(x) # (?,112,112,64) 激活函数relu
# x = tf.layers.max_pooling2d(x, (2,2), strides=(2,2), name='mpool1')
x1 = conv_block_2d(x, 3, [64, 64, 256], stage=2, block='1a', strides=(2, 2), is_training=is_training,
reuse=reuse) # 先stride2一次减小，L:112>>56，D:3>>64升维,k=1；再卷一次L:56>>56，D:64>>64,k=3捕获像素八邻域信息；再卷一次加深L:56>>56，D:64>>256二次升维,k=1；123串行产生A，4input图上sdride2减小为fmap大小，L:112>>56，D:3>>256,k=1产生B；5A+B对应元素相加[升维，拓宽，升维，+x]
x1 = identity_block2d(x1, 3, [64, 64, 256], stage=2, block='1b', is_training=is_training,
reuse=reuse) # 三次卷积stride都=1，k分别为[1,3,1]，D分别为 [64,64,256]，[降维，拓宽，升维]
x1 = identity_block2d(x1, 3, [64, 64, 256], stage=2, block='1c', is_training=is_training,
reuse=reuse) # 三次卷积stride都=1，k分别为[1,3,1]，D分别为 [64,64,256]，[降维，拓宽，升维]
x2 = conv_block_2d(x1, 3, [128, 128, 512], stage=3, block='2a', is_training=is_training, reuse=reuse)
x2 = identity_block2d(x2, 3, [128, 128, 512], stage=3, block='2b', is_training=is_training, reuse=reuse)
x2 = identity_block2d(x2, 3, [128, 128, 512], stage=3, block='2c', is_training=is_training, reuse=reuse)
x2 = identity_block2d(x2, 3, [128, 128, 512], stage=3, block='2d', is_training=is_training,
reuse=reuse) # (?,28,28,512)
# res50 有6个identity_block2d，res101 有23个identity_block2d，
x3 = conv_block_2d(x2, 3, [256, 256, 1024], stage=4, block='3a', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3b', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3c', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3d', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3e', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3f', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3g', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3h', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3i', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3j', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3k', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3l', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3m', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3n', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3o', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3p', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3q', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3r', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3s', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3t', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3u', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3v', is_training=is_training, reuse=reuse)
x3 = identity_block2d(x3, 3, [256, 256, 1024], stage=4, block='3w', is_training=is_training, reuse=reuse)
# (?,14,14,1024)
x4 = conv_block_2d(x3, 3, [512, 512, 2048], stage=5, block='4a', is_training=is_training, reuse=reuse)
x4 = identity_block2d(x4, 3, [512, 512, 2048], stage=5, block='4b', is_training=is_training, reuse=reuse)
x4 = identity_block2d(x4, 3, [512, 512, 2048], stage=5, block='4c', is_training=is_training,
reuse=reuse) # (?,7,7,2048)
if pooling_and_fc:
# pooling_output = tf.layers.max_pooling2d(x4, (7,7), strides=(1,1), name='mpool2')
pooling_output = tf.contrib.layers.flatten(x4) # (?, 100325=7*7*2058)
fc_output = tf.layers.dense(pooling_output, 512, name='fc1', reuse=reuse) # (?, 512)
fc_output = tf.layers.batch_normalization(fc_output, training=is_training, name='fbn')
return fc_output
# if __name__ == '__main__':
# example_data = [np.random.rand(112, 112, 3)]
# x = tf.placeholder(tf.float32, [None, 112, 112, 3])
# y = resnet50(x, is_training=True, reuse=False)
# print(y)
#
# with tf.Session() as sess:
# writer = tf.summary.FileWriter("logs/", sess.graph)
# init = tf.global_variables_initializer()
# sess.run(init)
# 例如一个50层的ResNet网络，其结构可以表示为2+48，其中2表示预处理，48则是conv卷积层的数目，采用三层的残差学习网络，由于3个卷积层为一个残差网络，故48/3=16个残差网络。
# 规定第一个block块和最后一个都只包括3个残差网络，如图50层的ResNet网络的结构为：（3+4+6+3）x 3 +2
# 相似地，101层的ResNet网络的结构为：（3+4+23+3）x 3 +2