forked from zcdliuwei/RetinaNet
-
Notifications
You must be signed in to change notification settings - Fork 0
/
backbone.py
202 lines (167 loc) · 6.79 KB
/
backbone.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
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
#-- coding: utf-8 --
import tensorflow as tf
import numpy as np
def res_net50(x):
x = conv_stage_1(x)
x = conv_stage_2(x)
x = conv_stage_3(x)
c3 = x
x = conv_stage_4_(x)
c4 = x
x = conv_stage_5(x)
c5 = x
return c3, c4, c5 #(bacthsize, 28, 28, 512) (bacthsize, 14, 14, 1024) (bacthsize, 7, 7, 2048)
pass
def res_net101(x):
x = conv_stage_1(x)
x = conv_stage_2(x)
x = conv_stage_3(x)
c3 = x
x = conv_stage_4(x)
c4 = x
x = conv_stage_5(x)
c5 = x
return c3, c4, c5 #(bacthsize, 28, 28, 512) (bacthsize, 14, 14, 1024) (bacthsize, 7, 7, 2048)
pass
def conv_stage_1(x):#224
x = conv_bn_relu(64, 7, 2,padding="same")(x)
return x #112
pass
def conv_stage_2(x):#112
x = tf.keras.layers.MaxPooling2D(3, 2, padding="same")(x)
# 由于上下组的卷积层通道数不同，使得短路连接不能直接相加，故需要在后四组连接上一组的第一个卷积层的短路连接通路添加投影卷积。
x_ = conv_bn(64 * 4, 1, 1)(x)
x_in = conv_bn_relu(64, 1, 1)(x)
x_in = tf.keras.layers.ZeroPadding2D(padding=1)(x_in)
x_in = conv_bn_relu(64, 3, 1)(x_in)
x_in = conv_bn(64 * 4, 1, 1)(x_in)
x_in = tf.keras.layers.Add()([x_in, x_])
x_in = tf.keras.layers.Activation("relu")(x_in)
for i in range(2):
x_in = res_block(x_in, 64, 1)
return x_in #56
pass
def conv_stage_3(x):#56
# 由于上下组的卷积层通道数不同，使得短路连接不能直接相加，故需要在后四组连接上一组的第一个卷积层的短路连接通路添加投影卷积。
x_ = conv_bn(128 * 4, 1, 2)(x)
x_in = conv_bn_relu(128, 1, 2)(x)
x_in = tf.keras.layers.ZeroPadding2D(padding=1)(x_in)
x_in = conv_bn_relu(128, 3, 1)(x_in)
x_in = conv_bn(128 * 4, 1, 1)(x_in)
x_in = tf.keras.layers.Add()([x_in, x_])
x_in = tf.keras.layers.Activation("relu")(x_in)
for i in range(3):
x_in = res_block(x_in, 128, 1)
return x_in #28
pass
def conv_stage_4_(x):#28
# 由于上下组的卷积层通道数不同，使得短路连接不能直接相加，故需要在后四组连接上一组的第一个卷积层的短路连接通路添加投影卷积。
x_ = conv_bn(256 * 4, 1, 2)(x)
x_in = conv_bn_relu(256, 1, 2)(x)
x_in = tf.keras.layers.ZeroPadding2D(padding=1)(x_in)
x_in = conv_bn_relu(256, 3, 1)(x_in)
x_in = conv_bn(256 * 4, 1, 1)(x_in)
x_in = tf.keras.layers.Add()([x_in, x_])
x_in = tf.keras.layers.Activation("relu")(x_in)
for i in range(5):
x_in = res_block(x_in, 256, 1)
return x_in #14
pass
def conv_stage_4(x):#28
# 由于上下组的卷积层通道数不同，使得短路连接不能直接相加，故需要在后四组连接上一组的第一个卷积层的短路连接通路添加投影卷积。
x_ = conv_bn(256 * 4, 1, 2)(x)
x_in = conv_bn_relu(256, 1, 2)(x)
x_in = tf.keras.layers.ZeroPadding2D(padding=1)(x_in)
x_in = conv_bn_relu(256, 3, 1)(x_in)
x_in = conv_bn(256 * 4, 1, 1)(x_in)
x_in = tf.keras.layers.Add()([x_in, x_])
x_in = tf.keras.layers.Activation("relu")(x_in)
for i in range(22):
x_in = res_block(x_in, 256, 1)
return x_in #14
pass
def conv_stage_5(x):#14
#由于上下组的卷积层通道数不同，使得短路连接不能直接相加，故需要在后四组连接上一组的第一个卷积层的短路连接通路添加投影卷积。
x_ = conv_bn(512 * 4, 1, 2)(x)
x_in = conv_bn_relu(512, 1, 2)(x)
x_in = tf.keras.layers.ZeroPadding2D(padding=1)(x_in)
x_in = conv_bn_relu(512, 3, 1)(x_in)
x_in = conv_bn(512 * 4, 1, 1)(x_in)
x_in = tf.keras.layers.Add()([x_in, x_])
x_in = tf.keras.layers.Activation("relu")(x_in)
for i in range(2):
x_in = res_block(x_in, 512, 1)
return x_in #7
pass
def res_block(x, filters,strides):
x_in = conv_bn_relu(filters, 1, strides)(x)
x_in = tf.keras.layers.ZeroPadding2D(padding=1)(x_in)
x_in = conv_bn_relu(filters, 3, strides)(x_in)
x_in = conv_bn(filters * 4, 1, strides)(x_in)
x_in = tf.keras.layers.Add()([x_in, x])
x_in = tf.keras.layers.Activation("relu")(x_in)
return x_in
pass
def conv_bn_relu(filters, filter_size,strides,padding='valid'):
return tf.keras.Sequential([
tf.keras.layers.Conv2D(filters, filter_size, strides, padding=padding,use_bias= False),
tf.keras.layers.BatchNormalization(epsilon=1e-5),
tf.keras.layers.Activation("relu")
])
def conv_bn(filters, filter_size, strides,padding='valid'):
return tf.keras.Sequential([
tf.keras.layers.Conv2D(filters, filter_size, strides, padding=padding,use_bias= False),
tf.keras.layers.BatchNormalization(epsilon=1e-5)
])
pass
def ResNet2D50(inputs, blocks=None):
if blocks is None:
blocks = [3, 4, 6, 3]
pass
x = tf.keras.layers.Conv2D(64, (7, 7), strides=(2, 2), use_bias=False, padding="same")(inputs)
x = tf.keras.layers.BatchNormalization(epsilon=1e-5)(x)
x = tf.keras.layers.Activation("relu")(x)
x = tf.keras.layers.MaxPooling2D((3, 3), strides=(2, 2), padding="same")(x)
features = 64
outputs = []
for stage_id, iterations in enumerate(blocks):
for block_id in range(iterations):
x = bottleneck(features, stage_id, block_id)(x)
pass
features *= 2
outputs.append(x)
pass
return outputs[1:]
pass
def bottleneck(filters,stage=0,block=0,kernel_size=3,stride=None):
if stride is None:
if block != 0 or stage == 0:
stride = 1
else:
stride = 2
def f(x):
y = tf.keras.layers.Conv2D(filters, (1, 1), strides=stride, use_bias=False)(x)
y = tf.keras.layers.BatchNormalization(epsilon=1e-5)(y)
y = tf.keras.layers.Activation("relu")(y)
y = tf.keras.layers.ZeroPadding2D(padding=1)(y)
y = tf.keras.layers.Conv2D(filters, kernel_size, use_bias=False)(y)
y = tf.keras.layers.BatchNormalization(epsilon=1e-5)(y)
y = tf.keras.layers.Activation("relu")(y)
y = tf.keras.layers.Conv2D(filters * 4, (1, 1), use_bias=False)(y)
y = tf.keras.layers.BatchNormalization(epsilon=1e-5)(y)
if block == 0:
shortcut = tf.keras.layers.Conv2D(filters * 4, (1, 1), strides=stride, use_bias=False)(x)
shortcut = tf.keras.layers.BatchNormalization(epsilon=1e-5)(shortcut)
else:
shortcut = x
y = tf.keras.layers.Add()([y, shortcut])
y = tf.keras.layers.Activation("relu")(y)
return y
pass
return f
pass
if __name__ == '__main__':
input =tf.constant(np.random.rand(2,224,224,3),dtype=tf.float32)
res = res_net101(input)
print(res[0].shape,res[1].shape,res[2].shape) #(bacthsize, 28, 28, 512) (bacthsize, 14, 14, 1024) (bacthsize, 7, 7, 2048)
pass