-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathvgg.py
137 lines (116 loc) · 5.16 KB
/
vgg.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
import tensorflow as tf
import numpy as np
REGULARIZER_COF = 1e-8
def _fc_variable( weight_shape,name="fc"):
with tf.variable_scope(name):
# check weight_shape
input_channels = int(weight_shape[0])
output_channels = int(weight_shape[1])
weight_shape = (input_channels, output_channels)
regularizer = tf.contrib.layers.l2_regularizer(scale=REGULARIZER_COF)
# define variables
weight = tf.get_variable("_w", weight_shape ,
initializer=tf.contrib.layers.xavier_initializer(),
regularizer =regularizer)
bias = tf.get_variable("_b", [weight_shape[1]],
initializer=tf.constant_initializer(0.0))
return weight, bias
def _conv_variable( weight_shape,name="conv"):
with tf.variable_scope(name):
# check weight_shape
w = int(weight_shape[0])
h = int(weight_shape[1])
input_channels = int(weight_shape[2])
output_channels = int(weight_shape[3])
weight_shape = (w,h,input_channels, output_channels)
regularizer = tf.contrib.layers.l2_regularizer(scale=REGULARIZER_COF)
# define variables
weight = tf.get_variable("w", weight_shape ,
initializer=tf.constant_initializer(0.0),
regularizer=regularizer)
bias = tf.get_variable("b", [output_channels],
initializer=tf.constant_initializer(0.0))
return weight, bias
def _conv2d( x, W, stride):
return tf.nn.conv2d(x, W, strides = [1, stride, stride, 1], padding = "SAME")
def _flatten(x):
n_b, n_h, n_w, n_f = [int(s) for s in x.get_shape()]
h = tf.reshape(x,[n_b,n_h*n_w*n_f])
return h
def _fc_layer(x, input_layer, output_layer, name="fc", isTraining=True):
w, b = _fc_variable([input_layer,output_layer],name=name)
h = tf.matmul(x, w) + b
return h
def _conv_layer(x, input_layer, output_layer, stride, filter_size=3, name="conv", isTraining=True):
conv_w, conv_b = _conv_variable([filter_size,filter_size,input_layer,output_layer],name=name)
h = _conv2d(x,conv_w,stride=stride) + conv_b
return h
def _max_pooling(x, ksize=[1,2,2,1], strides=[1,2,2,1], padding="SAME"):
return tf.nn.max_pool(x,ksize,strides,padding)
def vgg19(x,reuse=False,isTraining=True,convert=False):
# opencv read in BGR not RGB
VGG_MEAN = [103.939, 116.779, 123.68]
x = tf.image.resize_images(x,(224,224))
x = (x+1) * 127.5
b = x[:,:,:,0] - VGG_MEAN[0]
g = x[:,:,:,1] - VGG_MEAN[1]
r = x[:,:,:,2] - VGG_MEAN[2]
b = tf.reshape(b,[-1,224,224,1])
g = tf.reshape(g,[-1,224,224,1])
r = tf.reshape(r,[-1,224,224,1])
x = tf.concat([b,g,r],axis=3)
with tf.variable_scope("vgg19", reuse=reuse) as scope:
if reuse: scope.reuse_variables()
h = _conv_layer(x, 3, 64, 1, 3, "conv1_1", isTraining=isTraining)
h = tf.nn.relu(h)
h = _conv_layer(h, 64, 64, 1, 3, "conv1_2", isTraining=isTraining)
h_1 = h
h = tf.nn.relu(h)
h = _max_pooling(h)
h = _conv_layer(h, 64, 128, 1, 3, "conv2_1", isTraining=isTraining)
h = tf.nn.relu(h)
h = _conv_layer(h, 128, 128, 1, 3, "conv2_2", isTraining=isTraining)
h_2 = h
h = tf.nn.relu(h)
h = _max_pooling(h)
h = _conv_layer(h, 128, 256, 1, 3, "conv3_1", isTraining=isTraining)
h = tf.nn.relu(h)
h = _conv_layer(h, 256, 256, 1, 3, "conv3_2", isTraining=isTraining)
h = tf.nn.relu(h)
h = _conv_layer(h, 256, 256, 1, 3, "conv3_3", isTraining=isTraining)
h = tf.nn.relu(h)
h = _conv_layer(h, 256, 256, 1, 3, "conv3_4", isTraining=isTraining)
h_3 = h
h = tf.nn.relu(h)
h = _max_pooling(h)
h = _conv_layer(h, 256, 512, 1, 3, "conv4_1", isTraining=isTraining)
h = tf.nn.relu(h)
h = _conv_layer(h, 512, 512, 1, 3, "conv4_2", isTraining=isTraining)
h = tf.nn.relu(h)
h = _conv_layer(h, 512, 512, 1, 3, "conv4_3", isTraining=isTraining)
h = tf.nn.relu(h)
h = _conv_layer(h, 512, 512, 1, 3, "conv4_4", isTraining=isTraining)
h_4 = h
h = tf.nn.relu(h)
h = _max_pooling(h)
h = _conv_layer(h, 512, 512, 1, 3, "conv5_1", isTraining=isTraining)
h = tf.nn.relu(h)
h = _conv_layer(h, 512, 512, 1, 3, "conv5_2", isTraining=isTraining)
h = tf.nn.relu(h)
h = _conv_layer(h, 512, 512, 1, 3, "conv5_3", isTraining=isTraining)
h = tf.nn.relu(h)
h = _conv_layer(h, 512, 512, 1, 3, "conv5_4", isTraining=isTraining)
h_5 = h
h = tf.nn.relu(h)
h = _max_pooling(h)
if convert:
h = _flatten(h)
bs, f = h.get_shape().as_list()
h = _fc_layer(h, f, 4096, "fc6")
h = tf.nn.relu(h)
h = _fc_layer(h, 4096, 4096, "fc7")
h = tf.nn.relu(h)
out = _fc_layer(h, 4096, 1000, "fc8")
else:
pass
return h_1, h_2, h_3, h_4, h_5