forked from Pay20Y/GCAN
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathBackbone.py
189 lines (170 loc) · 8.75 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
import tensorflow as tf
from tensorflow.contrib import slim
# from module.nets import resnet_utils, resnet_v1
def mean_image_subtraction(images, means=[128.0, 128.0, 128.0]):
'''
image normalization
:param images:
:param means:
:return:
'''
with tf.variable_scope("mean_subtraction"):
num_channels = images.get_shape().as_list()[-1]
if len(means) != num_channels:
raise ValueError('len(means) must match the number of channels')
channels = tf.split(axis=3, num_or_size_splits=num_channels, value=images)
for i in range(num_channels):
channels[i] -= means[i]
return tf.concat(axis=3, values=channels)
class Backbone(object):
def __init__(self, weight_decay=1e-5, is_training=True):
self.weight_decay = weight_decay
self.is_training = is_training
def shortcut(self, inputs, output_dim, stride=1):
depth_in = slim.utils.last_dimension(inputs.get_shape(), min_rank=4)
if depth_in == output_dim:
shortcut = inputs
else:
shortcut = slim.conv2d(inputs, output_dim, [1, 1], stride=stride, activation_fn=None, scope='shortcut')
return shortcut
def basicblock(self, inputs, output_dim, kernel_size, stride, scope):
with tf.variable_scope(scope):
shortcut = self.shortcut(inputs, output_dim)
residual = slim.conv2d(inputs, output_dim, kernel_size, stride=stride, scope='conv1')
residual = slim.conv2d(residual, output_dim, kernel_size, stride=stride, activation_fn=None, scope='conv2')
return tf.nn.relu(shortcut + residual)
def __call__(self, input_image):
with tf.variable_scope("resnet_backbone"):
batch_norm_params = {
'decay': 0.997,
'epsilon': 1e-5,
'scale': True,
'is_training': self.is_training
}
with slim.arg_scope([slim.conv2d],
activation_fn=tf.nn.relu,
normalizer_fn=slim.batch_norm,
normalizer_params=batch_norm_params,
weights_regularizer=slim.l2_regularizer(self.weight_decay)):
input_image = mean_image_subtraction(input_image)
net = slim.conv2d(input_image, 64, 3, stride=1, rate=1, padding='SAME', scope="conv1")
net = slim.conv2d(net, 128, 3, stride=1, rate=1, padding='SAME', scope="conv2")
net = slim.max_pool2d(net, [2, 2], stride=2)
# 1 Block here
net = self.basicblock(net, output_dim=256, kernel_size=3, stride=1, scope="block1")
# 1 Block end
net = slim.conv2d(net, 256, 3, stride=1, rate=1, padding='SAME', scope="conv3")
net = slim.max_pool2d(net, [2, 2], stride=2)
# 2 Blocks here
net = self.basicblock(net, output_dim=256, kernel_size=3, stride=1, scope="block2")
net = self.basicblock(net, output_dim=256, kernel_size=3, stride=1, scope="block3")
# 2 Blocks end
net = slim.conv2d(net, 256, 3, stride=1, rate=1, padding='SAME', scope="conv4")
net = slim.max_pool2d(net, [2, 1], stride=[2, 1])
# 5 Blocks here
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block4")
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block5")
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block6")
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block7")
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block8")
# 5 Blocks end
net = slim.conv2d(net, 512, 3, stride=1, rate=1, padding='SAME', scope="conv5")
# 3 Blocks here
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block9")
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block10")
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block11")
# 3 Blocks end
net = slim.conv2d(net, 512, 3, stride=1, rate=1, padding='SAME', scope="conv6")
print("Backbone final output: ", net)
return net
# remove slim batch norm
class Backbone_v2(object):
def __init__(self, weight_decay=1e-5, is_training=True):
self.weight_decay = weight_decay
self.is_training = is_training
def shortcut(self, inputs, output_dim, stride=1):
depth_in = slim.utils.last_dimension(inputs.get_shape(), min_rank=4)
if depth_in == output_dim:
shortcut = inputs
else:
shortcut = slim.conv2d(inputs, output_dim, [1, 1], stride=stride, activation_fn=None, scope='shortcut')
shortcut = tf.layers.batch_normalization(shortcut, training=self.is_training, name='bn')
return shortcut
def basicblock(self, inputs, output_dim, kernel_size, stride, scope):
with tf.variable_scope(scope):
shortcut = self.shortcut(inputs, output_dim)
residual = slim.conv2d(inputs, output_dim, kernel_size, stride=stride, scope='conv1')
residual = tf.layers.batch_normalization(residual, training=self.is_training, name='conv1_bn')
residual = slim.conv2d(residual, output_dim, kernel_size, stride=stride, activation_fn=None, scope='conv2')
residual = tf.layers.batch_normalization(residual, training=self.is_training, name='conv2_bn')
return tf.nn.relu(shortcut + residual)
def __call__(self, input_image):
with tf.variable_scope("resnet_backbone"):
batch_norm_params = {
'decay': 0.997,
'epsilon': 1e-5,
'scale': True,
'is_training': self.is_training
}
with slim.arg_scope([slim.conv2d],
activation_fn=tf.nn.relu,
normalizer_fn=None,
normalizer_params=None,
weights_regularizer=slim.l2_regularizer(self.weight_decay)):
input_image = mean_image_subtraction(input_image)
net = slim.conv2d(input_image, 64, 3, stride=1, rate=1, padding='SAME', scope="conv1")
net = tf.layers.batch_normalization(net, training=self.is_training, name='conv1_bn')
net = slim.conv2d(net, 128, 3, stride=1, rate=1, padding='SAME', scope="conv2")
net = tf.layers.batch_normalization(net, training=self.is_training, name='conv2_bn')
net = slim.max_pool2d(net, [2, 2], stride=2)
# 1 Block here
net = self.basicblock(net, output_dim=256, kernel_size=3, stride=1, scope="block1")
# 1 Block end
net = slim.conv2d(net, 256, 3, stride=1, rate=1, padding='SAME', scope="conv3")
net = tf.layers.batch_normalization(net, training=self.is_training, name='conv3_bn')
net = slim.max_pool2d(net, [2, 2], stride=2)
# 2 Blocks here
net = self.basicblock(net, output_dim=256, kernel_size=3, stride=1, scope="block2")
net = self.basicblock(net, output_dim=256, kernel_size=3, stride=1, scope="block3")
# 2 Blocks end
net = slim.conv2d(net, 256, 3, stride=1, rate=1, padding='SAME', scope="conv4")
net = tf.layers.batch_normalization(net, training=self.is_training, name='conv4_bn')
net = slim.max_pool2d(net, [2, 1], stride=[2, 1])
# 5 Blocks here
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block4")
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block5")
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block6")
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block7")
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block8")
# 5 Blocks end
net = slim.conv2d(net, 512, 3, stride=1, rate=1, padding='SAME', scope="conv5")
net = tf.layers.batch_normalization(net, training=self.is_training, name='conv5_bn')
# 3 Blocks here
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block9")
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block10")
net = self.basicblock(net, output_dim=512, kernel_size=3, stride=1, scope="block11")
# 3 Blocks end
net = slim.conv2d(net, 512, 3, stride=1, rate=1, padding='SAME', scope="conv6")
net = tf.layers.batch_normalization(net, training=self.is_training, name='conv6_bn')
print("Backbone final output: ", net)
return net
if __name__ == '__main__':
import numpy as np
import os
os.environ['CUDA_VISIBLE_DEVICES'] = "1"
bb = Backbone()
input_images = tf.placeholder(tf.float32, shape=[32, 48 ,160, 3], name="input_images")
input_feature_label = tf.placeholder(tf.float32, shape=[32, 6, 40, 512], name="input_feature_label")
feature_map = bb(input_images)
loss = tf.reduce_mean(input_feature_label - feature_map)
print("feature_map: ", feature_map)
optimizer = tf.train.AdamOptimizer(learning_rate=1.0).minimize(loss)
_input_images = np.random.rand(32, 48, 160, 3)
_input_feature_label = np.random.rand(32, 6, 40, 512)
summary_writer = tf.summary.FileWriter("toy_summary")
with tf.Session() as sess:
summary_writer.add_graph(sess.graph)
sess.run(tf.global_variables_initializer())
for i in range(10):
_, loss_value = sess.run([optimizer, loss], feed_dict={input_images: _input_images, input_feature_label: _input_feature_label})
print(loss_value)