-
Notifications
You must be signed in to change notification settings - Fork 12
/
Copy pathresnet_model.py
132 lines (107 loc) · 5.06 KB
/
resnet_model.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
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# File: resnet_model.py
import tensorflow as tf
from tensorflow.contrib.layers import variance_scaling_initializer
from tensorpack.tfutils.argscope import argscope, get_arg_scope
from tensorpack.models import (
Conv2D, MaxPooling, GlobalAvgPooling, BatchNorm, FullyConnected,
LinearWrap)
def BNReLU(x, name=None):
"""
A shorthand of BatchNormalization + ReLU.
"""
x = get_bn()(x, None)
x = tf.nn.relu(x, name=name)
return x
def resnet_shortcut(l, n_out, stride, nl=tf.identity):
data_format = get_arg_scope()['Conv2D']['data_format']
n_in = l.get_shape().as_list()[1 if data_format == 'NCHW' else 3]
if n_in != n_out: # change dimension when channel is not the same
return Conv2D('convshortcut', l, n_out, 1, stride=stride, nl=nl)
else:
return l
def apply_preactivation(l, preact):
if preact == 'bnrelu':
shortcut = l # preserve identity mapping
l = BNReLU('preact', l)
else:
shortcut = l
return l, shortcut
def get_bn(zero_init=False):
"""
Zero init gamma is good for resnet. See https://arxiv.org/abs/1706.02677.
"""
if zero_init:
return lambda x, name: BatchNorm('bn', x, gamma_init=tf.zeros_initializer(),
internal_update=True)
else:
return lambda x, name: BatchNorm('bn', x, internal_update=True)
def preresnet_basicblock(l, ch_out, stride, preact):
l, shortcut = apply_preactivation(l, preact)
l = Conv2D('conv1', l, ch_out, 3, stride=stride, nl=BNReLU)
l = Conv2D('conv2', l, ch_out, 3)
return l + resnet_shortcut(shortcut, ch_out, stride)
def preresnet_bottleneck(l, ch_out, stride, preact):
# stride is applied on the second conv, following fb.resnet.torch
l, shortcut = apply_preactivation(l, preact)
l = Conv2D('conv1', l, ch_out, 1, nl=BNReLU)
l = Conv2D('conv2', l, ch_out, 3, stride=stride, nl=BNReLU)
l = Conv2D('conv3', l, ch_out * 4, 1)
return l + resnet_shortcut(shortcut, ch_out * 4, stride)
def preresnet_group(l, name, block_func, features, count, stride):
with tf.variable_scope(name):
for i in range(0, count):
with tf.variable_scope('block{}'.format(i)):
# first block doesn't need activation
l = block_func(l, features,
stride if i == 0 else 1,
'no_preact' if i == 0 else 'bnrelu')
# end of each group need an extra activation
l = BNReLU('bnlast', l)
return l
def resnet_basicblock(l, ch_out, stride):
shortcut = l
l = Conv2D('conv1', l, ch_out, 3, stride=stride, nl=BNReLU)
l = Conv2D('conv2', l, ch_out, 3, nl=get_bn(zero_init=True))
return l + resnet_shortcut(shortcut, ch_out, stride, nl=get_bn(zero_init=False))
def resnet_bottleneck(l, ch_out, stride, stride_first=False):
"""
stride_first: original resnet put stride on first conv. fb.resnet.torch put stride on second conv.
"""
shortcut = l
l = Conv2D('conv1', l, ch_out, 1, stride=stride if stride_first else 1, nl=BNReLU)
l = Conv2D('conv2', l, ch_out, 3, stride=1 if stride_first else stride, nl=BNReLU)
l = Conv2D('conv3', l, ch_out * 4, 1, nl=get_bn(zero_init=True))
return l + resnet_shortcut(shortcut, ch_out * 4, stride, nl=get_bn(zero_init=False))
def se_resnet_bottleneck(l, ch_out, stride):
shortcut = l
l = Conv2D('conv1', l, ch_out, 1, nl=BNReLU)
l = Conv2D('conv2', l, ch_out, 3, stride=stride, nl=BNReLU)
l = Conv2D('conv3', l, ch_out * 4, 1, nl=get_bn(zero_init=True))
squeeze = GlobalAvgPooling('gap', l)
squeeze = FullyConnected('fc1', squeeze, ch_out // 4, nl=tf.nn.relu)
squeeze = FullyConnected('fc2', squeeze, ch_out * 4, nl=tf.nn.sigmoid)
l = l * tf.reshape(squeeze, [-1, ch_out * 4, 1, 1])
return l + resnet_shortcut(shortcut, ch_out * 4, stride, nl=get_bn(zero_init=False))
def resnet_group(l, name, block_func, features, count, stride):
with tf.variable_scope(name):
for i in range(0, count):
with tf.variable_scope('block{}'.format(i)):
l = block_func(l, features, stride if i == 0 else 1)
# end of each block need an activation
l = tf.nn.relu(l)
return l
def resnet_backbone(image, num_blocks, group_func, block_func):
with argscope(Conv2D, nl=tf.identity, use_bias=False,
W_init=variance_scaling_initializer(mode='FAN_OUT')):
logits = (LinearWrap(image)
.Conv2D('conv0', 64, 7, stride=2, nl=BNReLU)
.MaxPooling('pool0', shape=3, stride=2, padding='SAME')
.apply(group_func, 'group0', block_func, 64, num_blocks[0], 1)
.apply(group_func, 'group1', block_func, 128, num_blocks[1], 2)
.apply(group_func, 'group2', block_func, 256, num_blocks[2], 2)
.apply(group_func, 'group3', block_func, 512, num_blocks[3], 2)
.GlobalAvgPooling('gap')
.FullyConnected('linear', 1000, nl=tf.identity)())
return logits