-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathmodel_base.py
executable file
·199 lines (164 loc) · 6.12 KB
/
model_base.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
"""Base ResNet model.
"""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import tensorflow as tf
class ResNet(object):
"""ResNet model."""
def __init__(self, is_training, data_format, batch_norm_decay, batch_norm_epsilon):
"""ResNet constructor.
Args:
is_training: if build training or inference model.
data_format: the data_format used during computation.
one of 'channels_first' or 'channels_last'.
"""
self._batch_norm_decay = batch_norm_decay
self._batch_norm_epsilon = batch_norm_epsilon
self._is_training = is_training
assert data_format in ('channels_first', 'channels_last')
self._data_format = data_format
def forward_pass(self, x):
raise NotImplementedError(
'forward_pass() is implemented in ResNet sub classes')
def _residual_v1(self,
x,
kernel_size,
in_filter,
out_filter,
stride,
activate_before_residual=False):
"""Residual unit with 2 sub layers, using Plan A for shortcut connection."""
del activate_before_residual
with tf.name_scope('residual_v1') as name_scope:
orig_x = x
x = self._conv(x, kernel_size, out_filter, stride)
x = self._batch_norm(x)
x = self._relu(x)
x = self._conv(x, kernel_size, out_filter, 1)
x = self._batch_norm(x)
if in_filter != out_filter:
orig_x = self._avg_pool(orig_x, stride, stride)
pad = (out_filter - in_filter) // 2
if self._data_format == 'channels_first':
orig_x = tf.pad(orig_x, [[0, 0], [pad, pad], [0, 0], [0, 0]])
else:
orig_x = tf.pad(orig_x, [[0, 0], [0, 0], [0, 0], [pad, pad]])
x = self._relu(tf.add(x, orig_x))
tf.logging.info('image after unit %s: %s', name_scope, x.get_shape())
return x
def _residual_v2(self,
x,
in_filter,
out_filter,
stride,
activate_before_residual=False):
"""Residual unit with 2 sub layers with preactivation, plan A shortcut."""
with tf.name_scope('residual_v2') as name_scope:
if activate_before_residual:
x = self._batch_norm(x)
x = self._relu(x)
orig_x = x
else:
orig_x = x
x = self._batch_norm(x)
x = self._relu(x)
x = self._conv(x, 3, out_filter, stride)
x = self._batch_norm(x)
x = self._relu(x)
x = self._conv(x, 3, out_filter, [1, 1, 1, 1])
if in_filter != out_filter:
pad = (out_filter - in_filter) // 2
orig_x = self._avg_pool(orig_x, stride, stride)
if self._data_format == 'channels_first':
orig_x = tf.pad(orig_x, [[0, 0], [pad, pad], [0, 0], [0, 0]])
else:
orig_x = tf.pad(orig_x, [[0, 0], [0, 0], [0, 0], [pad, pad]])
x = tf.add(x, orig_x)
tf.logging.info('image after unit %s: %s', name_scope, x.get_shape())
return x
def _bottleneck_residual_v2(self,
x,
in_filter,
out_filter,
stride,
activate_before_residual=False):
"""Bottleneck residual unit with 3 sub layers, plan B shortcut."""
with tf.name_scope('bottle_residual_v2') as name_scope:
if activate_before_residual:
x = self._batch_norm(x)
x = self._relu(x)
orig_x = x
else:
orig_x = x
x = self._batch_norm(x)
x = self._relu(x)
x = self._conv(x, 1, out_filter // 4, stride, is_atrous=True)
x = self._batch_norm(x)
x = self._relu(x)
# pad when stride isn't unit
x = self._conv(x, 3, out_filter // 4, 1, is_atrous=True)
x = self._batch_norm(x)
x = self._relu(x)
x = self._conv(x, 1, out_filter, 1, is_atrous=True)
if in_filter != out_filter:
orig_x = self._conv(orig_x, 1, out_filter, stride, is_atrous=True)
x = tf.add(x, orig_x)
tf.logging.info('image after unit %s: %s', name_scope, x.get_shape())
return x
def _conv(self, x, kernel_size, filters, strides, is_atrous=False):
"""Convolution."""
padding = 'SAME'
if not is_atrous and strides > 1:
pad = kernel_size - 1
pad_beg = pad // 2
pad_end = pad - pad_beg
if self._data_format == 'channels_first':
x = tf.pad(x, [[0, 0], [0, 0], [pad_beg, pad_end], [pad_beg, pad_end]])
else:
x = tf.pad(x, [[0, 0], [pad_beg, pad_end], [pad_beg, pad_end], [0, 0]])
padding = 'VALID'
return tf.layers.conv2d(
inputs=x,
kernel_size=kernel_size,
filters=filters,
strides=strides,
padding=padding,
use_bias=False,
data_format=self._data_format)
def _batch_norm(self, x):
if self._data_format == 'channels_first':
data_format = 'NCHW'
else:
data_format = 'NHWC'
return tf.contrib.layers.batch_norm(
x,
decay=self._batch_norm_decay,
center=True,
scale=True,
epsilon=self._batch_norm_epsilon,
is_training=self._is_training,
fused=True,
data_format=data_format)
def _relu(self, x):
return tf.nn.relu(x)
def _fully_connected(self, x, out_dim):
with tf.name_scope('fully_connected') as name_scope:
x = tf.layers.dense(x, out_dim)
tf.logging.info('image after unit %s: %s', name_scope, x.get_shape())
return x
def _avg_pool(self, x, pool_size, stride):
with tf.name_scope('avg_pool') as name_scope:
x = tf.layers.average_pooling2d(
x, pool_size, stride, 'SAME', data_format=self._data_format)
tf.logging.info('image after unit %s: %s', name_scope, x.get_shape())
return x
def _global_avg_pool(self, x):
with tf.name_scope('global_avg_pool') as name_scope:
assert x.get_shape().ndims == 4
if self._data_format == 'channels_first':
x = tf.reduce_mean(x, [2, 3])
else:
x = tf.reduce_mean(x, [1, 2])
tf.logging.info('image after unit %s: %s', name_scope, x.get_shape())
return x