-
Notifications
You must be signed in to change notification settings - Fork 0
/
resnet_functional.py
266 lines (226 loc) · 12.4 KB
/
resnet_functional.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
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
"""
File name: resnet_functional.py
Author: Benjamin Planche
Date created: 26.03.2019
Date last modified: 18:56 26.03.2019
Python Version: "3.6"
Copyright = "Copyright (C) 2018-2019 of Packt"
Credits = ["Eliot Andres, Benjamin Planche"]
License = "MIT"
Version = "1.0.0"
Maintainer = "non"
Status = "Prototype" # "Prototype", "Development", or "Production"
"""
#==============================================================================
# Imported Modules
#==============================================================================
import tensorflow as tf
from tensorflow.keras.models import Model
from tensorflow.keras.layers import (
Input, Activation, Dense, Flatten, Conv2D, MaxPooling2D,
GlobalAveragePooling2D, AveragePooling2D, BatchNormalization, add)
import tensorflow.keras.regularizers as regulizers
#==============================================================================
# Function Definitions
#==============================================================================
def _res_conv(filters, kernel_size=3, padding='same', strides=1, use_relu=True, use_bias=False, name='cbr',
kernel_initializer='he_normal', kernel_regularizer=regulizers.l2(1e-4)):
"""
Return a layer block chaining conv, batchnrom and reLU activation.
:param filters: Number of filters.
:param kernel_size: Kernel size.
:param padding: Convolution padding.
:param strides: Convolution strides.
:param use_relu: Flag to apply ReLu activation at the end.
:param use_bias: Flag to use bias or not in Conv layer.
:param name: Name suffix for the layers.
:param kernel_initializer: Kernel initialisation method name.
:param kernel_regularizer: Kernel regularizer.
:return: Callable layer block
"""
def layer_fn(x):
conv = Conv2D(
filters=filters, kernel_size=kernel_size, padding=padding, strides=strides, use_bias=use_bias,
kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer,
name=name + '_c')(x)
res = BatchNormalization(axis=-1, name=name + '_bn')(conv)
if use_relu:
res = Activation("relu", name=name + '_r')(res)
return res
return layer_fn
def _merge_with_shortcut(kernel_initializer='he_normal', kernel_regularizer=regulizers.l2(1e-4),
name='block'):
"""
Return a layer block which merge an input tensor and the corresponding
residual output tensor from another branch.
:param kernel_initializer: Kernel initialisation method name.
:param kernel_regularizer: Kernel regularizer.
:param name: Name suffix for the layers.
:return: Callable layer block
"""
def layer_fn(x, x_residual):
# We check if `x_residual` was scaled down. If so, we scale `x` accordingly with a 1x1 conv:
x_shape = tf.keras.backend.int_shape(x)
x_residual_shape = tf.keras.backend.int_shape(x_residual)
if x_shape == x_residual_shape:
shortcut = x
else:
strides = (
int(round(x_shape[1] / x_residual_shape[1])), # vertical stride
int(round(x_shape[2] / x_residual_shape[2])) # horizontal stride
)
x_residual_channels = x_residual_shape[3]
shortcut = Conv2D(
filters=x_residual_channels, kernel_size=(1, 1), padding="valid", strides=strides,
kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer,
name=name + '_shortcut_c')(x)
merge = add([shortcut, x_residual])
return merge
return layer_fn
def _residual_block_basic(filters, kernel_size=3, strides=1, use_bias=False, name='res_basic',
kernel_initializer='he_normal', kernel_regularizer=regulizers.l2(1e-4)):
"""
Return a basic residual layer block.
:param filters: Number of filters.
:param kernel_size: Kernel size.
:param strides: Convolution strides
:param use_bias: Flag to use bias or not in Conv layer.
:param kernel_initializer: Kernel initialisation method name.
:param kernel_regularizer: Kernel regularizer.
:return: Callable layer block
"""
def layer_fn(x):
x_conv1 = _res_conv(
filters=filters, kernel_size=kernel_size, padding='same', strides=strides,
use_relu=True, use_bias=use_bias,
kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer,
name=name + '_cbr_1')(x)
x_residual = _res_conv(
filters=filters, kernel_size=kernel_size, padding='same', strides=1,
use_relu=False, use_bias=use_bias,
kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer,
name=name + '_cbr_2')(x_conv1)
merge = _merge_with_shortcut(kernel_initializer, kernel_regularizer,name=name)(x, x_residual)
merge = Activation('relu')(merge)
return merge
return layer_fn
def _residual_block_bottleneck(filters, kernel_size=3, strides=1, use_bias=False, name='res_bottleneck',
kernel_initializer='he_normal', kernel_regularizer=regulizers.l2(1e-4)):
"""
Return a residual layer block with bottleneck, recommended for deep ResNets (depth > 34).
:param filters: Number of filters.
:param kernel_size: Kernel size.
:param strides: Convolution strides
:param use_bias: Flag to use bias or not in Conv layer.
:param kernel_initializer: Kernel initialisation method name.
:param kernel_regularizer: Kernel regularizer.
:return: Callable layer block
"""
def layer_fn(x):
x_bottleneck = _res_conv(
filters=filters, kernel_size=1, padding='valid', strides=strides,
use_relu=True, use_bias=use_bias,
kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer,
name=name + '_cbr1')(x)
x_conv = _res_conv(
filters=filters, kernel_size=kernel_size, padding='same', strides=1,
use_relu=True, use_bias=use_bias,
kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer,
name=name + '_cbr2')(x_bottleneck)
x_residual = _res_conv(
filters=filters * 4, kernel_size=1, padding='valid', strides=1,
use_relu=False, use_bias=use_bias,
kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer,
name=name + '_cbr3')(x_conv)
merge = _merge_with_shortcut(kernel_initializer, kernel_regularizer, name=name)(x, x_residual)
merge = Activation('relu')(merge)
return merge
return layer_fn
def _residual_macroblock(block_fn, filters, repetitions=3, kernel_size=3, strides_1st_block=1, use_bias=False,
kernel_initializer='he_normal', kernel_regularizer=regulizers.l2(1e-4),
name='res_macroblock'):
"""
Return a layer block, composed of a repetition of `N` residual blocks.
:param block_fn: Block layer method to be used.
:param repetitions: Number of times the block should be repeated inside.
:param filters: Number of filters.
:param kernel_size: Kernel size.
:param strides_1st_block: Convolution strides for the 1st block.
:param use_bias: Flag to use bias or not in Conv layer.
:param kernel_initializer: Kernel initialisation method name.
:param kernel_regularizer: Kernel regularizer.
:return: Callable layer block
"""
def layer_fn(x):
for i in range(repetitions):
block_name = "{}_{}".format(name, i)
strides = strides_1st_block if i == 0 else 1
x = block_fn(filters=filters, kernel_size=kernel_size,
strides=strides, use_bias=use_bias,
kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer,
name=block_name)(x)
return x
return layer_fn
def ResNet(input_shape, num_classes=1000, block_fn=_residual_block_basic, repetitions=(2, 2, 2, 2),
use_bias=False, kernel_initializer='he_normal', kernel_regularizer=regulizers.l2(1e-4)):
"""
Build a ResNet model for classification.
:param input_shape: Input shape (e.g. (224, 224, 3))
:param num_classes: Number of classes to predict
:param block_fn: Block layer method to be used.
:param repetitions: List of repetitions for each macro-blocks the network should contain.
:param use_bias: Flag to use bias or not in Conv layer.
:param kernel_initializer: Kernel initialisation method name.
:param kernel_regularizer: Kernel regularizer.
:return: ResNet model.
"""
# Input and 1st layers:
inputs = Input(shape=input_shape)
conv = _res_conv(
filters=64, kernel_size=7, strides=2, use_relu=True, use_bias=use_bias,
kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer)(inputs)
maxpool = MaxPooling2D(pool_size=3, strides=2, padding='same')(conv)
# Chain of residual blocks:
filters = 64
strides = 2
res_block = maxpool
for i, repet in enumerate(repetitions):
# We do not further reduce the input size for the 1st block (max-pool applied just before):
block_strides = strides if i != 0 else 1
macroblock_name = "block_{}".format(i)
res_block = _residual_macroblock(
block_fn=block_fn, repetitions=repet, name=macroblock_name,
filters=filters, strides_1st_block=block_strides, use_bias=use_bias,
kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer)(res_block)
filters = min(filters * 2, 1024) # we limit to 1024 filters max
# Final layers for prediction:
res_spatial_dim = tf.keras.backend.int_shape(res_block)[1:3]
avg_pool = AveragePooling2D(pool_size=res_spatial_dim, strides=1)(res_block)
flatten = Flatten()(avg_pool)
predictions = Dense(units=num_classes, kernel_initializer=kernel_initializer,
activation='softmax')(flatten)
# Model:
model = Model(inputs=inputs, outputs=predictions)
return model
def ResNet18(input_shape, num_classes=1000, use_bias=True,
kernel_initializer='he_normal', kernel_regularizer=None):
return ResNet(input_shape, num_classes, block_fn=_residual_block_basic, repetitions=(2, 2, 2, 2),
use_bias=use_bias, kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer)
def ResNet34(input_shape, num_classes=1000, use_bias=True,
kernel_initializer='he_normal', kernel_regularizer=None):
return ResNet(input_shape, num_classes, block_fn=_residual_block_basic, repetitions=(3, 4, 6, 3),
use_bias=use_bias, kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer)
def ResNet50(input_shape, num_classes=1000, use_bias=True,
kernel_initializer='he_normal', kernel_regularizer=None):
# Note: ResNet50 is similar to ResNet34,
# with the basic blocks replaced by bottleneck ones (3 conv layers each instead of 2)
return ResNet(input_shape, num_classes, block_fn=_residual_block_bottleneck, repetitions=(3, 4, 6, 3),
use_bias=use_bias, kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer)
def ResNet101(input_shape, num_classes=1000, use_bias=True,
kernel_initializer='he_normal', kernel_regularizer=None):
return ResNet(input_shape, num_classes, block_fn=_residual_block_bottleneck, repetitions=(3, 4, 23, 3),
use_bias=use_bias, kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer)
def ResNet152(input_shape, num_classes=1000, use_bias=True,
kernel_initializer='he_normal', kernel_regularizer=None):
return ResNet(input_shape, num_classes, block_fn=_residual_block_bottleneck, repetitions=(3, 8, 36, 3),
use_bias=use_bias, kernel_initializer=kernel_initializer, kernel_regularizer=kernel_regularizer)