forked from commaai/speedchallenge
-
Notifications
You must be signed in to change notification settings - Fork 1
/
keras_regnet.py
180 lines (144 loc) · 7.69 KB
/
keras_regnet.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
import tensorflow as tf
from keras_senet import squeeze_excite_block
from aa_downsample import BlurPool2D
def x_block(input, num_channels, group_width, stride=1):
if not stride == 1 or not input.shape[-1] == num_channels:
shortcut = tf.keras.layers.Conv2D(num_channels, 1, strides=stride, padding='same')(input)
shortcut = tf.keras.layers.BatchNormalization()(shortcut)
else:
shortcut = input
# Residual
res = tf.keras.layers.Conv2D(num_channels, 1, padding='same')(input)
res = tf.keras.layers.BatchNormalization()(res)
res = tf.keras.layers.Activation(tf.nn.swish)(res)
res = tf.keras.layers.Conv2D(num_channels, 3, groups=num_channels//group_width, strides=stride, padding='same')(res)
res = tf.keras.layers.BatchNormalization()(res)
res = tf.keras.layers.Activation(tf.nn.swish)(res)
res = tf.keras.layers.Conv2D(num_channels, 1, padding='same')(res)
res = tf.keras.layers.BatchNormalization()(res)
# Merge
out = tf.keras.layers.add([res, shortcut])
out = tf.keras.layers.Activation(tf.nn.swish)(out)
return out
def y_block(input, num_channels, group_width, stride=1):
if not stride == 1 or not input.shape[-1] == num_channels:
shortcut = tf.keras.layers.AveragePooling2D()(input)
shortcut = tf.keras.layers.Conv2D(num_channels, 1, padding='same')(shortcut)
shortcut = tf.keras.layers.BatchNormalization()(shortcut)
else:
shortcut = input
# Residual
res = tf.keras.layers.Conv2D(num_channels, 1, padding='same')(input)
res = tf.keras.layers.BatchNormalization()(res)
res = tf.keras.layers.Activation(tf.nn.swish)(res)
res = tf.keras.layers.Conv2D(num_channels, 3, groups=num_channels//group_width, padding='same')(res)
res = tf.keras.layers.BatchNormalization()(res)
res = tf.keras.layers.Activation(tf.nn.swish)(res)
if not stride == 1:
res = BlurPool2D()(res)
res = squeeze_excite_block(res, ratio=4)
res = tf.keras.layers.Conv2D(num_channels, 1, padding='same')(res)
res = tf.keras.layers.BatchNormalization()(res)
# Merge
out = tf.keras.layers.add([res, shortcut])
out = tf.keras.layers.Activation(tf.nn.swish)(out)
return out
def regnety_400mf(image):
encoder_filters = [32, 48, 104, 208, 440]
stride = 2
group_width = 8
conv1 = tf.keras.layers.Conv2D(encoder_filters[0], 3, strides=stride, padding='same')(image)
conv1 = tf.keras.layers.BatchNormalization()(conv1)
conv1 = tf.keras.layers.Activation(tf.nn.swish)(conv1)
conv2 = y_block(conv1, encoder_filters[1], group_width, stride=stride)
conv3 = y_block(conv2, encoder_filters[2], group_width, stride=stride)
conv3 = y_block(conv3, encoder_filters[2], group_width)
conv3 = y_block(conv3, encoder_filters[2], group_width)
conv4 = y_block(conv3, encoder_filters[3], group_width, stride=stride)
conv4 = y_block(conv4, encoder_filters[3], group_width)
conv4 = y_block(conv4, encoder_filters[3], group_width)
conv4 = y_block(conv4, encoder_filters[3], group_width)
conv4 = y_block(conv4, encoder_filters[3], group_width)
conv4 = y_block(conv4, encoder_filters[3], group_width)
conv5 = y_block(conv4, encoder_filters[4], group_width, stride=stride)
conv5 = y_block(conv5, encoder_filters[4], group_width)
conv5 = y_block(conv5, encoder_filters[4], group_width)
conv5 = y_block(conv5, encoder_filters[4], group_width)
conv5 = y_block(conv5, encoder_filters[4], group_width)
conv5 = y_block(conv5, encoder_filters[4], group_width)
return conv5
def regnetx_200mf(image):
encoder_filters = [32, 24, 56, 152, 368]
stride = 2
group_width = 8
conv1 = tf.keras.layers.Conv2D(encoder_filters[0], 3, strides=stride, padding='same')(image)
conv1 = tf.keras.layers.BatchNormalization()(conv1)
conv1 = tf.keras.layers.Activation(tf.nn.swish)(conv1)
conv2 = x_block(conv1, encoder_filters[1], group_width, stride=stride)
conv3 = x_block(conv2, encoder_filters[2], group_width, stride=stride)
conv4 = x_block(conv3, encoder_filters[3], group_width, stride=stride)
conv4 = x_block(conv4, encoder_filters[3], group_width)
conv4 = x_block(conv4, encoder_filters[3], group_width)
conv4 = x_block(conv4, encoder_filters[3], group_width)
conv5 = x_block(conv4, encoder_filters[4], group_width, stride=stride)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
return conv5
def regnetx_400mf(image):
encoder_filters = [32, 32, 64, 160, 384]
stride = 2
group_width = 16
conv1 = tf.keras.layers.Conv2D(encoder_filters[0], 3, strides=stride, padding='same')(image)
conv1 = tf.keras.layers.BatchNormalization()(conv1)
conv1 = tf.keras.layers.Activation(tf.nn.swish)(conv1)
conv2 = x_block(conv1, encoder_filters[1], group_width, stride=stride)
conv3 = x_block(conv2, encoder_filters[2], group_width, stride=stride)
conv3 = x_block(conv3, encoder_filters[2], group_width)
conv4 = x_block(conv3, encoder_filters[3], group_width, stride=stride)
conv4 = x_block(conv4, encoder_filters[3], group_width)
conv4 = x_block(conv4, encoder_filters[3], group_width)
conv4 = x_block(conv4, encoder_filters[3], group_width)
conv4 = x_block(conv4, encoder_filters[3], group_width)
conv4 = x_block(conv4, encoder_filters[3], group_width)
conv4 = x_block(conv4, encoder_filters[3], group_width)
conv5 = x_block(conv4, encoder_filters[4], group_width, stride=stride)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
return conv5
def regnetx_600mf(image):
encoder_filters = [32, 48, 96, 240, 528]
stride = 2
group_width = 24
conv1 = tf.keras.layers.Conv2D(encoder_filters[0], 3, strides=stride, padding='same')(image)
conv1 = tf.keras.layers.BatchNormalization()(conv1)
conv1 = tf.keras.layers.Activation(tf.nn.swish)(conv1)
conv2 = x_block(conv1, encoder_filters[1], group_width, stride=stride)
conv3 = x_block(conv2, encoder_filters[2], group_width, stride=stride)
conv3 = x_block(conv3, encoder_filters[2], group_width)
conv3 = x_block(conv3, encoder_filters[2], group_width)
conv4 = x_block(conv3, encoder_filters[3], group_width, stride=stride)
conv4 = x_block(conv4, encoder_filters[3], group_width)
conv4 = x_block(conv4, encoder_filters[3], group_width)
conv4 = x_block(conv4, encoder_filters[3], group_width)
conv4 = x_block(conv4, encoder_filters[3], group_width)
conv5 = x_block(conv4, encoder_filters[4], group_width, stride=stride)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
conv5 = x_block(conv5, encoder_filters[4], group_width)
return conv5