-
Notifications
You must be signed in to change notification settings - Fork 7
/
mobilenetv3.py
142 lines (131 loc) · 7.63 KB
/
mobilenetv3.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
import tensorflow as tf
class Bneck(tf.keras.layers.Layer):
def __init__(self,
filters,
expansion_filters,
kernel_size,
alpha=1.0,
strides=(1, 1),
use_se=False,
activation=tf.nn.relu6,
**kwargs):
super(Bneck, self).__init__(**kwargs)
self.filters = _make_divisible(filters * alpha, 8)
self.expansion_filters = expansion_filters
self.kernel_size = kernel_size
self.strides = strides
self.use_se = use_se
self.activation = activation
self.expand_conv2d = tf.keras.layers.Conv2D(self.expansion_filters, 1, padding='same', use_bias=False,kernel_regularizer=tf.keras.regularizers.l2(1e-5))
self.expand_bn = tf.keras.layers.BatchNormalization()
self.zero_padding2d = tf.keras.layers.ZeroPadding2D(((self.kernel_size - 1) // 2, (self.kernel_size - 1) // 2))
self.depthwise_conv2d = tf.keras.layers.DepthwiseConv2D(self.kernel_size, strides=self.strides, use_bias=False,
padding='same' if self.strides == 1 else 'valid',depthwise_regularizer=tf.keras.regularizers.l2(1e-5))
self.depthwise_bn = tf.keras.layers.BatchNormalization()
self.se = SeBlock()
self.project_conv2d = tf.keras.layers.Conv2D(self.filters, kernel_size=1, padding='same', use_bias=False,kernel_regularizer=tf.keras.regularizers.l2(1e-5))
self.project_bn = tf.keras.layers.BatchNormalization()
self.add = tf.keras.layers.Add()
def build(self, input_shape):
self.in_channels = input_shape[-1]
self.built = True
def call(self, inputs):
x = self.expand_conv2d(inputs)
x = self.expand_bn(x)
x = self.activation(x)
if self.strides == 2:
x = self.zero_padding2d(x)
x = self.depthwise_conv2d(x)
x = self.depthwise_bn(x)
if self.use_se:
x = self.se(x)
x = self.activation(x)
x = self.project_conv2d(x)
x = self.project_bn(x)
if self.in_channels == self.filters and self.strides == 1:
x = self.add([inputs, x])
return x
class SeBlock(tf.keras.layers.Layer):
def __init__(self, reduction=4,**kwargs):
super(SeBlock,self).__init__(**kwargs)
self.reduction = reduction
def build(self, input_shape):
self.average_pool = tf.keras.layers.AveragePooling2D((int(input_shape[1]),int(input_shape[2])))
self.dense1 = tf.keras.layers.Dense(int(input_shape[-1]) // self.reduction, use_bias=False,activation=tf.nn.relu)
self.dense2 = tf.keras.layers.Dense(int(input_shape[-1]), use_bias=False,activation=tf.keras.activations.hard_sigmoid)
self.built = True
def call(self, inputs):
x = self.average_pool(inputs)
x = self.dense1(x)
x = self.dense2(x)
return inputs *x
def h_swish(inputs):
return inputs * tf.nn.relu6(inputs + 3) / 6
class HSwish(tf.keras.layers.Layer):
def call(self, inputs):
return h_swish(inputs)
def _make_divisible(v, divisor, min_value=None):
if min_value is None:
min_value = divisor
new_v = max(min_value, int(v + divisor / 2) // divisor * divisor)
# Make sure that round down does not go down by more than 10%.
if new_v < 0.9 * v:
new_v += divisor
return new_v
def MobilenetV3(input_shape,num_classes, size="large", include_top=True,alpha=1.0):
input = tf.keras.layers.Input([*input_shape, 3])
first_block_filters = _make_divisible(16 * alpha, 8)
if size not in ['large', 'small']:
raise ValueError('size should be large or small')
if size == "large":
x = tf.keras.layers.Conv2D(first_block_filters, 3, strides=2, padding='same', use_bias=False,kernel_regularizer=tf.keras.regularizers.l2(1e-5))(input)
x = tf.keras.layers.BatchNormalization()(x)
x = HSwish()(x)
x = Bneck(16, 16, 3, alpha=alpha, strides=1, use_se=False, activation=tf.nn.relu6)(x)
x = Bneck(24, 64, 3, alpha=alpha, strides=2, use_se=False, activation=tf.nn.relu6)(x)
x = Bneck(24, 72, 3, alpha=alpha, strides=1, use_se=False, activation=tf.nn.relu6)(x)
x = Bneck(40, 72, 5, alpha=alpha, strides=2, use_se=True, activation=tf.nn.relu6)(x)
x = Bneck(40, 120, 5, alpha=alpha, strides=1, use_se=True, activation=tf.nn.relu6)(x)
x = Bneck(40, 120, 5, alpha=alpha, strides=1, use_se=True, activation=tf.nn.relu6)(x)
x = Bneck(80, 240, 3, alpha=alpha, strides=2, use_se=False, activation=h_swish)(x)
x = Bneck(80, 200, 3, alpha=alpha, strides=1, use_se=False, activation=h_swish)(x)
x = Bneck(80, 184, 3, alpha=alpha, strides=1, use_se=False, activation=h_swish)(x)
x = Bneck(80, 184, 3, alpha=alpha, strides=1, use_se=False, activation=h_swish)(x)
x = Bneck(112, 480, 3, alpha=alpha, strides=1, use_se=True, activation=h_swish)(x)
x = Bneck(112, 672, 3, alpha=alpha, strides=1, use_se=True, activation=h_swish)(x)
x = Bneck(160, 672, 5, alpha=alpha, strides=2, use_se=True, activation=h_swish)(x)
x = Bneck(160, 960, 5, alpha=alpha, strides=1, use_se=True, activation=h_swish)(x)
x = Bneck(160, 960, 5, alpha=alpha, strides=1, use_se=True, activation=h_swish)(x)
x = tf.keras.layers.Conv2D(_make_divisible(960 * alpha, 8), 1, use_bias=False,kernel_regularizer=tf.keras.regularizers.l2(1e-5))(x)
x = tf.keras.layers.BatchNormalization()(x)
output = HSwish()(x)
else:
x = tf.keras.layers.Conv2D(first_block_filters, 3, strides=2, padding='same', use_bias=False,kernel_regularizer=tf.keras.regularizers.l2(1e-5))(input)
x = tf.keras.layers.BatchNormalization()(x)
x = HSwish()(x)
x = Bneck(16, 16, 3, alpha=alpha, strides=2, use_se=True, activation=tf.nn.relu6)(x)
x = Bneck(24, 72, 3, alpha=alpha, strides=2, use_se=False, activation=tf.nn.relu6)(x)
x = Bneck(24, 88, 3, alpha=alpha, strides=1, use_se=False, activation=tf.nn.relu6)(x)
x = Bneck(40, 96, 5, alpha=alpha, strides=2, use_se=True, activation=h_swish)(x)
x = Bneck(40, 240, 5, alpha=alpha, strides=1, use_se=True, activation=h_swish)(x)
x = Bneck(40, 240, 5, alpha=alpha, strides=1, use_se=True, activation=h_swish)(x)
x = Bneck(48, 120, 5, alpha=alpha, strides=1, use_se=True, activation=h_swish)(x)
x = Bneck(48, 144, 5, alpha=alpha, strides=1, use_se=True, activation=h_swish)(x)
x = Bneck(96, 288, 5, alpha=alpha, strides=2, use_se=True, activation=h_swish)(x)
x = Bneck(96, 576, 5, alpha=alpha, strides=1, use_se=True, activation=h_swish)(x)
x = Bneck(96, 576, 5, alpha=alpha, strides=1, use_se=True, activation=h_swish)(x)
x = tf.keras.layers.Conv2D(_make_divisible(576 * alpha, 8), 1, use_bias=False,kernel_regularizer=tf.keras.regularizers.l2(1e-5))(x)
x = tf.keras.layers.BatchNormalization()(x)
x=SeBlock()(x)
output = HSwish()(x)
if include_top:
output = tf.keras.layers.AveragePooling2D(pool_size=x.shape[1:3])(output)
if alpha > 1.0:
last_block_filters = _make_divisible(1280 * alpha, 8)
else:
last_block_filters = 1280
output = tf.keras.layers.Conv2D(last_block_filters,1, use_bias=False,activation=h_swish,kernel_regularizer=tf.keras.regularizers.l2(1e-5))(output)
output = tf.keras.layers.Dropout(0.8)(output)
output = tf.keras.layers.Conv2D(num_classes,1, use_bias=True,activation=tf.keras.activations.softmax,kernel_regularizer=tf.keras.regularizers.l2(1e-5))(output)
output = tf.keras.layers.Flatten()(output)
return tf.keras.Model(input,output)