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
|
"""Mobile inverted residual block."""
from typing import Tuple
import torch
from torch import nn, Tensor
from torch.nn import functional as F
from .utils import get_same_padding_conv2d
class MBConvBlock(nn.Module):
"""Mobile Inverted Residual Bottleneck block."""
def __init__(
self,
in_channels: int,
kernel_size: int,
stride: int,
bn_momentum: float,
bn_eps: float,
se_ratio: float,
id_skip: bool,
expand_ratio: int,
image_size: Tuple[int, int],
) -> None:
super().__init__()
self.kernel_size = kernel_size
self.bn_momentum = bn_momentum
self.bn_eps = bn_eps
self.id_skip = id_skip
self.has_se = se_ratio is not None and 0.0 < se_ratio < 1.0
def _build(self, image_size: Tuple[int, int], in_channels: int, kernel_size: int, stride: int, expand_ratio: int) -> None:
inner_channels = in_channels * expand_ratio
self._inverted_bottleneck = (
self._configure_inverted_bottleneck(
image_size=image_size,
in_channels=in_channels,
out_channels=inner_channels,
)
if expand_ratio != 1
else None
)
self._depthwise = self._configure_depthwise(
image_size=image_size,
in_channels=in_channels,
out_channels=inner_channels,
groups=inner_channels,
kernel_size=kernel_size,
stride=stride,
)
image_size = calculate_output_image_size(image_size, stride)
self._squeeze_excite = (
self._configure_squeeze_excite(
in_channels=inner_channels, out_channels=inner_channels, se_ratio=se_ratio
)
if self.has_se
else None
)
self._pointwise = self._configure_pointwise(
image_size=image_size, in_channels=inner_channels, out_channels=out_channels
)
def _configure_inverted_bottleneck(
self,
image_size: Tuple[int, int],
in_channels: int,
out_channels: int,
) -> nn.Sequential:
"""Expansion phase."""
Conv2d = get_same_padding_conv2d(image_size=image_size)
return nn.Sequential(
Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
bias=False,
),
nn.BatchNorm2d(
num_features=out_channels, momentum=self.bn_momentum, eps=self.bn_eps
),
nn.SiLU(inplace=True),
)
def _configure_depthwise(
self,
image_size: Tuple[int, int],
in_channels: int,
out_channels: int,
groups: int,
kernel_size: int,
stride: int,
) -> nn.Sequential:
Conv2d = get_same_padding_conv2d(image_size=image_size)
return nn.Sequential(
Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
groups=groups,
bias=False,
),
nn.BatchNorm2d(
num_features=out_channels, momentum=self.bn_momentum, eps=self.bn_eps
),
nn.SiLU(inplace=True),
)
def _configure_squeeze_excite(
self, in_channels: int, out_channels: int, se_ratio: float
) -> nn.Sequential:
Conv2d = get_same_padding_conv2d(image_size=(1, 1))
num_squeezed_channels = max(1, int(in_channels * se_ratio))
return nn.Sequential(
Conv2d(
in_channels=in_channels,
out_channels=num_squeezed_channels,
kernel_size=1,
),
nn.SiLU(inplace=True),
Conv2d(
in_channels=num_squeezed_channels,
out_channels=out_channels,
kernel_size=1,
),
)
def _configure_pointwise(
self, image_size: Tuple[int, int], in_channels: int, out_channels: int
) -> nn.Sequential:
Conv2d = get_same_padding_conv2d(image_size=image_size)
return nn.Sequential(
Conv2d(
in_channels=in_channels,
out_channels=out_channels,
kernel_size=1,
bias=False,
),
nn.BatchNorm2d(
num_features=out_channels, momentum=self.bn_momentum, eps=self.bn_eps
),
)
def forward(self, x: Tensor, drop_connection_rate: Optional[float]) -> Tensor:
residual = x
if self._inverted_bottleneck is not None:
x = self._inverted_bottleneck(x)
x = self._depthwise(x)
if self._squeeze_excite is not None:
x_squeezed = F.adaptive_avg_pool2d(x, 1)
x_squeezed = self._squeeze_excite(x)
x = torch.sigmoid(x_squeezed) * x
x = self._pointwise(x)
# Stochastic depth
|
