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
267
|
"""Mobile inverted residual block."""
from typing import Optional, Tuple, Union
import torch
from torch import nn, Tensor
import torch.nn.functional as F
from text_recognizer.networks.efficientnet.utils import stochastic_depth
class BaseModule(nn.Module):
"""Base sub module class."""
def __init__(self, bn_momentum: float, bn_eps: float) -> None:
super().__init__()
self.bn_momentum = bn_momentum
self.bn_eps = bn_eps
self._build()
def _build(self) -> None:
pass
def forward(self, x: Tensor) -> Tensor:
"""Forward pass."""
return self.block(x)
class InvertedBottleneck(BaseModule):
"""Inverted bottleneck module."""
def __init__(
self,
bn_momentum: float,
bn_eps: float,
in_channels: int,
out_channels: int,
) -> None:
self.in_channels = in_channels
self.out_channels = out_channels
super().__init__(bn_momentum, bn_eps)
def _build(self) -> None:
self.block = nn.Sequential(
nn.Conv2d(
in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=1,
bias=False,
),
nn.BatchNorm2d(
num_features=self.out_channels,
momentum=self.bn_momentum,
eps=self.bn_eps,
),
nn.Mish(inplace=True),
)
class Depthwise(BaseModule):
"""Depthwise convolution module."""
def __init__(
self,
bn_momentum: float,
bn_eps: float,
channels: int,
kernel_size: int,
stride: int,
) -> None:
self.channels = channels
self.kernel_size = kernel_size
self.stride = stride
super().__init__(bn_momentum, bn_eps)
def _build(self) -> None:
self.block = nn.Sequential(
nn.Conv2d(
in_channels=self.channels,
out_channels=self.channels,
kernel_size=self.kernel_size,
stride=self.stride,
groups=self.channels,
bias=False,
),
nn.BatchNorm2d(
num_features=self.channels, momentum=self.bn_momentum, eps=self.bn_eps
),
nn.Mish(inplace=True),
)
class SqueezeAndExcite(BaseModule):
"""Sequeeze and excite module."""
def __init__(
self,
bn_momentum: float,
bn_eps: float,
in_channels: int,
channels: int,
se_ratio: float,
) -> None:
self.in_channels = in_channels
self.channels = channels
self.se_ratio = se_ratio
super().__init__(bn_momentum, bn_eps)
def _build(self) -> None:
num_squeezed_channels = max(1, int(self.in_channels * self.se_ratio))
self.block = nn.Sequential(
nn.Conv2d(
in_channels=self.channels,
out_channels=num_squeezed_channels,
kernel_size=1,
),
nn.Mish(inplace=True),
nn.Conv2d(
in_channels=num_squeezed_channels,
out_channels=self.channels,
kernel_size=1,
),
)
class Pointwise(BaseModule):
"""Pointwise module."""
def __init__(
self,
bn_momentum: float,
bn_eps: float,
in_channels: int,
out_channels: int,
) -> None:
self.in_channels = in_channels
self.out_channels = out_channels
super().__init__(bn_momentum, bn_eps)
def _build(self) -> None:
self.block = nn.Sequential(
nn.Conv2d(
in_channels=self.in_channels,
out_channels=self.out_channels,
kernel_size=1,
bias=False,
),
nn.BatchNorm2d(
num_features=self.out_channels,
momentum=self.bn_momentum,
eps=self.bn_eps,
),
)
class MBConvBlock(nn.Module):
"""Mobile Inverted Residual Bottleneck block."""
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: Tuple[int, int],
stride: Tuple[int, int],
bn_momentum: float,
bn_eps: float,
se_ratio: float,
expand_ratio: int,
) -> None:
super().__init__()
self.in_channels = in_channels
self.out_channels = out_channels
self.kernel_size = kernel_size
self.stride = stride
self.bn_momentum = bn_momentum
self.bn_eps = bn_eps
self.se_ratio = se_ratio
self.pad = self._configure_padding()
self.expand_ratio = expand_ratio
self._inverted_bottleneck: Optional[InvertedBottleneck]
self._depthwise: nn.Sequential
self._squeeze_excite: nn.Sequential
self._pointwise: nn.Sequential
self._build()
def _configure_padding(self) -> Tuple[int, int, int, int]:
"""Set padding for convolutional layers."""
if self.stride == (2, 2):
return (
(self.kernel_size - 1) // 2 - 1,
(self.kernel_size - 1) // 2,
) * 2
return ((self.kernel_size - 1) // 2,) * 4
def _build(self) -> None:
has_se = self.se_ratio is not None and 0.0 < self.se_ratio < 1.0
inner_channels = self.in_channels * self.expand_ratio
self._inverted_bottleneck = (
InvertedBottleneck(
in_channels=self.in_channels,
out_channels=inner_channels,
bn_momentum=self.bn_momentum,
bn_eps=self.bn_eps,
)
if self.expand_ratio != 1
else None
)
self._depthwise = Depthwise(
channels=inner_channels,
kernel_size=self.kernel_size,
stride=self.stride,
bn_momentum=self.bn_momentum,
bn_eps=self.bn_eps,
)
self._squeeze_excite = (
SqueezeAndExcite(
in_channels=self.in_channels,
channels=inner_channels,
se_ratio=self.se_ratio,
bn_momentum=self.bn_momentum,
bn_eps=self.bn_eps,
)
if has_se
else None
)
self._pointwise = Pointwise(
in_channels=inner_channels,
out_channels=self.out_channels,
bn_momentum=self.bn_momentum,
bn_eps=self.bn_eps,
)
def _stochastic_depth(
self, x: Tensor, residual: Tensor, stochastic_dropout_rate: Optional[float]
) -> Tensor:
if self.stride == (1, 1) and self.in_channels == self.out_channels:
if stochastic_dropout_rate:
x = stochastic_depth(
x, p=stochastic_dropout_rate, training=self.training
)
x += residual
return x
def forward(
self, x: Tensor, stochastic_dropout_rate: Optional[float] = None
) -> Tensor:
"""Forward pass."""
residual = x
if self._inverted_bottleneck is not None:
x = self._inverted_bottleneck(x)
x = F.pad(x, self.pad)
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.tanh(F.softplus(x_squeezed)) * x
x = self._pointwise(x)
# Stochastic depth
x = self._stochastic_depth(x, residual, stochastic_dropout_rate)
return x
|
