-
Notifications
You must be signed in to change notification settings - Fork 80
/
ai8x_blocks.py
292 lines (259 loc) · 12.6 KB
/
ai8x_blocks.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
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
###################################################################################################
#
# Copyright (C) 2020-2023 Maxim Integrated Products, Inc. All Rights Reserved.
#
# Maxim Integrated Products, Inc. Default Copyright Notice:
# https://www.maximintegrated.com/en/aboutus/legal/copyrights.html
#
###################################################################################################
"""
Contains implementations of popular neural network blocks by taking MAX7800X limits into account.
"""
import torch
from torch import nn
from torch.nn import functional as F
import ai8x
class Fire(nn.Module):
"""
AI8X - Fire Layer
"""
def __init__(self, in_planes, squeeze_planes, expand1x1_planes, expand3x3_planes,
bias=True, **kwargs):
super().__init__()
self.squeeze_layer = ai8x.FusedConv2dReLU(in_channels=in_planes,
out_channels=squeeze_planes, kernel_size=1,
bias=bias, **kwargs)
self.expand1x1_layer = ai8x.FusedConv2dReLU(in_channels=squeeze_planes,
out_channels=expand1x1_planes, kernel_size=1,
bias=bias, **kwargs)
self.expand3x3_layer = ai8x.FusedConv2dReLU(in_channels=squeeze_planes,
out_channels=expand3x3_planes, kernel_size=3,
padding=1, bias=bias, **kwargs)
def forward(self, x): # pylint: disable=arguments-differ
"""Forward prop"""
x = self.squeeze_layer(x)
return torch.cat([self.expand1x1_layer(x), self.expand3x3_layer(x)], 1)
class ResidualBottleneck(nn.Module):
"""
AI8X - Residual Bottleneck Layer.
This module uses ReLU activation not ReLU6 as the original study suggests [1],
because of MAX7800X capabilities.
Args:
in_channels: number of input channels
out_channels: number of output channels
expansion_factor: expansion_factor
stride: stirde size (default=1)
bias: determines if bias used at non-depthwise layers.
depthwise_bias: determines if bias used at depthwise layers.
References:
[1] https://arxiv.org/pdf/1801.04381.pdf (MobileNetV2)
"""
def __init__(self, in_channels, out_channels, expansion_factor, stride=1, bias=False,
depthwise_bias=False, **kwargs):
super().__init__()
self.stride = stride
hidden_channels = int(round(in_channels * expansion_factor))
if hidden_channels == in_channels:
self.conv1 = ai8x.Empty()
else:
self.conv1 = ai8x.FusedConv2dBNReLU(in_channels, hidden_channels, 1, padding=0,
bias=bias, **kwargs)
if stride == 1:
if depthwise_bias:
self.conv2 = ai8x.FusedDepthwiseConv2dBNReLU(hidden_channels, hidden_channels, 3,
padding=1, stride=stride,
bias=depthwise_bias, **kwargs)
else:
self.conv2 = ai8x.FusedDepthwiseConv2dReLU(hidden_channels, hidden_channels, 3,
padding=1, stride=stride,
bias=depthwise_bias, **kwargs)
else:
if depthwise_bias:
self.conv2 = ai8x.FusedMaxPoolDepthwiseConv2dBNReLU(hidden_channels,
hidden_channels,
3, padding=1, pool_size=stride,
pool_stride=stride,
bias=depthwise_bias,
**kwargs)
else:
self.conv2 = ai8x.FusedMaxPoolDepthwiseConv2dReLU(hidden_channels,
hidden_channels,
3, padding=1, pool_size=stride,
pool_stride=stride,
bias=depthwise_bias,
**kwargs)
self.conv3 = ai8x.FusedConv2dBN(hidden_channels, out_channels, 1, bias=bias, **kwargs)
if (stride == 1) and (in_channels == out_channels):
self.resid = ai8x.Add()
else:
self.resid = self.NoResidual()
class NoResidual(nn.Module):
"""
Does nothing.
"""
def forward(self, *x): # pylint: disable=arguments-differ
"""Forward prop"""
return x[0]
def forward(self, x): # pylint: disable=arguments-differ
"""Forward prop"""
y = self.conv1(x)
y = self.conv2(y)
y = self.conv3(y)
return self.resid(y, x)
class ConvResidualBottleneck(nn.Module):
"""
AI8X module based on Residual Bottleneck Layer.
Depthwise convolution is replaced with standard convolution.
This module uses ReLU activation not ReLU6 as the original study suggests [1],
because of MAX7800X capabilities.
Args:
in_channels: number of input channels
out_channels: number of output channels
expansion_factor: expansion_factor
stride: stirde size (default=1)
bias: determines if bias used at non-depthwise layers.
depthwise_bias: determines if bias used at depthwise layers.
References:
[1] https://arxiv.org/pdf/1801.04381.pdf (MobileNetV2)
"""
def __init__(self, in_channels, out_channels, expansion_factor, stride=1, bias=False,
depthwise_bias=False, **kwargs):
super().__init__()
self.stride = stride
hidden_channels = int(round(in_channels * expansion_factor))
if hidden_channels == in_channels:
self.conv1 = ai8x.Empty()
else:
self.conv1 = ai8x.FusedConv2dBNReLU(in_channels, hidden_channels, 1, padding=0,
bias=bias, **kwargs)
if stride == 1:
if depthwise_bias:
self.conv2 = ai8x.FusedConv2dBN(hidden_channels, out_channels, 3,
padding=1, stride=stride,
bias=depthwise_bias, **kwargs)
else:
self.conv2 = ai8x.Conv2d(hidden_channels, out_channels, 3,
padding=1, stride=stride,
bias=depthwise_bias, **kwargs)
else:
if depthwise_bias:
self.conv2 = ai8x.FusedMaxPoolConv2dBN(hidden_channels,
out_channels, 3,
padding=1, pool_size=stride,
pool_stride=stride,
bias=depthwise_bias, **kwargs)
else:
self.conv2 = ai8x.FusedMaxPoolConv2d(hidden_channels,
out_channels, 3,
padding=1, pool_size=stride,
pool_stride=stride,
bias=depthwise_bias, **kwargs)
if (stride == 1) and (in_channels == out_channels):
self.resid = ai8x.Add()
else:
self.resid = self.NoResidual()
class NoResidual(nn.Module):
"""
Does nothing.
"""
def forward(self, *x): # pylint: disable=arguments-differ
"""Forward prop"""
return x[0]
def forward(self, x): # pylint: disable=arguments-differ
"""Forward prop"""
y = self.conv1(x)
y = self.conv2(y)
return self.resid(y, x)
class MBConvBlock(nn.Module):
"""Mobile Inverted Residual Bottleneck Block.
Args:
image_size (tuple or list): [image_height, image_width].
in_channels: number of input channels
out_channels: number of output channels
kernel_size: kernel size (default 3)
stride: stride size (default 1)
se_ratio: squeeze and excitation (SE) ratio (0-1)
expand_ratio: expansion ratio (default 1)
fused: eliminates depthwise convolution layer
References:
[1] https://arxiv.org/pdf/2104.00298.pdf (EfficientNetV2)
"""
def __init__(self,
in_channels,
out_channels,
kernel_size=3,
stride=1,
bias=False,
se_ratio=None,
expand_ratio=1,
fused=False,
**kwargs):
super().__init__()
self.has_se = (se_ratio is not None) and (0 < se_ratio <= 1)
self.in_channels = in_channels
self.out_channels = out_channels
self.stride = stride
self.expand_ratio = expand_ratio
self.fused = fused
# Expansion phase (Inverted Bottleneck)
inp = in_channels # number of input channels
out = in_channels * expand_ratio # number of output channels
if expand_ratio != 1:
if fused is True:
self.expand_conv = ai8x.FusedConv2dBNReLU(inp, out, kernel_size=kernel_size,
padding=1, batchnorm='Affine', bias=bias,
eps=1e-03, momentum=0.01, **kwargs)
else:
self.expand_conv = ai8x.FusedConv2dBNReLU(inp, out, 1,
batchnorm='Affine', bias=bias,
eps=1e-03, momentum=0.01, **kwargs)
# Depthwise Convolution phase
if fused is not True:
self.depthwise_conv = ai8x.FusedConv2dBNReLU(in_channels=out, out_channels=out,
groups=out, # groups makes it depthwise
padding=1, kernel_size=kernel_size,
stride=stride, batchnorm='Affine',
bias=bias, eps=1e-03, momentum=0.01,
**kwargs)
# Squeeze and Excitation phase
if self.has_se:
num_squeezed_channels = max(1, int(in_channels * se_ratio))
self.se_reduce = ai8x.FusedConv2dReLU(in_channels=out,
out_channels=num_squeezed_channels,
kernel_size=1, stride=1, bias=bias, **kwargs)
self.se_expand = ai8x.Conv2d(in_channels=num_squeezed_channels, out_channels=out,
kernel_size=1, stride=1, bias=bias, **kwargs)
# Output Convolution phase
final_out = out_channels
self.project_conv = ai8x.FusedConv2dBN(in_channels=out, out_channels=final_out,
kernel_size=1, batchnorm='Affine', bias=bias,
eps=1e-03, momentum=0.01, **kwargs)
# Skip connection
self.resid = ai8x.Add()
def forward(self, inputs):
"""MBConvBlock's forward function.
Args:
inputs (tensor): Input tensor.
Returns:
Output of this block after processing.
"""
# Expansion Convolution layer
x = inputs
if self.expand_ratio != 1:
x = self.expand_conv(inputs)
# Depthwise Convolution layer
if self.fused is not True:
x = self.depthwise_conv(x)
# Squeeze and Excitation layers
if self.has_se:
x_squeezed = F.adaptive_avg_pool2d(x, 1)
x_squeezed = self.se_reduce(x_squeezed)
x_squeezed = self.se_expand(x_squeezed)
x = torch.sigmoid(x_squeezed) * x
# Output Convolution layer
x = self.project_conv(x)
# Skip connection
input_filters, output_filters = self.in_channels, self.out_channels
if self.stride == 1 and input_filters == output_filters:
x = self.resid(x, inputs)
return x