File size: 7,798 Bytes
747451d | 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 | # /*---------------------------------------------------------------------------------------------
# * Copyright (c) 2025 STMicroelectronics.
# * All rights reserved.
# *
# * This software is licensed under terms that can be found in the LICENSE file in
# * the root directory of this software component.
# * If no LICENSE file comes with this software, it is provided AS-IS.
# *--------------------------------------------------------------------------------------------*/
import torch
import torch.nn as nn
activation_choice = "relu"
activation = {"relu" : nn.ReLU(inplace=True),
"hswish" : nn.Hardswish(inplace=True) ,
"silu" : nn.SiLU(inplace=True) ,
}
class BasicBlock(nn.Module):
def __init__(self, in_channels, out_channels, stride=1, downsample=None, conv1_custom=None, conv2_custom=None):
super(BasicBlock, self).__init__()
self.conv1 = conv1_custom if conv1_custom is not None else nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=stride, padding=1, bias=False)
self.bn1 = nn.BatchNorm2d(out_channels, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.drop_block = nn.Identity()
self.act1 = activation[activation_choice]
self.aa = nn.Identity()
self.conv2 = conv2_custom if conv2_custom is not None else nn.Conv2d(out_channels, out_channels, kernel_size=3, padding=1, bias=False)
self.bn2 = nn.BatchNorm2d(out_channels)
self.act2 = activation[activation_choice]
self.downsample = downsample
def forward(self, x):
identity = x
out = self.conv1(x)
out = self.bn1(out)
out = self.drop_block(out)
out = self.act1(out)
out = self.aa(out)
out = self.conv2(out)
out = self.bn2(out)
if self.downsample is not None:
identity = self.downsample(x)
out += identity
out = self.act2(out)
return out
class STResNetMicro(nn.Module):
def __init__(
self,
out_features=("dark3", "dark4", "dark5"),
):
super().__init__()
assert out_features, "please provide output features of STResNetMicro"
self.out_features = out_features
self.conv1 = nn.Sequential(
nn.Conv2d(3, 3, kernel_size=1, bias=False),
nn.Conv2d(3, 8, kernel_size=7, stride=2, padding=3, bias=False),
nn.Conv2d(8, 64, kernel_size=1, bias=False)
)
self.bn1 = nn.BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.act1 = activation[activation_choice]
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
self.layer1 = nn.Sequential(
BasicBlock(64, 64),
BasicBlock(64, 64)
)
self.layer2 = nn.Sequential(
BasicBlock(64, 128, stride=2,
downsample=nn.Sequential(
nn.Conv2d(64, 128, kernel_size=1, stride=2, bias=False),
nn.BatchNorm2d(128)
),
conv2_custom=nn.Sequential(
nn.Conv2d(128, 40, kernel_size=1, bias=False),
nn.Conv2d(40, 40, kernel_size=3, padding=1, bias=False),
nn.Conv2d(40, 128, kernel_size=1, bias=False)
)
),
BasicBlock(128, 128,
conv1_custom=nn.Sequential(
nn.Conv2d(128, 88, kernel_size=1, bias=False),
nn.Conv2d(88, 88, kernel_size=3, padding=1, bias=False),
nn.Conv2d(88, 128, kernel_size=1, bias=False)
),
conv2_custom=nn.Sequential(
nn.Conv2d(128, 32, kernel_size=1, bias=False),
nn.Conv2d(32, 32, kernel_size=3, padding=1, bias=False),
nn.Conv2d(32, 128, kernel_size=1, bias=False)
)
)
)
self.layer3 = nn.Sequential(
BasicBlock(128, 256,
stride=2,
downsample=nn.Sequential(
nn.Sequential(
nn.Conv2d(128, 16, kernel_size=1, bias=False),
nn.Conv2d(16, 16, kernel_size=1, stride=2, bias=False),
nn.Conv2d(16, 256, kernel_size=1, bias=False)
),
nn.BatchNorm2d(256)
),
conv1_custom=nn.Sequential(
nn.Conv2d(128, 88, kernel_size=1, bias=False),
nn.Conv2d(88, 88, kernel_size=3, stride=2, padding=1, bias=False),
nn.Conv2d(88, 256, kernel_size=1, bias=False)
),
conv2_custom=nn.Sequential(
nn.Conv2d(256, 72, kernel_size=1, bias=False),
nn.Conv2d(72, 72, kernel_size=3, padding=1, bias=False),
nn.Conv2d(72, 256, kernel_size=1, bias=False)
)
),
BasicBlock(256, 256,
conv1_custom=nn.Sequential(
nn.Conv2d(256, 80, kernel_size=1, bias=False),
nn.Conv2d(80, 80, kernel_size=3, padding=1, bias=False),
nn.Conv2d(80, 256, kernel_size=1, bias=False)
),
conv2_custom=nn.Sequential(
nn.Conv2d(256, 32, kernel_size=1, bias=False),
nn.Conv2d(32, 32, kernel_size=3, padding=1, bias=False),
nn.Conv2d(32, 256, kernel_size=1, bias=False)
)
)
)
self.layer4 = nn.Sequential(
BasicBlock(256, 512,
stride=2,
downsample=nn.Sequential(
nn.Sequential(
nn.Conv2d(256, 24, kernel_size=1, bias=False),
nn.Conv2d(24, 24, kernel_size=1, stride=2, bias=False),
nn.Conv2d(24, 512, kernel_size=1, bias=False)
),
nn.BatchNorm2d(512)
),
conv1_custom=nn.Sequential(
nn.Conv2d(256, 80, kernel_size=1, bias=False),
nn.Conv2d(80, 80, kernel_size=3, stride=2, padding=1, bias=False),
nn.Conv2d(80, 512, kernel_size=1, bias=False)
),
conv2_custom=nn.Sequential(
nn.Conv2d(512, 8, kernel_size=1, bias=False),
nn.Conv2d(8, 8, kernel_size=3, padding=1, bias=False),
nn.Conv2d(8, 512, kernel_size=1, bias=False)
)
),
BasicBlock(512, 512,
conv1_custom=nn.Sequential(
nn.Conv2d(512, 72, kernel_size=1, bias=False),
nn.Conv2d(72, 72, kernel_size=3, padding=1, bias=False),
nn.Conv2d(72, 512, kernel_size=1, bias=False)
),
conv2_custom=nn.Sequential(
nn.Conv2d(512, 64, kernel_size=1, bias=False),
nn.Conv2d(64, 64, kernel_size=3, padding=1, bias=False),
nn.Conv2d(64, 512, kernel_size=1, bias=False)
)
)
)
self.global_pool = nn.AdaptiveAvgPool2d((1, 1))
self.flatten = nn.Flatten()
self.fc = nn.Linear(512, 1000)
def forward(self, x):
x = self.conv1(x)
x = self.bn1(x)
x = self.act1(x)
x = self.maxpool(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.layer4(x)
x = self.global_pool(x)
x = self.flatten(x)
x = self.fc(x)
return x
|