File size: 6,882 Bytes
854f0d0 |
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 |
import torch.nn as nn
import torch.nn.functional as F
def conv1x1(in_planes, out_planes, stride=1):
"""1x1 convolution without padding"""
return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, padding=0, bias=False)
def conv3x3(in_planes, out_planes, stride=1):
"""3x3 convolution with padding"""
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False)
class BasicBlock(nn.Module):
def __init__(self, in_planes, planes, stride=1):
super().__init__()
self.conv1 = conv3x3(in_planes, planes, stride)
self.conv2 = conv3x3(planes, planes)
self.bn1 = nn.BatchNorm2d(planes)
self.bn2 = nn.BatchNorm2d(planes)
self.relu = nn.ReLU(inplace=True)
if stride == 1:
self.downsample = None
else:
self.downsample = nn.Sequential(
conv1x1(in_planes, planes, stride=stride),
nn.BatchNorm2d(planes)
)
def forward(self, x):
y = x
y = self.relu(self.bn1(self.conv1(y)))
y = self.bn2(self.conv2(y))
if self.downsample is not None:
x = self.downsample(x)
return self.relu(x+y)
class ResNetFPN_8_2(nn.Module):
"""
ResNet+FPN, output resolution are 1/8 and 1/2.
Each block has 2 layers.
"""
def __init__(self, config):
super().__init__()
# Config
block = BasicBlock
initial_dim = config['initial_dim']
block_dims = config['block_dims']
# Class Variable
self.in_planes = initial_dim
# Networks
self.conv1 = nn.Conv2d(1, initial_dim, kernel_size=7, stride=2, padding=3, bias=False)
self.bn1 = nn.BatchNorm2d(initial_dim)
self.relu = nn.ReLU(inplace=True)
self.layer1 = self._make_layer(block, block_dims[0], stride=1) # 1/2
self.layer2 = self._make_layer(block, block_dims[1], stride=2) # 1/4
self.layer3 = self._make_layer(block, block_dims[2], stride=2) # 1/8
# 3. FPN upsample
self.layer3_outconv = conv1x1(block_dims[2], block_dims[2])
self.layer2_outconv = conv1x1(block_dims[1], block_dims[2])
self.layer2_outconv2 = nn.Sequential(
conv3x3(block_dims[2], block_dims[2]),
nn.BatchNorm2d(block_dims[2]),
nn.LeakyReLU(),
conv3x3(block_dims[2], block_dims[1]),
)
self.layer1_outconv = conv1x1(block_dims[0], block_dims[1])
self.layer1_outconv2 = nn.Sequential(
conv3x3(block_dims[1], block_dims[1]),
nn.BatchNorm2d(block_dims[1]),
nn.LeakyReLU(),
conv3x3(block_dims[1], block_dims[0]),
)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def _make_layer(self, block, dim, stride=1):
layer1 = block(self.in_planes, dim, stride=stride)
layer2 = block(dim, dim, stride=1)
layers = (layer1, layer2)
self.in_planes = dim
return nn.Sequential(*layers)
def forward(self, x):
# ResNet Backbone
x0 = self.relu(self.bn1(self.conv1(x)))
x1 = self.layer1(x0) # 1/2
x2 = self.layer2(x1) # 1/4
x3 = self.layer3(x2) # 1/8
# FPN
x3_out = self.layer3_outconv(x3)
x3_out_2x = F.interpolate(x3_out, scale_factor=2., mode='bilinear', align_corners=True)
x2_out = self.layer2_outconv(x2)
x2_out = self.layer2_outconv2(x2_out+x3_out_2x)
x2_out_2x = F.interpolate(x2_out, scale_factor=2., mode='bilinear', align_corners=True)
x1_out = self.layer1_outconv(x1)
x1_out = self.layer1_outconv2(x1_out+x2_out_2x)
return [x3_out, x1_out]
class ResNetFPN_16_4(nn.Module):
"""
ResNet+FPN, output resolution are 1/16 and 1/4.
Each block has 2 layers.
"""
def __init__(self, config):
super().__init__()
# Config
block = BasicBlock
initial_dim = config['initial_dim']
block_dims = config['block_dims']
# Class Variable
self.in_planes = initial_dim
# Networks
self.conv1 = nn.Conv2d(1, initial_dim, kernel_size=7, stride=2, padding=3, bias=False)
self.bn1 = nn.BatchNorm2d(initial_dim)
self.relu = nn.ReLU(inplace=True)
self.layer1 = self._make_layer(block, block_dims[0], stride=1) # 1/2
self.layer2 = self._make_layer(block, block_dims[1], stride=2) # 1/4
self.layer3 = self._make_layer(block, block_dims[2], stride=2) # 1/8
self.layer4 = self._make_layer(block, block_dims[3], stride=2) # 1/16
# 3. FPN upsample
self.layer4_outconv = conv1x1(block_dims[3], block_dims[3])
self.layer3_outconv = conv1x1(block_dims[2], block_dims[3])
self.layer3_outconv2 = nn.Sequential(
conv3x3(block_dims[3], block_dims[3]),
nn.BatchNorm2d(block_dims[3]),
nn.LeakyReLU(),
conv3x3(block_dims[3], block_dims[2]),
)
self.layer2_outconv = conv1x1(block_dims[1], block_dims[2])
self.layer2_outconv2 = nn.Sequential(
conv3x3(block_dims[2], block_dims[2]),
nn.BatchNorm2d(block_dims[2]),
nn.LeakyReLU(),
conv3x3(block_dims[2], block_dims[1]),
)
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
def _make_layer(self, block, dim, stride=1):
layer1 = block(self.in_planes, dim, stride=stride)
layer2 = block(dim, dim, stride=1)
layers = (layer1, layer2)
self.in_planes = dim
return nn.Sequential(*layers)
def forward(self, x):
# ResNet Backbone
x0 = self.relu(self.bn1(self.conv1(x)))
x1 = self.layer1(x0) # 1/2
x2 = self.layer2(x1) # 1/4
x3 = self.layer3(x2) # 1/8
x4 = self.layer4(x3) # 1/16
# FPN
x4_out = self.layer4_outconv(x4)
x4_out_2x = F.interpolate(x4_out, scale_factor=2., mode='bilinear', align_corners=True)
x3_out = self.layer3_outconv(x3)
x3_out = self.layer3_outconv2(x3_out+x4_out_2x)
x3_out_2x = F.interpolate(x3_out, scale_factor=2., mode='bilinear', align_corners=True)
x2_out = self.layer2_outconv(x2)
x2_out = self.layer2_outconv2(x2_out+x3_out_2x)
return [x4_out, x2_out]
|