1.文章原文地址
Deep Residual Learning for Image Recognition
2.文章摘要
神经网络的层次越深越难训练。我们提出了一个残差学习框架来简化网络的训练，这些网络比之前使用的网络都要深的多。我们明确地将层变为学习关于层输入的残差函数，而不是学习未参考的函数。我们提供了综合的实验证据来表明这个残差网络更容易优化，以及通过极大提升网络深度可以获得更好的准确率。在ImageNet数据集上，我们评估了残差网络，该网络有152层，层数是VGG网络的8倍，但是有更低的复杂度。几个残差网络的集成在ImageNet数据集上取得了3.57%错误率。这个结果在ILSVRC2015分类任务上取得第一名的成绩。我们也使用了100和1000层网络用在了数据集CIFAR-10上加以分析。
在许多视觉识别任务中，表征的深度是至关重要的。仅仅通过极端深的表征，我们在COCO目标检测数据集上得到了28%的相对提高。深度残差网络是我们提交到ILSVRC & COCO2015竞赛的网络基础，在这里我们获得了ImageNet检测任务、ImageNet定位任务，COCO检测任务和COCO分割任务的第一名。
3.网络结构
4.Pytorch实现
1 import torch.nn as nn 2 from torch.utils.model_zoo import load_url as load_state_dict_from_url 3 4 5 __all__ = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101', 6 'resnet152', 'resnext50_32x4d', 'resnext101_32x8d'] 7 8 9 model_urls = { 10 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', 11 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', 12 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', 13 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', 14 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', 15 } 16 17 18 def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1): 19 """3x3 convolution with padding""" 20 return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, 21 padding=dilation, groups=groups, bias=False, dilation=dilation) 22 23 24 def conv1x1(in_planes, out_planes, stride=1): 25 """1x1 convolution""" 26 return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False) 27 28 29 class BasicBlock(nn.Module): 30 expansion = 1 31 32 def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1, 33 base_width=64, dilation=1, norm_layer=None): 34 super(BasicBlock, self).__init__() 35 if norm_layer is None: 36 norm_layer = nn.BatchNorm2d 37 if groups != 1 or base_width != 64: 38 raise ValueError('BasicBlock only supports groups=1 and base_width=64') 39 if dilation > 1: 40 raise NotImplementedError("Dilation > 1 not supported in BasicBlock") 41 # Both self.conv1 and self.downsample layers downsample the input when stride != 1 42 self.conv1 = conv3x3(inplanes, planes, stride) 43 self.bn1 = norm_layer(planes) 44 self.relu = nn.ReLU(inplace=True) 45 self.conv2 = conv3x3(planes, planes) 46 self.bn2 = norm_layer(planes) 47 self.downsample = downsample 48 self.stride = stride 49 50 def forward(self, x): 51 identity = x 52 53 out = self.conv1(x) 54 out = self.bn1(out) 55 out = self.relu(out) 56 57 out = self.conv2(out) 58 out = self.bn2(out) 59 60 if self.downsample is not None: 61 identity = self.downsample(x) 62 63 out += identity 64 out = self.relu(out) 65 66 return out 67 68 69 class Bottleneck(nn.Module): 70 expansion = 4 71 72 def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1, 73 base_width=64, dilation=1, norm_layer=None): 74 super(Bottleneck, self).__init__() 75 if norm_layer is None: 76 norm_layer = nn.BatchNorm2d 77 width = int(planes * (base_width / 64.)) * groups 78 # Both self.conv2 and self.downsample layers downsample the input when stride != 1 79 self.conv1 = conv1x1(inplanes, width) 80 self.bn1 = norm_layer(width) 81 self.conv2 = conv3x3(width, width, stride, groups, dilation) 82 self.bn2 = norm_layer(width) 83 self.conv3 = conv1x1(width, planes * self.expansion) 84 self.bn3 = norm_layer(planes * self.expansion) 85 self.relu = nn.ReLU(inplace=True) 86 self.downsample = downsample 87 self.stride = stride 88 89 def forward(self, x): 90 identity = x 91 92 out = self.conv1(x) 93 out = self.bn1(out) 94 out = self.relu(out) 95 96 out = self.conv2(out) 97 out = self.bn2(out) 98 out = self.relu(out) 99 100 out = self.conv3(out) 101 out = self.bn3(out) 102 103 if self.downsample is not None: 104 identity = self.downsample(x) 105 106 out += identity 107 out = self.relu(out) 108 109 return out 110 111 112 class ResNet(nn.Module): 113 114 def __init__(self, block, layers, num_classes=1000, zero_init_residual=False, 115 groups=1, width_per_group=64, replace_stride_with_dilation=None, 116 norm_layer=None): 117 super(ResNet, self).__init__() 118 if norm_layer is None: 119 norm_layer = nn.BatchNorm2d 120 self._norm_layer = norm_layer 121 122 self.inplanes = 64 123 self.dilation = 1 124 if replace_stride_with_dilation is None: 125 # each element in the tuple indicates if we should replace 126 # the 2x2 stride with a dilated convolution instead 127 replace_stride_with_dilation = [False, False, False] 128 if len(replace_stride_with_dilation) != 3: 129 raise ValueError("replace_stride_with_dilation should be None " 130 "or a 3-element tuple, got {}".format(replace_stride_with_dilation)) 131 self.groups = groups 132 self.base_width = width_per_group 133 self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3, 134 bias=False) 135 self.bn1 = norm_layer(self.inplanes) 136 self.relu = nn.ReLU(inplace=True) 137 self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) 138 self.layer1 = self._make_layer(block, 64, layers[0]) 139 self.layer2 = self._make_layer(block, 128, layers[1], stride=2, 140 dilate=replace_stride_with_dilation[0]) 141 self.layer3 = self._make_layer(block, 256, layers[2], stride=2, 142 dilate=replace_stride_with_dilation[1]) 143 self.layer4 = self._make_layer(block, 512, layers[3], stride=2, 144 dilate=replace_stride_with_dilation[2]) 145 self.avgpool = nn.AdaptiveAvgPool2d((1, 1)) 146 self.fc = nn.Linear(512 * block.expansion, num_classes) 147 148 for m in self.modules(): 149 if isinstance(m, nn.Conv2d): 150 nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') 151 elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)): 152 nn.init.constant_(m.weight, 1) 153 nn.init.constant_(m.bias, 0) 154 155 # Zero-initialize the last BN in each residual branch, 156 # so that the residual branch starts with zeros, and each residual block behaves like an identity. 157 # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677 158 if zero_init_residual: 159 for m in self.modules(): 160 if isinstance(m, Bottleneck): 161 nn.init.constant_(m.bn3.weight, 0) 162 elif isinstance(m, BasicBlock): 163 nn.init.constant_(m.bn2.weight, 0) 164 165 def _make_layer(self, block, planes, blocks, stride=1, dilate=False): 166 norm_layer = self._norm_layer 167 downsample = None 168 previous_dilation = self.dilation 169 if dilate: 170 self.dilation *= stride 171 stride = 1 172 if stride != 1 or self.inplanes != planes * block.expansion: 173 downsample = nn.Sequential( 174 conv1x1(self.inplanes, planes * block.expansion, stride), 175 norm_layer(planes * block.expansion), 176 ) 177 178 layers = [] 179 layers.append(block(self.inplanes, planes, stride, downsample, self.groups, 180 self.base_width, previous_dilation, norm_layer)) 181 self.inplanes = planes * block.expansion 182 for _ in range(1, blocks): 183 layers.append(block(self.inplanes, planes, groups=self.groups, 184 base_width=self.base_width, dilation=self.dilation, 185 norm_layer=norm_layer)) 186 187 return nn.Sequential(*layers) 188 189 def forward(self, x): 190 x = self.conv1(x) 191 x = self.bn1(x) 192 x = self.relu(x) 193 x = self.maxpool(x) 194 195 x = self.layer1(x) 196 x = self.layer2(x) 197 x = self.layer3(x) 198 x = self.layer4(x) 199 200 x = self.avgpool(x) 201 x = x.reshape(x.size(0), -1) 202 x = self.fc(x) 203 204 return x 205 206 207 def _resnet(arch, inplanes, planes, pretrained, progress, **kwargs): 208 model = ResNet(inplanes, planes, **kwargs) 209 if pretrained: 210 state_dict = load_state_dict_from_url(model_urls[arch], 211 progress=progress) 212 model.load_state_dict(state_dict) 213 return model 214 215 216 def resnet18(pretrained=False, progress=True, **kwargs): 217 """Constructs a ResNet-18 model. 218 Args: 219 pretrained (bool): If True, returns a model pre-trained on ImageNet 220 progress (bool): If True, displays a progress bar of the download to stderr 221 """ 222 return _resnet('resnet18', BasicBlock, [2, 2, 2, 2], pretrained, progress, 223 **kwargs) 224 225 226 def resnet34(pretrained=False, progress=True, **kwargs): 227 """Constructs a ResNet-34 model. 228 Args: 229 pretrained (bool): If True, returns a model pre-trained on ImageNet 230 progress (bool): If True, displays a progress bar of the download to stderr 231 """ 232 return _resnet('resnet34', BasicBlock, [3, 4, 6, 3], pretrained, progress, 233 **kwargs) 234 235 236 def resnet50(pretrained=False, progress=True, **kwargs): 237 """Constructs a ResNet-50 model. 238 Args: 239 pretrained (bool): If True, returns a model pre-trained on ImageNet 240 progress (bool): If True, displays a progress bar of the download to stderr 241 """ 242 return _resnet('resnet50', Bottleneck, [3, 4, 6, 3], pretrained, progress, 243 **kwargs) 244 245 246 def resnet101(pretrained=False, progress=True, **kwargs): 247 """Constructs a ResNet-101 model. 248 Args: 249 pretrained (bool): If True, returns a model pre-trained on ImageNet 250 progress (bool): If True, displays a progress bar of the download to stderr 251 """ 252 return _resnet('resnet101', Bottleneck, [3, 4, 23, 3], pretrained, progress, 253 **kwargs) 254 255 256 def resnet152(pretrained=False, progress=True, **kwargs): 257 """Constructs a ResNet-152 model. 258 Args: 259 pretrained (bool): If True, returns a model pre-trained on ImageNet 260 progress (bool): If True, displays a progress bar of the download to stderr 261 """ 262 return _resnet('resnet152', Bottleneck, [3, 8, 36, 3], pretrained, progress, 263 **kwargs) 264 265 266 def resnext50_32x4d(**kwargs): 267 kwargs['groups'] = 32 268 kwargs['width_per_group'] = 4 269 return _resnet('resnext50_32x4d', Bottleneck, [3, 4, 6, 3], 270 pretrained=False, progress=True, **kwargs) 271 272 273 def resnext101_32x8d(**kwargs): 274 kwargs['groups'] = 32 275 kwargs['width_per_group'] = 8 276 return _resnet('resnext101_32x8d', Bottleneck, [3, 4, 23, 3], 277 pretrained=False, progress=True, **kwargs)
参考
https://github.com/pytorch/vision/tree/master/torchvision/models