ResNet

代码：link

class ResNet的forward，source code
值得注意的是，self.maxpool中ks=3

ResNet34

(224, 224, 3)

→【self.conv1, Cout=64, ks=7, s=2, p=3】→(112, 112, 64)
→【self.bn1】→【self.relu】→(112, 112, 64)
→【self.maxpool, ks=3, s=2, p=1】→(56, 56, 64)

→【self.layer1】→(56, 56, 64)
→【self.layer2】→(28, 28, 128)
→【self.layer3】→(14, 14, 256)
→【self.layer4】→(7, 7, 512)

→【self.avgpool, Adaptive, output_size=(1, 1)】→(1, 1, 512)
→【reshape】→(512,)→【self.fc】→(1000,)

self.layer1中包含2个Identity Block，每个Identity Block的结构如下

(56, 56, 64)

分支1：→【Conv, Cout=64, ks=3, s=1, p=1】→【bn】→【relu】→(56, 56, 64)
→【Conv, Cout=64, ks=3, s=1, p=1】→【bn】→(56, 56, 64)

分支2：→【identity maping】→(56, 56, 64)

合并add：→(56, 56, 64)→【relu】→(56, 56, 64)

self.layer2中包含1个Conv Block和3个Identity Block，其中Conv Block的结构如下

(56, 56 ,64)

分支1：→【Conv, Cout=128, ks=3, s=2, p=1】→【bn】→【relu】→(28, 28, 128)
→【Conv, Cout=128, ks=3, s=1, p=1】→【bn】→(28, 28, 128)

分支2(downsample)：→【Conv, Cout=128, ks=1, s=2】→【bn】→(28, 28, 128)

合并add：→(28, 28, 128)→【relu】→(28, 28, 128)

ResNet50

(224, 224, 3)

→【self.conv1, Cout=64, ks=7, s=2, p=3】→(112, 112, 64)
→【self.bn1】→【self.relu】→(112, 112, 64)
→【self.maxpool, ks=3, s=2, p=1】→(56, 56, 64)

→【self.layer1】→(56, 56, 256)
→【self.layer2】→(28, 28, 512)
→【self.layer3】→(14, 14, 1024)
→【self.layer4】→(7, 7, 2048)

→【self.avgpool, Adaptive, output_size=(1, 1)】→(1, 1, 2048)
→【reshape】→(2048,)→【self.fc】→(1000,)

self.layer1的第一个Block比较特殊，空间维度不变，通道数从64增加到256

(56, 56 ,64)

分支1：→【Conv, Cout=64, ks=1, s=1】→【bn】→【relu】→(56, 56, 64)
→【Conv, Cout=64, ks=3, s=1, p=1】→【bn】→【relu】→(56, 56, 64)
→【Conv, Cout=256, ks=1, s=1】→【bn】→(56, 56, 256)

分支2：→【Conv, Cout=256, ks=1, s=1】→【bn】→(56, 56, 256)

合并add：→(56, 56, 256)→【relu】→(56, 56, 256)

然后是2个Bottleneck型的Identity Block，结构如下

(56, 56, 256)

分支1：→【Conv, Cout=64, ks=1, s=1】→【bn】→【relu】→(56, 56, 64)
→【Conv, Cout=64, ks=3, s=1, p=1】→【bn】→【relu】→(56, 56, 64)
→【Conv, Cout=256, ks=1, s=1】→【bn】→(56, 56, 256)

分支2：→【identity maping】→(56, 56, 256)

合并add：→(56, 56, 256)→【relu】→(56, 56, 256)

self.layer2中包含1个Bottleneck型的Conv Block和3个Bottleneck型的Identity Block
其中Bottleneck型的Conv Block的结构如下

(56, 56, 256)

分支1：→【Conv, Cout=128, ks=1, s=1】→【bn】→【relu】→(56, 56, 128)
→【Conv, Cout=128, ks=3, s=2, p=1】→【bn】→【relu】→(28, 28, 128)
→【Conv, Cout=512, ks=1, s=1】→【bn】→(28, 28, 512)

分支2(downsample)：→【Conv, Cout=512, ks=1, s=2】→【bn】→(28, 28, 512)

合并add：→(28, 28, 512)→【relu】→(28, 28, 512)

Bottleneck型的Identity Block的结构如下

(28, 28, 512)

分支1：→【Conv, Cout=128, ks=1, s=1】→【bn】→【relu】→(28, 28, 128)
→【Conv, Cout=128, ks=3, s=1, p=1】→【bn】→【relu】→(28, 28, 128)
→【Conv, Cout=512, ks=1, s=1】→【bn】→(28, 28, 512)

分支2：→【identity maping】→(28, 28, 512)

合并add：→(28, 28, 512)→【relu】→(28, 28, 512)

第2个参数指定通道数，实际的通道数还要乘上block.expansion（BasicBlock.expansion=1, Bottleneck.expansion=4）
stride=2表示堆叠的第1个Block为

self.layer1 = self._make_layer(block, 64, layers[0])
self.layer2 = self._make_layer(block, 128, layers[1], stride=2)
self.layer3 = self._make_layer(block, 256, layers[2], stride=2)
self.layer4 = self._make_layer(block, 512, layers[3], stride=2)