RESNET code details

In order to analyze the network structure ： Downloaded netron

snap install netron # System ubuntu

chart 1 Resnet The overall architecture

Above picture ,[ ] It means a block .

about 18/34 layer Resnet

block In the official code is ： BasicBlock .BasicBlock according to identity whether Down sampling Divided into two . In the diagram above , The green arrow part namely conv3_1, conv4_1, and conv5_1 It's about identity Conduct Down sampling , Realization ： (1x1 conv, s=2) , And double each channel . stay In the code , adopt Pass in  _make_layer Of  stride Is it 1 determine .

about 50/101/152 layer Resnet

block In the official code is ： Bottleneck .Bottleneck according to identity whether Down sampling Divided into two . stay chart 1 in , The green arrow part namely conv3_1, conv4_1, and conv5_1 It's about identity Conduct Down sampling , Realization ： (1x1 conv, s=2) , And double each channel . stay In the code , adopt Pass in  _make_layer Of  stride Is it 1 determine .

  notes ： about  50/101/152 layer Resnet, according to chart 1 We know that through 3x3 Of max pool Then output the characteristic matrix shape Should be [56, 56, 64], But we conv2_x The corresponding The output characteristic matrix of the main branch shape yes [56, 56, 256]. therefore The first level residual structure requires shape from [56, 56, 64] --> [56, 56, 256]. Be careful , Here only adjust channel dimension , The height and width remain the same （ and conv3_1, conv4_1, conv5_1 Corresponding to a series of residual structures First floor dotted line residual structure It's not just about channel double + Halve the height and width ）.

Code ：

if stride != 1 or self.inplanes != planes * block.expansion:
    downsample = nn.Sequential(
        conv1x1(self.inplanes, planes * block.expansion, stride),
        nn.BatchNorm2d(planes * block.expansion),
    )

self.inplanes != planes * block.expansion ： 64！= 64 * 4

  Reference resources ：

resnet The paper

torchvision.models — Torchvision 0.11.0 documentation