What is the corresponding part of resnet101's "res4b22" in torchvision's pretrained model

vision
1

image
image707×367 60.4 KB

When i turn to the resnet source code in torchvision.model

model = ResNet(Bottleneck, [3, 4, 23, 3], **kwargs)

And the constructor of Resnet

        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
                               bias=False)
        self.bn1 = nn.BatchNorm2d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        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)

Compared to this image, it seems that the layer4 is corresponding with conv5
这里写图片描述
So, how should i perform fine-tune mentioned in the paper on pytorch resnet101 model?
Is the “res4b22” corresponding to

list(model.layer3.children())[-1]

What about the “res5a”, “res5b” and “res5c”, Are they corresponding to the 3 Bottleneck of conv5_x?

2

res4b22 correspond to model.layer3[-1]
And res5* correspond to model.layer5.
ref: https://github.com/KaimingHe/deep-residual-networks/blob/master/prototxt/ResNet-101-deploy.prototxt

3

The naming rule of ResNet as follows:
stage 2 and stage 5 both have 3 layers, so they are named as ‘a’, ‘b’, and ‘c’.
stage 3 and stage 4 has more than 3 layers. And there are clear differences between the first layer and rest layers of those two stages. So, their first layer is named as ‘a’, and rest layers are all 'b’s. In order to differentiate those 'b’s, they are named as ‘b1’, ‘b2’, and so on. Hence, for ResNet-101, stage 4 has 23 layers. They are one ‘a’ layer and 22 ‘b’ layers, ranging from ‘b1’ to ‘b22’.