[pytorch] resnet18, resnet20, resnet34, resnet50 network structure and Implementation

Pick classic early Pytorch Analyze the official implementation code

https://github.com/pytorch/vision/blob/9a481d0bec2700763a799ff148fe2e083b575441/torchvision/models/resnet.py
Various ResNet The Internet is made up of BasicBlock perhaps bottleneck Composed of , They are the basic modules of the deep residual network

ResNet The main body

ResNet Most of the various structures of are 1 layer conv+4 individual block+1 layer fc

class ResNet(nn.Module):

    def __init__(self, block, layers, zero_init_residual=False):
        super(ResNet, self).__init__()
        self.inplanes = 64
        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)
		self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(512 * block.expansion, num_classes)
        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.init.constant_(m.weight, 1)
                nn.init.constant_(m.bias, 0)

        # Zero-initialize the last BN in each residual branch,
        # so that the residual branch starts with zeros, and each residual block behaves like an identity.
        # This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
        if zero_init_residual:
            for m in self.modules():
                if isinstance(m, Bottleneck):
                    nn.init.constant_(m.bn3.weight, 0)
                elif isinstance(m, BasicBlock):
                    nn.init.constant_(m.bn2.weight, 0)

    def _make_layer(self, block, planes, blocks, stride=1):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
         # normly happened when stride = 2
            downsample = nn.Sequential(
                conv1x1(self.inplanes, planes * block.expansion, stride),
                nn.BatchNorm2d(planes * block.expansion),
            )

        layers = []
        layers.append(block(self.inplanes, planes, stride, downsample))
        self.inplanes = planes * block.expansion
        for _ in range(1, blocks):  
        # only the first block need downsample thus there is no downsample and stride = 2
            layers.append(block(self.inplanes, planes))

        return nn.Sequential(*layers)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        c2 = self.layer1(x)
        c3 = self.layer2(c2)
        c4 = self.layer3(c3)
        c5 = self.layer4(c4)
        x = self.avgpool(x)
        x = x.view(x.size(0), -1)
        x = self.fc(x)
        return c5

It should be noted that the last avgpool Is the global average pooling

BasicBlock

class BasicBlock(nn.Module):
    expansion = 1

    def __init__(self, inplanes, planes, stride=1, downsample=None):
    	# here planes names channel number
        super(BasicBlock, self).__init__()
        self.conv1 = conv3x3(inplanes, planes, stride)
        self.bn1 = nn.BatchNorm2d(planes)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(planes, planes)
        self.bn2 = nn.BatchNorm2d(planes)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        identity = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out

chart 1. BasicBlock chart ¹

ResNet18

The corresponding is [2,2,2,2]

def resnet18(pretrained=False, **kwargs):
    """Constructs a ResNet-18 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """
    model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs)
    if pretrained:
        print('Loading the pretrained model ...')
        # strict = False as we don't need fc layer params.
        model.load_state_dict(model_zoo.load_url(model_urls['resnet18']), strict=False)
    return model

ResNet34

def resnet34(pretrained=False, **kwargs):
    """Constructs a ResNet-34 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """
    model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs)
    if pretrained:
        print('Loading the pretrained model ...')
        model.load_state_dict(model_zoo.load_url(model_urls['resnet34']), strict=False)
    return model

ResNet20

This needs to be emphasized , natural ResNet20 It should be put forward in the article , in the light of cifar Data set design n=3 When , 1+6*3+1=20

class ResNet4Cifar(nn.Module):
    def __init__(self, block, num_block, num_classes=10):
        super().__init__()
        self.in_channels = 16
        self.conv1 = nn.Sequential(
            nn.Conv2d(3, 16, kernel_size=3, padding=1, bias=False),
            nn.BatchNorm2d(16),
            nn.ReLU(inplace=True))
        # we use a different inputsize than the original paper
        # so conv2_x's stride is 1
        self.conv2_x = self._make_layer(block, 16, num_block[0], 1)
        self.conv3_x = self._make_layer(block, 32, num_block[1], 2)
        self.conv4_x = self._make_layer(block, 64, num_block[2], 2)
        self.avg_pool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(64 * block.expansion, num_classes)

    def _make_layer(self, block, out_channels, num_blocks, stride):
        strides = [stride] + [1] * (num_blocks - 1)
        layers = []
        for stride in strides:
            layers.append(block(self.in_channels, out_channels, stride))
            self.in_channels = out_channels * block.expansion
        return nn.Sequential(*layers)

    def forward(self, x):
        output = self.conv1(x)
        output = self.conv2_x(output)
        output = self.conv3_x(output)
        output = self.conv4_x(output)
        output = self.avg_pool(output)
        output = output.view(output.size(0), -1)
        output = self.fc(output)
        return output

def resnet20(num_classes=10, **kargs):
    """ return a ResNet 20 object """
    return ResNet4Cifar(BasicBlock, [3, 3, 3], num_classes=num_classes)

We also calculate the parameter quantity as 0.27M, Consistent with the paper , Yes [1,3,32,32] The input of , The output dimension is [1,64,8,8]

chart 2 ResNet20 Parameter calculation

But there are also some articles that only change the first three layers 3x3 Convolution layer , The number of channels is not used 16、32、64, Is still 4 Layer of 64、128、256、512
, So the parameter quantity is 11.25M. For different tasks , But if you don't pay attention to the original network structure , This can be ignored .

Bottleneck Block

Bottleneck Block Used in 1×1 Convolution layer . If the number of input channels is 256,1×1 The convolution layer will reduce the number of channels to 64, after 3×3 After the convolution layer , Then increase the number of channels to 256.1×1 The advantage of convolution layer is in deeper networks , Handle inputs with a large number of channels with a small amount of parameters .
This structure is used in ResNet50、ResNet101 in .

chart 2. Bottleneck chart ¹

class Bottleneck(nn.Module):
    expansion = 4

    def __init__(self, inplanes, planes, stride=1, downsample=None):
        super(Bottleneck, self).__init__()
        self.conv1 = conv1x1(inplanes, planes)
        self.bn1 = nn.BatchNorm2d(planes)
        self.conv2 = conv3x3(planes, planes, stride)
        self.bn2 = nn.BatchNorm2d(planes)
        self.conv3 = conv1x1(planes, planes * self.expansion)
        self.bn3 = nn.BatchNorm2d(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x):
        identity = x

        out = self.conv1(x)
        out = self.bn1(out)
        out = self.relu(out)

        out = self.conv2(out)
        out = self.bn2(out)
        out = self.relu(out)

        out = self.conv3(out)
        out = self.bn3(out)

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out

ResNet50

chart 3. ResNet50 chart ²

The same network structure as above , hold Bottleneck It's OK to pile it up according to the number of layers

def resnet50(pretrained=False, **kwargs):
    """Constructs a ResNet-50 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """
    model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs)
    if pretrained:
        print('Loading the pretrained model ...')
        model.load_state_dict(model_zoo.load_url(model_urls['resnet50']), strict=False)
    return model

ResNet What problem has it solved

It is recommended to see the problem Resnet What problem is being solved ？
Post some of my favorite answers ：

A. about $L$ Layer network , There are no residuals Plain Net The attenuation of gradient correlation is in $\frac{1}{2^L}$ , and ResNet But the attenuation of it is only $\frac{1}{\sqrt{L}}$ . Even if BN After that, the mode of the gradient is stable in the normal range , However, the gradient correlation actually decreases continuously with the increase of the number of layers . And it turns out ,ResNet It can effectively reduce the attenuation of this correlation .

B. about “ Gradient dispersion ” From the point of view , Introduce an input into the output x The identity mapping of , Then the gradient will correspondingly introduce a constant 1, Such a network is indeed not prone to abnormal gradient values , In a sense , It plays a role in stabilizing the gradient .

C. Hop join addition can realize the combination of different resolution features , Because the shallow layer is easy to have high-resolution but low-level semantic features , The deep features have high-level semantics , But the resolution is very low . The introduction of jump connection actually makes the model itself more “ flexible ” Structure , That is, in the training process itself , The model can be selected in each part is “ More convolution and nonlinear transformation ” still “ More inclined to do nothing ”, Or a combination of the two . The model can adapt its own structure after training .³

D. When a residual network is used , Is to join skip connection structure , At this time, a building block The task of : F(x) := H(x), Turned into F(x) := H(x)-x Compare these two functions to be fitted , Fitting residual diagram is easier to optimize , in other words :F(x) := H(x)-x Than F(x) := H(x) Easier to optimize ⁴. An example of a differential amplifier is given ：F Is the network mapping before summation ,H Is the network mapping from input to summation . For example 5 Mapping to 5.1, So before introducing the residuals F’(5)=5.1, After introducing the residual, it is H(5)=5.1, H(5)=F(5)+5, F(5)=0.1. there F’ and F All represent network parameter mapping , The residual mapping is more sensitive to the change of output . such as s Output from 5.1 Change to 5.2, mapping F’ The output of has increased 1/51=2%, For the residual structure, the output is from 5.1 To 5.2, mapping F It's from 0.1 To 0.2, Added 100%. Obviously, the output change of the latter plays a greater role in adjusting the weight , So it works better . The idea of residuals is to remove the same main part , To highlight small changes .

There are many sayings , Easy to use, it's over ~

---

1. 【pytorch series 】ResNet Medium BasicBlock And bottleneck︎︎

2. ResNet50 Network structure diagram and detailed structure ︎

3. https://www.zhihu.com/question/64494691/answer/786270699︎

4. https://www.zhihu.com/question/64494691/answer/271335912︎