ICode9

精准搜索请尝试: 精确搜索
首页 > 其他分享> 文章详细

里程碑的残差结构|ResNet(三)

2022-02-24 16:00:05  阅读:272  来源: 互联网

标签:layer self ResNet 残差 stride block out norm 里程碑


开局一张图,首先抛出resnet18的网络架构(完整版放在文章最下方)

resnet18(部分)

下面,再配合pytorch官方代码,解析一下resnet18。以resnet18为切入点,由浅入深,理解resnet架构

源码解析

demo

import torch
import torchvision.models as models
resnet18 = models.resnet18()

input = torch.randn(32,3,224,224)

output = resnet18(input)
print(resnet18)

ctrl+鼠标左键点击resnet18,进入resnet.py文件下
映入眼帘的是resnet18的构造函数

#构造函数  conv1为 1 层,conv2、conv3、conv4、conv5均为 4 层(每个 layer 有 2 个 BasicBlock,每个 BasicBlock 有 2 个卷积层),总共为 16 层,最后一层全连接层,总层数=1 + 4 × 4 + 1 = 18,依此类推。
def resnet18(pretrained: bool = False, progress: bool = True, **kwargs: Any) -> ResNet:
    r"""ResNet-18 model from
    `"Deep Residual Learning for Image Recognition" <https://arxiv.org/pdf/1512.03385.pdf>`_.

    Args:
        pretrained (bool): If True, returns a model pre-trained on ImageNet
        progress (bool): If True, displays a progress bar of the download to stderr
    """
    return _resnet('resnet18', BasicBlock, [2, 2, 2, 2], pretrained, progress,
                   **kwargs)

构造函数用到了_resnet()方法来创建网络

def _resnet(
    arch: str,
    block: Type[Union[BasicBlock, Bottleneck]],
    layers: List[int],
    pretrained: bool,
    progress: bool,
    **kwargs: Any
) -> ResNet:
    model = ResNet(block, layers, **kwargs)
    if pretrained: #加载预训练参数
        state_dict = load_state_dict_from_url(model_urls[arch],
                                              progress=progress)
        model.load_state_dict(state_dict)
    return model

_resnet()方法中,又调用了ResNet()方法创建模型

class ResNet(nn.Module):

    def __init__(
        self,
        block: Type[Union[BasicBlock, Bottleneck]],
        layers: List[int],
        num_classes: int = 1000,
        zero_init_residual: bool = False,
        groups: int = 1,
        width_per_group: int = 64,
        replace_stride_with_dilation: Optional[List[bool]] = None,
        norm_layer: Optional[Callable[..., nn.Module]] = None
    ) -> None:
        super(ResNet, self).__init__()
        if norm_layer is None:#判断是否传入 norm_layer,没有传入,则使用 BatchNorm2d
            norm_layer = nn.BatchNorm2d
        self._norm_layer = norm_layer

        self.inplanes = 64
        self.dilation = 1
        if replace_stride_with_dilation is None: #判断是否传入空洞卷积参数 replace_stride_with_dilation,如果不指定,则赋值为 [False, False, False],表示不使用空洞卷积。
            # each element in the tuple indicates if we should replace
            # the 2x2 stride with a dilated convolution instead
            replace_stride_with_dilation = [False, False, False]
        if len(replace_stride_with_dilation) != 3:
            raise ValueError("replace_stride_with_dilation should be None "
                             "or a 3-element tuple, got {}".format(replace_stride_with_dilation))
        self.groups = groups #读取分组卷积的参数
        self.base_width = width_per_group
        self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3,
                               bias=False)
        self.bn1 = norm_layer(self.inplanes)
        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]) #这个 layer 的参数没有指定 stride,默认 stride=1,因此这个 layer 不会改变图片大小
        self.layer2 = self._make_layer(block, 128, layers[1], stride=2,
                                       dilate=replace_stride_with_dilation[0]) #这个layer的stride=2,会降采样
        self.layer3 = self._make_layer(block, 256, layers[2], stride=2,
                                       dilate=replace_stride_with_dilation[1]) #这个layer的stride=2,会降采样
        self.layer4 = self._make_layer(block, 512, layers[3], stride=2,
                                       dilate=replace_stride_with_dilation[2]) #这个layer的stride=2,会降采样
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(512 * block.expansion, num_classes)

        #网络参数初始化
        for m in self.modules():
            if isinstance(m, nn.Conv2d): #卷积层采用 kaiming_normal_() 初始化方法  参考:https://arxiv.org/abs/1502.01852
                nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
            elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)): #bn 层和 GroupNorm 层初始化为 weight=1,bias=0
                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
        #每个 BasicBlock 和 Bottleneck 的最后一层 bn 的 weight=0,可以提升准确率 0.2~0.3%。
        if zero_init_residual:
            for m in self.modules():
                if isinstance(m, Bottleneck):
                    nn.init.constant_(m.bn3.weight, 0)  # type: ignore[arg-type]
                elif isinstance(m, BasicBlock):
                    nn.init.constant_(m.bn2.weight, 0)  # type: ignore[arg-type]

    def _make_layer(self, block: Type[Union[BasicBlock, Bottleneck]], planes: int, blocks: int,
                    stride: int = 1, dilate: bool = False) -> nn.Sequential:
        norm_layer = self._norm_layer
        downsample = None
        previous_dilation = self.dilation
        if dilate:#判断空洞卷积,计算 previous_dilation 参数
            self.dilation *= stride
            stride = 1
        if stride != 1 or self.inplanes != planes * block.expansion: #判断 stride 是否为 1,输入通道和输出通道是否相等。如果这两个条件都不成立,那么表明需要建立一个 1 X 1 的卷积层,来改变通道数和改变图片大小。具体是建立 downsample 层,包括 conv1x1 -> norm_layer。
            downsample = nn.Sequential(
                conv1x1(self.inplanes, planes * block.expansion, stride),
                norm_layer(planes * block.expansion),
            )

        layers = []
        #建立第一个 block,把 downsample 传给 block 作为降采样的层,并且 stride 也使用传入的 stride(stride=2)
        layers.append(block(self.inplanes, planes, stride, downsample, self.groups,
                            self.base_width, previous_dilation, norm_layer))
        self.inplanes = planes * block.expansion #改变通道数self.inplanes = planes * block.expansion。备注:在 BasicBlock 里,expansion=1,因此这一步不会改变通道数;在 Bottleneck 里,expansion=4,因此这一步会改变通道数
        #图片经过第一个 block后,就会改变通道数和图片大小。接下来 for 循环添加剩下的 block。从第 2 个 block 起,输入和输出通道数是相等的,因此就不用传入 downsample 和 stride(那么 block 的 stride 默认使用 1)
        for _ in range(1, blocks): #继续添加这个 layer 里接下来的 BasicBlock
            layers.append(block(self.inplanes, planes, groups=self.groups,
                                base_width=self.base_width, dilation=self.dilation,
                                norm_layer=norm_layer))

        return nn.Sequential(*layers)

    def _forward_impl(self, x: Tensor) -> Tensor: #整个resnet18的前向传播结构,可以参考文章最后的完整结构图来理解
        # See note [TorchScript super()]
        x = self.conv1(x)
        print("131254214")
        print(x.size())
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)

        x = self.avgpool(x)
        x = torch.flatten(x, 1)
        x = self.fc(x)

        return x

    def forward(self, x: Tensor) -> Tensor:
        return self._forward_impl(x)

可以看到,上面每个layer都是使用_make_layer()方法来创建层,下面来解析_make_layer()

    def _make_layer(self, block: Type[Union[BasicBlock, Bottleneck]], planes: int, blocks: int,
                    stride: int = 1, dilate: bool = False) -> nn.Sequential:
        norm_layer = self._norm_layer
        downsample = None
        previous_dilation = self.dilation
        if dilate:#判断空洞卷积,计算 previous_dilation 参数
            self.dilation *= stride
            stride = 1
        if stride != 1 or self.inplanes != planes * block.expansion: #判断 stride 是否为 1,输入通道和输出通道是否相等。如果这两个条件都不成立,那么表明需要建立一个 1 X 1 的卷积层,来改变通道数和改变图片大小。具体是建立 downsample 层,包括 conv1x1 -> norm_layer。
            downsample = nn.Sequential(
                conv1x1(self.inplanes, planes * block.expansion, stride),
                norm_layer(planes * block.expansion),
            )

        layers = []
        #建立第一个 block,把 downsample 传给 block 作为降采样的层,并且 stride 也使用传入的 stride(stride=2)
        layers.append(block(self.inplanes, planes, stride, downsample, self.groups,
                            self.base_width, previous_dilation, norm_layer))
        self.inplanes = planes * block.expansion #改变通道数self.inplanes = planes * block.expansion。备注:在 BasicBlock 里,expansion=1,因此这一步不会改变通道数;在 Bottleneck 里,expansion=4,因此这一步会改变通道数
        #图片经过第一个 block后,就会改变通道数和图片大小。接下来 for 循环添加剩下的 block。从第 2 个 block 起,输入和输出通道数是相等的,因此就不用传入 downsample 和 stride(那么 block 的 stride 默认使用 1)
        for _ in range(1, blocks): #继续添加这个 layer 里接下来的 BasicBlock
            layers.append(block(self.inplanes, planes, groups=self.groups,
                                base_width=self.base_width, dilation=self.dilation,
                                norm_layer=norm_layer))

        return nn.Sequential(*layers)

上面的过程如下图所示:

示例BasicBlock

resnet18用的是BasicBlock

class BasicBlock(nn.Module):
    expansion: int = 1

    def __init__(
        self,
        inplanes: int,
        planes: int,
        stride: int = 1,
        downsample: Optional[nn.Module] = None,
        groups: int = 1,
        base_width: int = 64,
        dilation: int = 1,
        norm_layer: Optional[Callable[..., nn.Module]] = None
    ) -> None:
        super(BasicBlock, self).__init__()
        if norm_layer is None: #判断是否传入了 norm_layer 层,如果没有,则使用 BatchNorm2d。
            norm_layer = nn.BatchNorm2d
        if groups != 1 or base_width != 64: #校验参数:groups == 1,base_width == 64,dilation == 1。也就是说,在 BasicBlock 中,不使用空洞卷积和分组卷积。
            raise ValueError('BasicBlock only supports groups=1 and base_width=64')
        if dilation > 1:
            raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
        # Both self.conv1 and self.downsample layers downsample the input when stride != 1
        self.conv1 = conv3x3(inplanes, planes, stride) #定义第 1 组 conv3x3 -> norm_layer -> relu,这里使用传入的 stride 和 inplanes。(如果是 layer2 ,layer3 ,layer4 里的第一个 BasicBlock,那么 stride=2,这里会降采样和改变通道数)。
        self.bn1 = norm_layer(planes)
        self.relu = nn.ReLU(inplace=True)
        self.conv2 = conv3x3(planes, planes) #定义第 2 组 conv3x3 -> norm_layer -> relu,这里不使用传入的 stride (默认为 1),输入通道数和输出通道数使用planes,也就是不需要降采样和改变通道数。
        self.bn2 = norm_layer(planes)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x: Tensor) -> Tensor:
        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

备注:resnet18,34用的是BasicBlock,而resnet50,101等都用的是Bottleneck结构

Bottleneck结构源码如下:

class Bottleneck(nn.Module):
    # Bottleneck in torchvision places the stride for downsampling at 3x3 convolution(self.conv2)
    # while original implementation places the stride at the first 1x1 convolution(self.conv1)
    # according to "Deep residual learning for image recognition"https://arxiv.org/abs/1512.03385.
    # This variant is also known as ResNet V1.5 and improves accuracy according to
    # https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch.

    expansion: int = 4

    def __init__(
        self,
        inplanes: int,
        planes: int,
        stride: int = 1,
        downsample: Optional[nn.Module] = None,
        groups: int = 1,
        base_width: int = 64,
        dilation: int = 1,
        norm_layer: Optional[Callable[..., nn.Module]] = None
    ) -> None:
        super(Bottleneck, self).__init__()
        if norm_layer is None:
            norm_layer = nn.BatchNorm2d
        width = int(planes * (base_width / 64.)) * groups #计算 width,等于传入的 planes,用于中间的3 × 3卷积。
        # Both self.conv2 and self.downsample layers downsample the input when stride != 1
        self.conv1 = conv1x1(inplanes, width) #定义第 1 组 conv1x1 -> norm_layer,这里不使用传入的 stride,使用 width,作用是进行降维,减少通道数。
        self.bn1 = norm_layer(width)
        self.conv2 = conv3x3(width, width, stride, groups, dilation) #定义第 2 组 conv3x3 -> norm_layer,这里使用传入的 stride,输入通道数和输出通道数使用width。(如果是 layer2 ,layer3 ,layer4 里的第一个 Bottleneck,那么 stride=2,这里会降采样)。
        self.bn2 = norm_layer(width)
        self.conv3 = conv1x1(width, planes * self.expansion) #定义第 3 组 conv1x1 -> norm_layer,这里不使用传入的 stride,使用 planes * self.expansion,作用是进行升维,增加通道数。
        self.bn3 = norm_layer(planes * self.expansion)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = downsample
        self.stride = stride

    def forward(self, x: Tensor) -> Tensor:
        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

BasicBlock与Bottleneck的不同:

  • BasicBlock 中有 1 个 3 × 3 3 \times 3 3×3卷积层,如果是 layer 的第一个 BasicBlock,那么第一个卷积层的 stride=2,作用是进行降采样。
  • Bottleneck 中有 2 个 1 × 1 1 \times 1 1×1卷积层, 1 个 3 × 3 3 \times 3 3×3卷积层。先经过第 1 个 1 × 1 1 \times 1 1×1卷积层,进行降维,然后经过 3 × 3 3 \times 3 3×3卷积层(如果是 layer 的第一个 Bottleneck,那么 3 × 3 3 \times 3 3×3卷积层的 stride=2,作用是进行降采样),最后经过 1 × 1 1 \times 1 1×1卷积层,进行升维 。

最后抛出resnet18的完整架构图:

resnet18完整结构

标签:layer,self,ResNet,残差,stride,block,out,norm,里程碑
来源: https://blog.csdn.net/wl1780852311/article/details/123113586

本站声明: 1. iCode9 技术分享网(下文简称本站)提供的所有内容,仅供技术学习、探讨和分享;
2. 关于本站的所有留言、评论、转载及引用,纯属内容发起人的个人观点,与本站观点和立场无关;
3. 关于本站的所有言论和文字,纯属内容发起人的个人观点,与本站观点和立场无关;
4. 本站文章均是网友提供,不完全保证技术分享内容的完整性、准确性、时效性、风险性和版权归属;如您发现该文章侵犯了您的权益,可联系我们第一时间进行删除;
5. 本站为非盈利性的个人网站,所有内容不会用来进行牟利,也不会利用任何形式的广告来间接获益,纯粹是为了广大技术爱好者提供技术内容和技术思想的分享性交流网站。

专注分享技术,共同学习,共同进步。侵权联系[81616952@qq.com]

Copyright (C)ICode9.com, All Rights Reserved.

ICode9版权所有