深度学习——残差网络ResNet基础结构理解+注释

import torch
import torch.nn as nn

class BasicBlockV1b(nn.Module):
    expansion = 1

    def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None,
                 previous_dilation=1, norm_layer=nn.BatchNorm2d):
        super(BasicBlockV1b, self).__init__()
        # 定义了一个二维卷积用于提取特征
        self.conv1 = nn.Conv2d(inplanes, planes, 3, stride,
                               dilation, dilation, bias=False)
        # 定义批归一化层，用于加速训练过程并提高模型稳定性
        self.bn1 = norm_layer(planes)
        self.relu = nn.ReLU(True)
        self.conv2 = nn.Conv2d(planes, planes, 3, 1, previous_dilation,
                               dilation=previous_dilation, bias=False)
        self.bn2 = norm_layer(planes)
        # 设置成下采样，用于调整残差块的输入与输出维度
        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)
        # 检查是否有下采样操作，通常是卷积层和批归一化层的组合，用于调整identity的维度与out匹配
        if self.downsample is not None:
            identity = self.downsample(x)
        # 将处理过的输出与跳跃连接输出相加，能减少梯度消失问题
        out += identity
        out = self.relu(out)

        return out

# 定义更复杂的残差块
class BottleneckV1b(nn.Module):
    expansion = 4

    def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None,
                 previous_dilation=1, norm_layer=nn.BatchNorm2d):
        super(BottleneckV1b, self).__init__()
        self.conv1 = nn.Conv2d(inplanes, planes, 1, bias=False)
        self.bn1 = norm_layer(planes)
        self.conv2 = nn.Conv2d(planes, planes, 3, stride,
                               dilation, dilation, bias=False)
        self.bn2 = norm_layer(planes)
        self.conv3 = nn.Conv2d(planes, planes * self.expansion, 1, bias=False)
        self.bn3 = norm_layer(planes * self.expansion)
        self.relu = nn.ReLU(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


# 定义了残差网络ResNetV1类的构造函数
class ResNetV1(nn.Module):
    # block指定了ResNet使用的块类型； layers包含每个块的数量; num_classes分类任务的类别数；
    # deep_stem是否使用深度茎结构； zero_init_residual是否对残差块的最后一个批归一化使用0初始化
    # norm_layer用于指定批归一化层
    def __init__(self, block, layers, num_classes=1000, deep_stem=False,
                 zero_init_residual=False, norm_layer=nn.BatchNorm2d):
        # 输出步长：模型下采样相关参数，决定网络最终特征图与输入图的空间分辨率比
        output_stride = 16
        scale = 1.0
        # dilations和strides是根据output_stride设置的扩张和步长，影响卷积层空间尺度
        if output_stride == 32:
            dilations = [1, 1]
            strides = [2, 2]
        elif output_stride == 16:
            dilations = [1, 2]
            strides = [2, 1]
        elif output_stride == 8:
            dilations = [2, 4]
            strides = [1, 1]
        else:
            raise NotImplementedError
        # 设置模型的初始通道数
        self.inplanes = int((128 if deep_stem else 64) * scale)
        super(ResNetV1, self).__init__()
        # 深度茎结构初始化
        if deep_stem:
            # resnet vc 使用小卷积层构造更深的初始层
            mid_channel = int(64 * scale)
            self.conv1 = nn.Sequential(
                # 输入通道转换到mid_channel，卷积核大小3，步长2，填充1
                nn.Conv2d(3, mid_channel, 3, 2, 1, bias=False),
                norm_layer(mid_channel),
                nn.ReLU(True),
                # 保持通道数不变
                nn.Conv2d(mid_channel, mid_channel, 3, 1, 1, bias=False),
                norm_layer(mid_channel),
                nn.ReLU(True),
                # 将通道增加到inplanes
                nn.Conv2d(mid_channel, self.inplanes, 3, 1, 1, bias=False)
            )
        else:
            self.conv1 = nn.Conv2d(3, self.inplanes, 7, 2, 3, bias=False)
        self.bn1 = norm_layer(self.inplanes)
        self.relu = nn.ReLU(True)
        self.maxpool = nn.MaxPool2d(3, 2, 1)
        # 构建不同的残差层
        self.layer1 = self._make_layer(block, int(64 * scale), layers[0], norm_layer=norm_layer)
        self.layer2 = self._make_layer(block, int(128 * scale), layers[1], stride=2, norm_layer=norm_layer)

        self.layer3 = self._make_layer(block, int(256 * scale), layers[2], stride=strides[0], dilation=dilations[0],
                                       norm_layer=norm_layer)
        self.layer4 = self._make_layer(block, int(512 * scale), layers[3], stride=strides[1], dilation=dilations[1],
                                       norm_layer=norm_layer, multi_grid=True,
                                       multi_dilation=[4,8,16])
        # 设置模型最后一层输出通道，用于后续全连接
        self.last_inp_channels = int(512 * block.expansion * scale)
        # 定义全局平均池化和全连接将特征图大小转换为1*1，用于分类任务
        self.avgpool = nn.AdaptiveAvgPool2d((1, 1))
        self.fc = nn.Linear(int(512 * block.expansion * scale), num_classes)
        # 用于循环初始化网络中的权重，卷积层使用Kaiming初始化+批归一化层权重1，偏置0
        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)
        # 残差连接的零初始化
        if zero_init_residual:
            for m in self.modules():
                if isinstance(m, BottleneckV1b):
                    nn.init.constant_(m.bn3.weight, 0)
                elif isinstance(m, BasicBlockV1b):
                    nn.init.constant_(m.bn2.weight, 0)
    # 用于构建每个残差层：block残差块类型； blocks输出通道数；dilation：卷积层空洞卷积大小
    def _make_layer(self, block, planes, blocks, stride=1, dilation=1, norm_layer=nn.BatchNorm2d,
                    multi_grid=False, multi_dilation=None):
        downsample = None
        if stride != 1 or self.inplanes != planes * block.expansion:
            downsample = nn.Sequential(
                nn.Conv2d(self.inplanes, planes * block.expansion, 1, stride, bias=False),
                norm_layer(planes * block.expansion),
            )

        layers = []
        if not multi_grid:  # 不使用多网格策略
            # 根据空洞卷积大小添加一个残差块到层中
            if dilation in (1, 2):
                layers.append(block(self.inplanes, planes, stride, dilation=1, downsample=downsample,
                                    previous_dilation=dilation, norm_layer=norm_layer))
            elif dilation == 4:
                layers.append(block(self.inplanes, planes, stride, dilation=2, downsample=downsample,
                                    previous_dilation=dilation, norm_layer=norm_layer))
            else:
                raise RuntimeError("=> unknown dilation size: {}".format(dilation))
        else:  # 这个是使用多网格策略
            layers.append(block(self.inplanes, planes, stride, dilation=multi_dilation[0],
                                downsample=downsample, previous_dilation=dilation, norm_layer=norm_layer))
        # 更新self.inplanes为当前层输出的通道数
        self.inplanes = planes * block.expansion

        if multi_grid: # 使用多网格策略，表示在一层中使用不同的空洞卷积策略
            # 计算多网格策略中不同空洞尺寸数量
            div = len(multi_dilation)
            for i in range(1, blocks):
                layers.append(block(self.inplanes, planes, dilation=multi_dilation[i % div],
                                    previous_dilation=dilation, norm_layer=norm_layer))
        else:
            for _ in range(1, blocks):
                layers.append(block(self.inplanes, planes, dilation=dilation,
                                    previous_dilation=dilation, norm_layer=norm_layer))

        return nn.Sequential(*layers)

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

        x1 = self.layer1(x)
        x2 = self.layer2(x1)
        x3 = self.layer3(x2)
        x4 = self.layer4(x3)

        # for classification
        # x = self.avgpool(c4)
        # x = x.view(x.size(0), -1)
        # x = self.fc(x)

        return x1, x2, x3, x4

def resnet50():
    num_block = [3, 4, 6, 3]
    model = ResNetV1(BottleneckV1b, num_block)
    return model

先提前理解了下这个网络的结构，然后看代码，边注释，看懂之后，比着写了一遍，就此记录一下！！！