MMdetection解读系列之backbone-resnet

科普知识

ResNet

mmdetection中实现此结构

   arch_settings = {
        18: (BasicBlock, (2, 2, 2, 2)),
        34: (BasicBlock, (3, 4, 6, 3)),
        50: (Bottleneck, (3, 4, 6, 3)),
        101: (Bottleneck, (3, 4, 23, 3)),
        152: (Bottleneck, (3, 8, 36, 3))
    }
    # 注意BasicBlock Bottleneck均为一个类

RestNet 18/34采用BasicBlock作为基本单元，而ResNet 50/101/152采用Bottlenet Block作为基本单元

Basic Block就是左边的图，包括两个3×3的卷积操作。bollteneck是右边的图，为了减少参数，它采用了两个1×1的卷积。

参数解读

""ResNet backbone.

    Args:
        depth (int): Depth of resnet, from {18, 34, 50, 101, 152}. ResNet深度，总所周知，常用的为50,101
        stem_channels (int | None): Number of stem channels. If not specified,
            it will be the same as `base_channels`. Default: None.
        base_channels (int): Number of base channels of res layer. Default: 64. 众所周知，这个为64详情见Resnet Conv1
        in_channels (int): Number of input image channels. Default: 3.众所周知
        num_stages (int): Resnet stages. Default: 4.众所周知 详情见conv1_x conv2_x .......称为一个stage
        strides (Sequence[int]): Strides of the first block of each stage.众所周知  每个stage第一个block的步长
        dilations (Sequence[int]): Dilation of each stage.众所周知，用于空洞卷积
        out_indices (Sequence[int]): Output from which stages. 众所周知，每个stage的输出
        style (str): `pytorch` or `caffe`. If set to "pytorch", the stride-two ##使用的是pytorch/caffe关系到使用3*3/1*1的卷积
            layer is the 3x3 conv layer, otherwise the stride-two layer is ##在步长为2时
            the first 1x1 conv layer.
        deep_stem (bool): Replace 7x7 conv in input stem with 3 3x3 conv ##在初始的时候即conv1使用3*3的卷积替代7*7的卷积
        avg_down (bool): Use AvgPool instead of stride conv when
            downsampling in the bottleneck.## 使用平均池化替代卷积下采样 H=ceil((H-F+2*P)/S+1)
        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
            -1 means not freezing any parameters. ##从stage4开始冻结几个stage拒绝梯度传播更新参数
        norm_cfg (dict): Dictionary to construct and config norm layer.## 正则化方式
        norm_eval (bool): Whether to set norm layers to eval mode, namely,## 用于测试模型时，将正则化设置为false，
            freeze running stats (mean and var). Note: Effect on Batch Norm
            and its variants only.
        plugins (list[dict]): List of plugins for stages, each dict contains: ##扩展模块

            - cfg (dict, required): Cfg dict to build plugin.## 需要扩展哪些操作，比如说dfn
            - position (str, required): Position inside block to insert ##在那里插入这个操作，参数为opsName
              plugin, options are 'after_conv1', 'after_conv2', 'after_conv3'.
            - stages (tuple[bool], optional): Stages to apply plugin, length ## 在哪一个stage使用这个操作
              should be same as 'num_stages'.
        with_cp (bool): Use checkpoint or not. Using checkpoint will save some
            memory while slowing down the training speed. ## 众所周知checkpoint，看个人需求。
        zero_init_residual (bool): Whether to use zero init for last norm layer
            in resblocks to let them behave as identity.## 不知道

Resnet初始化

    def __init__(self,
                 depth,
                 in_channels