### mmcv.cnn 中 ConvModule 类的源码解析 ConvModule 是 mmcv 库中的一个重要组件，用于简化卷积层及其后续操作（如标准化、激活函数等）的组合。以下是 ConvModule 的主要实现细节： #### 导入必要的库和模块 ```python import torch from torch import nn from torch.nn.modules.batchnorm import _BatchNorm ``` #### 定义 ConvModule 类 ```python class ConvModule(nn.Module): """A conv block that bundles conv/norm/activation layers. This block simplifies the usage of convolution layers, which are commonly used with a norm layer (e.g., BatchNorm) and activation function. Args: in_channels (int): Number of channels in the input feature map. out_channels (int): Number of channels produced by the convolution. kernel_size (int or tuple[int]): Size of the convolving kernel. stride (int or tuple[int], optional): Stride of the convolution. Default: 1. padding (int or tuple[int], optional): Zero-padding added to both sides of the input. Default: 0. dilation (int or tuple[int], optional): Spacing between kernel elements. Default: 1. groups (int, optional): Number of blocked connections from input channels to output channels. Default: 1. bias (bool or str, optional): If specified as `auto`, it will be decided by the norm_cfg. Bias will not be set when norm_cfg is one of the following: - bn: per-channel affine transformation (True) - sync_bn: per-channel affine transformation (True) - gn: number of groups greater than 1 (True). Otherwise, it will remain unchanged. Default: "auto". conv_cfg (dict, optional): Config dict for convolution layer. Default: None, which means using conv2d. norm_cfg (dict, optional): Config dict for normalization layer. Default: None. act_cfg (dict, optional): Config dict for activation layer. Default: dict(type='ReLU'). inplace (bool, optional): Whether to use inplace version of the activation functions. Default: True. order (tuple[str], optional): The order of conv/norm/activation layers. It is a sequence of {"conv", "norm", "act"}. Example orders are ("conv", "norm", "act") and ("act", "conv", "norm"). """ def __init__(self, in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias='auto', conv_cfg=None, norm_cfg=None, act_cfg=dict(type='ReLU'), inplace=True, order=('conv', 'norm', 'act')): super(ConvModule, self).__init__() assert isinstance(order, tuple) and len(order) == 3 assert set(order) == {'conv', 'norm', 'act'} self.conv_cfg = conv_cfg self.norm_cfg = norm_cfg self.act_cfg = act_cfg self.inplace = inplace self.order = order # Build convolutional layer self.with_norm = norm_cfg is not None self.with_activatation = act_cfg is not None if bias == 'auto': bias = not self.with_norm self.conv = build_conv_layer( conv_cfg, in_channels, out_channels, kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias) # Build normalization layers if self.with_norm: if order.index('norm') > order.index('conv'): norm_channels = out_channels else: norm_channels = in_channels self.norm_name, norm = build_norm_layer(norm_cfg, norm_channels) self.add_module(self.norm_name, norm) # Build activation layer if self.with_activatation: act_type = act_cfg.pop('type') if act_type not in ['LeakyReLU']: self.activate = build_activation_layer(act_cfg) else: @torch.jit.script class LeakyReLUJITImplementation(torch.nn.LeakyReLU): pass self.activate = LeakyReLUJITImplementation(**act_cfg) # Use msra init defaultly self.init_weights() ``` 此代码片段展示了如何定义并初始化一个通用的卷积模块，其中包含了卷积层、规范化层以及激活函数的选择与配置。 为了进一步理解 ConvModule 的工作原理，可以查看其内部方法的具体实现方式[^1]。