以MobileNet_V2为例，看一下在mmdet中backbone是如何实现的，在此之前需要对MobileNet_V2的架构有所了解，可参考MobileNet教程。

一、整体框架

二、初始化参数

三、ConvModule类

四、self.make_layer()

一、整体框架

源码路径为mmdet/models/backbones/mobilenet_v2.py，代码内容及粗略注释如下：

# Copyright (c) OpenMMLab. All rights reserved.
import warnings
import torch.nn as nn
from mmcv.cnn import ConvModule
from mmcv.runner import BaseModule
from torch.nn.modules.batchnorm import _BatchNorm
from ..builder import BACKBONES
from ..utils import InvertedResidual, make_divisible@BACKBONES.register_module()
class MobileNetV2(BaseModule):"""MobileNetV2 backbone.Args:widen_factor (float): Width multiplier, multiply number ofchannels in each layer by this amount. Default: 1.0.out_indices (Sequence[int], optional): Output from which stages.Default: (1, 2, 4, 7).frozen_stages (int): Stages to be frozen (all param fixed).Default: -1, which means not freezing any parameters.conv_cfg (dict, optional): Config dict for convolution layer.Default: None, which means using conv2d.norm_cfg (dict): Config dict for normalization layer.Default: dict(type='BN').act_cfg (dict): Config dict for activation layer.Default: dict(type='ReLU6').norm_eval (bool): Whether to set norm layers to eval mode, namely,freeze running stats (mean and var). Note: Effect on Batch Normand its variants only. Default: False.with_cp (bool): Use checkpoint or not. Using checkpoint will save somememory while slowing down the training speed. Default: False.pretrained (str, optional): model pretrained path. Default: Noneinit_cfg (dict or list[dict], optional): Initialization config dict.Default: None"""# Parameters to build layers. 4 parameters are needed to construct a# layer, from left to right: expand_ratio, channel, num_blocks, stride.# 网络配置参数arch_settings = [[1, 16, 1, 1],[6, 24, 2, 2],[6, 32, 3, 2],[6, 64, 4, 2],[6, 96, 3, 1],[6, 160, 3, 2],[6, 320, 1, 1]]# 初始化参数def __init__(self,# MobileNet通道扩展因子widen_factor=1.,# 输出特征矩阵所在特征层的索引out_indices=(1, 2, 4, 7),# 冻结层索引frozen_stages=-1,conv_cfg=None,# 归一化配置norm_cfg=dict(type='BN'),# 激活函数配置act_cfg=dict(type='ReLU6'),norm_eval=False,with_cp=False,pretrained=None,# 初始化父类初始化函数的配置init_cfg=None):# 继承父类，即BaseModule的初始化函数super(MobileNetV2, self).__init__(init_cfg)self.pretrained = pretrainedassert not (init_cfg and pretrained), \'init_cfg and pretrained cannot be specified at the same time'if isinstance(pretrained, str):warnings.warn('DeprecationWarning: pretrained is deprecated, ''please use "init_cfg" instead')self.init_cfg = dict(type='Pretrained', checkpoint=pretrained)elif pretrained is None:if init_cfg is None:self.init_cfg = [dict(type='Kaiming', layer='Conv2d'),dict(type='Constant',val=1,layer=['_BatchNorm', 'GroupNorm'])]else:raise TypeError('pretrained must be a str or None')self.widen_factor = widen_factorself.out_indices = out_indices# 特征层索引只能为0~7if not set(out_indices).issubset(set(range(0, 8))):raise ValueError('out_indices must be a subset of range'f'(0, 8). But received {out_indices}')# 规定冻结层的索引范围if frozen_stages not in range(-1, 8):raise ValueError('frozen_stages must be in range(-1, 8). 'f'But received {frozen_stages}')self.out_indices = out_indicesself.frozen_stages = frozen_stagesself.conv_cfg = conv_cfgself.norm_cfg = norm_cfgself.act_cfg = act_cfgself.norm_eval = norm_evalself.with_cp = with_cpself.in_channels = make_divisible(32 * widen_factor, 8)# 构建第一个卷积层，其中包括卷积+BN+激活函数self.conv1 = ConvModule(in_channels=3,out_channels=self.in_channels,kernel_size=3,stride=2,padding=1,conv_cfg=self.conv_cfg,norm_cfg=self.norm_cfg,act_cfg=self.act_cfg)self.layers = []# 根据网络配置参数循环构建网络for i, layer_cfg in enumerate(self.arch_settings):expand_ratio, channel, num_blocks, stride = layer_cfgout_channels = make_divisible(channel * widen_factor, 8)# 构建倒残差层（可包含多个倒残差模块）inverted_res_layer = self.make_layer(out_channels=out_channels,num_blocks=num_blocks,stride=stride,expand_ratio=expand_ratio)# 命名每一个层结构layer_name = f'layer{i + 1}'# 将倒残差模块作为值，其名称作为键，传入一个有序字典self.add_module(layer_name, inverted_res_layer)# 将倒残差模块名称存入列表self.layers.append(layer_name)if widen_factor > 1.0:self.out_channel = int(1280 * widen_factor)else:self.out_channel = 1280# 构建最后一个卷积层layer = ConvModule(in_channels=self.in_channels,out_channels=self.out_channel,kernel_size=1,stride=1,padding=0,conv_cfg=self.conv_cfg,norm_cfg=self.norm_cfg,act_cfg=self.act_cfg)self.add_module('conv2', layer)self.layers.append('conv2')def make_layer(self, out_channels, num_blocks, stride, expand_ratio):"""Stack InvertedResidual blocks to build a layer for MobileNetV2.Args:out_channels (int): out_channels of block.num_blocks (int): number of blocks.stride (int): stride of the first block. Default: 1expand_ratio (int): Expand the number of channels of thehidden layer in InvertedResidual by this ratio. Default: 6."""layers = []# 根据网络配置参数中倒残差模块的个数构建层，具体参见MobileNet教程for i in range(num_blocks):if i >= 1:stride = 1layers.append(# 直接调用InvertedResidual模块循环搭建倒残差模块InvertedResidual(self.in_channels,out_channels,mid_channels=int(round(self.in_channels * expand_ratio)),stride=stride,with_expand_conv=expand_ratio != 1,conv_cfg=self.conv_cfg,norm_cfg=self.norm_cfg,act_cfg=self.act_cfg,with_cp=self.with_cp))self.in_channels = out_channelsreturn nn.Sequential(*layers)# 根据传入的冻结层索引冻结相应的层def _freeze_stages(self):if self.frozen_stages >= 0:for param in self.conv1.parameters():param.requires_grad = Falsefor i in range(1, self.frozen_stages + 1):layer = getattr(self, f'layer{i}')layer.eval()for param in layer.parameters():param.requires_grad = False# 前向传播函数def forward(self, x):"""Forward function."""x = self.conv1(x)outs = []for i, layer_name in enumerate(self.layers):layer = getattr(self, layer_name)x = layer(x)# 根据需要哪些特征层对应特征图的索引，将特征图存储到列表中# 这些特征图将被传入rpn或其他结构进行目标框预测if i in self.out_indices:outs.append(x)return tuple(outs)# 在冻结BN层的同时将模型转变为训练模式def train(self, mode=True):"""Convert the model into training mode while keep normalization layerfrozen."""super(MobileNetV2, self).train(mode)self._freeze_stages()if mode and self.norm_eval:for m in self.modules():# trick: eval have effect on BatchNorm onlyif isinstance(m, _BatchNorm):m.eval()

二、初始化参数

在配置文件中backbone部分的参数就是对应的初始化参数。以ssdlite框架为例（路径为configs/ssd/ssdlite_mobilenetv2_scratch_600e_coco.py），配置文件中backbone部分如下：

backbone=dict(# backbone名称type='MobileNetV2',# 输出特征矩阵所在特征层的索引out_indices=(4, 7),# 对正则化的配置norm_cfg=dict(type='BN', eps=0.001, momentum=0.03),# 对父类初始化函数的配置init_cfg=dict(type='TruncNormal', layer='Conv2d', std=0.03)),

backbone中的type指的是使用的网络名称，mmdet支持的所有backbone名称均可在mmdet/models/backbones/__init__.py文件中找到：

# Copyright (c) OpenMMLab. All rights reserved.
from .csp_darknet import CSPDarknet
from .darknet import Darknet
from .detectors_resnet import DetectoRS_ResNet
from .detectors_resnext import DetectoRS_ResNeXt
from .hourglass import HourglassNet
from .hrnet import HRNet
from .mobilenet_v2 import MobileNetV2
from .regnet import RegNet
from .res2net import Res2Net
from .resnest import ResNeSt
from .resnet import ResNet, ResNetV1d
from .resnext import ResNeXt
from .ssd_vgg import SSDVGG
from .swin import SwinTransformer
from .trident_resnet import TridentResNet__all__ = ['RegNet', 'ResNet', 'ResNetV1d', 'ResNeXt', 'SSDVGG', 'HRNet','MobileNetV2', 'Res2Net', 'HourglassNet', 'DetectoRS_ResNet','DetectoRS_ResNeXt', 'Darknet', 'ResNeSt', 'TridentResNet', 'CSPDarknet','SwinTransformer'
]

其余参数将被传入MobileNetV2类中用于创建MobileNet模型。

三、ConvModule类

该类用于构建卷积层，包含卷积+归一化+激活函数，其中也包括很多函数没有细讲，掌握该类即可自行构建卷积层，在此不再深究原理，英语好的小伙伴可直接看英文注释，比较详细，另外也附注了中文注释：

import warnings
import torch.nn as nn
from ..utils import constant_init, kaiming_init
from .activation import build_activation_layer
from .conv import build_conv_layer
from .norm import build_norm_layer
from .padding import build_padding_layer
from .registry import PLUGIN_LAYERS@PLUGIN_LAYERS.register_module()
class ConvModule(nn.Module):"""A conv block that bundles conv/norm/activation layers.This block simplifies the usage of convolution layers, which are commonlyused with a norm layer (e.g., BatchNorm) and activation layer (e.g., ReLU).It is based upon three build methods: `build_conv_layer()`,`build_norm_layer()` and `build_activation_layer()`.Besides, we add some additional features in this module.1. Automatically set `bias` of the conv layer.2. Spectral norm is supported.3. More padding modes are supported. Before PyTorch 1.5, nn.Conv2d onlysupports zero and circular padding, and we add "reflect" padding mode.Args:in_channels (int): Number of channels in the input feature map.Same as that in ``nn._ConvNd``.out_channels (int): Number of channels produced by the convolution.Same as that in ``nn._ConvNd``.kernel_size (int | tuple[int]): Size of the convolving kernel.Same as that in ``nn._ConvNd``.stride (int | tuple[int]): Stride of the convolution.Same as that in ``nn._ConvNd``.padding (int | tuple[int]): Zero-padding added to both sides ofthe input. Same as that in ``nn._ConvNd``.dilation (int | tuple[int]): Spacing between kernel elements.Same as that in ``nn._ConvNd``.groups (int): Number of blocked connections from input channels tooutput channels. Same as that in ``nn._ConvNd``.bias (bool | str): If specified as `auto`, it will be decided by thenorm_cfg. Bias will be set as True if `norm_cfg` is None, otherwiseFalse. Default: "auto".conv_cfg (dict): Config dict for convolution layer. Default: None,which means using conv2d.norm_cfg (dict): Config dict for normalization layer. Default: None.act_cfg (dict): Config dict for activation layer.Default: dict(type='ReLU').inplace (bool): Whether to use inplace mode for activation.Default: True.with_spectral_norm (bool): Whether use spectral norm in conv module.Default: False.padding_mode (str): If the `padding_mode` has not been supported bycurrent `Conv2d` in PyTorch, we will use our own padding layerinstead. Currently, we support ['zeros', 'circular'] with officialimplementation and ['reflect'] with our own implementation.Default: 'zeros'.order (tuple[str]): The order of conv/norm/activation layers. It is asequence of "conv", "norm" and "act". Common examples are("conv", "norm", "act") and ("act", "conv", "norm").Default: ('conv', 'norm', 'act')."""_abbr_ = 'conv_block'def __init__(self,# 输入通道数in_channels,# 输出通道数out_channels,# 卷积核尺寸kernel_size,# 步距stride=1,# 填充padding=0,# 卷积核中元素间隔的距离，大于1则为空洞卷积dilation=1,# 将通道分成的组数groups=1,# 偏置bias='auto',conv_cfg=None,norm_cfg=None,act_cfg=dict(type='ReLU'),inplace=True,# 是否使用谱归一化with_spectral_norm=False,padding_mode='zeros',order=('conv', 'norm', 'act')):super(ConvModule, self).__init__()assert conv_cfg is None or isinstance(conv_cfg, dict)assert norm_cfg is None or isinstance(norm_cfg, dict)assert act_cfg is None or isinstance(act_cfg, dict)# 官方提供的两种填充模式official_padding_mode = ['zeros', 'circular']self.conv_cfg = conv_cfgself.norm_cfg = norm_cfgself.act_cfg = act_cfgself.inplace = inplaceself.with_spectral_norm = with_spectral_normself.with_explicit_padding = padding_mode not in official_padding_modeself.order = orderassert isinstance(self.order, tuple) and len(self.order) == 3assert set(order) == set(['conv', 'norm', 'act'])self.with_norm = norm_cfg is not Noneself.with_activation = act_cfg is not None# if the conv layer is before a norm layer, bias is unnecessary.# 若卷积之后有归一化层，则没有必要使用偏置if bias == 'auto':bias = not self.with_normself.with_bias = bias# 如果同时使用归一化和偏置会输出警告信息if self.with_norm and self.with_bias:warnings.warn('ConvModule has norm and bias at the same time')# 若不采用官方提供的填充模式，则自定义填充方式if self.with_explicit_padding:pad_cfg = dict(type=padding_mode)self.padding_layer = build_padding_layer(pad_cfg, padding)# reset padding to 0 for conv moduleconv_padding = 0 if self.with_explicit_padding else padding# build convolution layer# 构建卷积层self.conv = build_conv_layer(conv_cfg,in_channels,out_channels,kernel_size,stride=stride,padding=conv_padding,dilation=dilation,groups=groups,bias=bias)# export the attributes of self.conv to a higher level for convenience# 为了方便，将self.conv的属性赋值给更高一级的属性，即ConvModule类属性self.in_channels = self.conv.in_channelsself.out_channels = self.conv.out_channelsself.kernel_size = self.conv.kernel_sizeself.stride = self.conv.strideself.padding = paddingself.dilation = self.conv.dilationself.transposed = self.conv.transposedself.output_padding = self.conv.output_paddingself.groups = self.conv.groups# 使用谱归一化if self.with_spectral_norm:self.conv = nn.utils.spectral_norm(self.conv)# build normalization layers# 构建归一化层if self.with_norm:# norm layer is after conv layer# 判断归一化层在卷积层之前还是之后if order.index('norm') > order.index('conv'):norm_channels = out_channelselse:norm_channels = in_channelsself.norm_name, norm = build_norm_layer(norm_cfg, norm_channels)self.add_module(self.norm_name, norm)else:self.norm_name = None# build activation layer# 构建激活层if self.with_activation:act_cfg_ = act_cfg.copy()# nn.Tanh has no 'inplace' argumentif act_cfg_['type'] not in ['Tanh', 'PReLU', 'Sigmoid', 'HSigmoid', 'Swish']:act_cfg_.setdefault('inplace', inplace)self.activate = build_activation_layer(act_cfg_)# Use msra init by default# 使用默认的参数初始化方法self.init_weights()@propertydef norm(self):if self.norm_name:return getattr(self, self.norm_name)else:return None# 参数初始化方法def init_weights(self):# 1. It is mainly for customized conv layers with their own#    initialization manners by calling their own ``init_weights()``,#    and we do not want ConvModule to override the initialization.# 2. For customized conv layers without their own initialization#    manners (that is, they don't have their own ``init_weights()``)#    and PyTorch's conv layers, they will be initialized by#    this method with default ``kaiming_init``.# Note: For PyTorch's conv layers, they will be overwritten by our#    initialization implementation using default ``kaiming_init``.if not hasattr(self.conv, 'init_weights'):if self.with_activation and self.act_cfg['type'] == 'LeakyReLU':nonlinearity = 'leaky_relu'a = self.act_cfg.get('negative_slope', 0.01)else:nonlinearity = 'relu'a = 0kaiming_init(self.conv, a=a, nonlinearity=nonlinearity)if self.with_norm:constant_init(self.norm, 1, bias=0)# 定义前向传播函数def forward(self, x, activate=True, norm=True):for layer in self.order:if layer == 'conv':if self.with_explicit_padding:x = self.padding_layer(x)x = self.conv(x)elif layer == 'norm' and norm and self.with_norm:x = self.norm(x)elif layer == 'act' and activate and self.with_activation:x = self.activate(x)return x

四、self.make_layer()

该方法根据直接根据网络配置参中每一层残差块的重复个数，调用InvertedResidual()方法循环搭建倒残差模块。在MMdetection中预定义好了一些基础模块，可以直接调用，具体参见mmdet/models/utils/__init__.py：

# Copyright (c) OpenMMLab. All rights reserved.
from .builder import build_linear_layer, build_transformer
from .conv_upsample import ConvUpsample
from .csp_layer import CSPLayer
from .gaussian_target import gaussian_radius, gen_gaussian_target
from .inverted_residual import InvertedResidual
from .make_divisible import make_divisible
from .misc import interpolate_as
from .normed_predictor import NormedConv2d, NormedLinear
from .positional_encoding import (LearnedPositionalEncoding,SinePositionalEncoding)
from .res_layer import ResLayer, SimplifiedBasicBlock
from .se_layer import SELayer
from .transformer import (DetrTransformerDecoder, DetrTransformerDecoderLayer,DynamicConv, Transformer)__all__ = ['ResLayer', 'gaussian_radius', 'gen_gaussian_target','DetrTransformerDecoderLayer', 'DetrTransformerDecoder', 'Transformer','build_transformer', 'build_linear_layer', 'SinePositionalEncoding','LearnedPositionalEncoding', 'DynamicConv', 'SimplifiedBasicBlock','NormedLinear', 'NormedConv2d', 'make_divisible', 'InvertedResidual','SELayer', 'interpolate_as', 'ConvUpsample', 'CSPLayer'
]

以上就是MMdetection中backbone实现的基本流程，可参考上述流程进行自定义backbone搭建，当然自定义涉及到一些注册机制，等以后我搞懂了再更新。本人初学者，如有纰漏之处还请批评指正！

MMdetection中backbone的实现-MobileNetV2相关推荐

目标检测 SSD: Single Shot MultiBox Detector - SSD在MMDetection中的实现
目标检测 SSD: Single Shot MultiBox Detector - SSD在MMDetection中的实现 flyfish 目标检测 SSD: Single Shot MultiBox ...
计算机视觉系列-轻松掌握 MMDetection 中全景分割算法 MaskFormer（一）
计算机视觉系列-轻松掌握 MMDetection 中全景分割算法 MaskFormer(一) 目录全景分割简介 MaskFormer 简介 MaskFormer 配置代码计算机视觉学习笔记系列 ...
COCO 数据集格式及mmdetection中的转换方法
COCO 数据集格式及mmdetection中的转换方法 COCO格式 CV中的目标检测任务不同于分类,其标签的形式稍为复杂,有几种常用检测数据集格式,本文将简要介绍最为常见的COCO数据集的格式. ...
Anchor和RPN的浅薄理解(三)-mmdetection中Anchor生成源码分析
在 MMDetection 中,RPN 网络使用 AnchorGenerator 类生成 Anchor,在config文件中 AnchorGenerator 的默认设置如下: anchor_gener ...
目标检测和分割中Backbone技术总结
目标检测和分割中Backbone技术总结 VGG (2014) ResNet系列 ResNet (2015) ResNeXt ResNeSt Inception系列 Inception v1 Ince ...
【mmdetection系列】mmdetection之backbone讲解
如果是自己来写这部分代码的话,真的是很容易,如果这个框架中有自带的backbone结构的话,那可以从其他地方找到对应的pytorch版本实现,或者自己写. 配置部分在configs/_base_/mo ...
在Colab平台上用mmdetection中的Cascade RCNN训练自己的voc数据集
最近在学习目标检测,要用Cascde RCNN训练自己的数据集,因为本地电脑没有GPU,所以就借用Colab云端来进行模型训练,想要记录一下操作过程,以及希望可以给更多电脑没有GPU,但是需要用mmd ...
关于神经网络中Backbone，Neck，Bottleneck，Head的理解
文章目录摘要 1. Backbone 2. Neck 3. Bottleneck 4. Head 5.GAP或者avgpool: 6.Embedding 摘要梳理了一些长见的名词,方便大家够好的理 ...
【MMDet Note】MMDetection中AnchorGenerator代码理解与解读
文章目录前言一.总概二.代码解读 1.AnchorGenerator类 2. 属性@property 3.gen_base_anchors方法 4. gen_single_level_base_ ...

MMdetection中backbone的实现-MobileNetV2

一、整体框架

二、初始化参数

三、ConvModule类

四、self.make_layer()

MMdetection中backbone的实现-MobileNetV2相关推荐

最新文章

热门文章