经典目标检测YOLO系列(一)YOLOV1的复现(1)总体架构

经典目标检测YOLO系列(一)实现YOLOV1网络(1)总体架构

实现原版的YOLOv1并没有多大的意义，因此，根据《YOLO目标检测》(ISBN:9787115627094)一书，在不脱离YOLOv1的大部分核心理念的前提下，重构一款较新的YOLOv1检测器，来对YOLOV1有更加深刻的认识。

书中源码连接:GitHub - yjh0410/RT-ODLab: YOLO Tutorial

对比原始YOLOV1网络，主要改进点如下：

1. 将主干网络替换为ResNet-18
2. 添加了后续YOLO中使用的neck，即SPPF模块
3. 删除全连接层，修改为检测头+预测层的组合，并且使用普遍用在RetinaNet、FCOS、YOLOX等通用目标检测网络中的解耦检测头（Decoupled head）
4. 修改损失函数，将YOLOV1原本的MSE loss，分类分支替换为BCE loss，回归分支替换为GIou loss。

我们按照现在目标检测的整体架构，来搭建自己的YOLOV1网络，概览图如下：

1、替换主干网络

• 使用ResNet-18作为主干网络(backbone)，替换YOLOV1原版的GoogLeNet风格的主干网络。替换后，图片的下采样从64倍，降低为32倍。我们更改图像大小由原版中的448×448变为416×416。
• 这里不同于分类网络，我们去掉网络中的平均池化层以及分类层。一张图片经过主干网络得到13×13×512的特征图，相对于原本输出的7×7网格，这里输出的网格要更加精细。

# RT-ODLab/models/detectors/yolov1/yolov1_backbone.pyimport torch
import torch.nn as nn
import torch.utils.model_zoo as model_zoo# 只会导入 __all__ 列出的成员，可以其他成员都被排除在外
__all__ = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152']model_urls = {'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth','resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth','resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth','resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth','resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth',
}# --------------------- Basic Module -----------------------
def conv3x3(in_planes, out_planes, stride=1):"""3x3 convolution with padding"""return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,padding=1, bias=False)def conv1x1(in_planes, out_planes, stride=1):"""1x1 convolution"""return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)class BasicBlock(nn.Module):expansion = 1def __init__(self, inplanes, planes, stride=1, downsample=None):super(BasicBlock, self).__init__()self.conv1 = conv3x3(inplanes, planes, stride)self.bn1 = nn.BatchNorm2d(planes)self.relu = nn.ReLU(inplace=True)self.conv2 = conv3x3(planes, planes)self.bn2 = nn.BatchNorm2d(planes)self.downsample = downsampleself.stride = stridedef forward(self, x):identity = xout = 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 += identityout = self.relu(out)return outclass Bottleneck(nn.Module):expansion = 4def __init__(self, inplanes, planes, stride=1, downsample=None):super(Bottleneck, self).__init__()self.conv1 = conv1x1(inplanes, planes)self.bn1 = nn.BatchNorm2d(planes)self.conv2 = conv3x3(planes, planes, stride)self.bn2 = nn.BatchNorm2d(planes)self.conv3 = conv1x1(planes, planes * self.expansion)self.bn3 = nn.BatchNorm2d(planes * self.expansion)self.relu = nn.ReLU(inplace=True)self.downsample = downsampleself.stride = stridedef forward(self, x):identity = xout = 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 += identityout = self.relu(out)return out# --------------------- ResNet -----------------------
class ResNet(nn.Module):def __init__(self, block, layers, zero_init_residual=False):super(ResNet, self).__init__()self.inplanes = 64self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,bias=False)self.bn1 = nn.BatchNorm2d(64)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])self.layer2 = self._make_layer(block, 128, layers[1], stride=2)self.layer3 = self._make_layer(block, 256, layers[2], stride=2)self.layer4 = self._make_layer(block, 512, layers[3], stride=2)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)# 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/1706.02677if zero_init_residual:for m in self.modules():if isinstance(m, Bottleneck):nn.init.constant_(m.bn3.weight, 0)elif isinstance(m, BasicBlock):nn.init.constant_(m.bn2.weight, 0)def _make_layer(self, block, planes, blocks, stride=1):downsample = Noneif stride != 1 or self.inplanes != planes * block.expansion:downsample = nn.Sequential(conv1x1(self.inplanes, planes * block.expansion, stride),nn.BatchNorm2d(planes * block.expansion),)layers = []layers.append(block(self.inplanes, planes, stride, downsample))self.inplanes = planes * block.expansionfor _ in range(1, blocks):layers.append(block(self.inplanes, planes))return nn.Sequential(*layers)def forward(self, x):"""Input:x: (Tensor) -> [B, C, H, W]Output:c5: (Tensor) -> [B, C, H/32, W/32]"""c1 = self.conv1(x)     # [B, C, H/2, W/2]c1 = self.bn1(c1)      # [B, C, H/2, W/2]c1 = self.relu(c1)     # [B, C, H/2, W/2]c2 = self.maxpool(c1)  # [B, C, H/4, W/4]c2 = self.layer1(c2)   # [B, C, H/4, W/4]c3 = self.layer2(c2)   # [B, C, H/8, W/8]c4 = self.layer3(c3)   # [B, C, H/16, W/16]c5 = self.layer4(c4)   # [B, C, H/32, W/32]return c5# --------------------- Fsnctions -----------------------
def resnet18(pretrained=False, **kwargs):"""Constructs a ResNet-18 model.Args:pretrained (bool): If True, returns a model pre-trained on ImageNet"""model = ResNet(BasicBlock, [2, 2, 2, 2], **kwargs)if pretrained:# strict = False as we don't need fc layer params.model.load_state_dict(model_zoo.load_url(model_urls['resnet18']), strict=False)return modeldef resnet34(pretrained=False, **kwargs):"""Constructs a ResNet-34 model.Args:pretrained (bool): If True, returns a model pre-trained on ImageNet"""model = ResNet(BasicBlock, [3, 4, 6, 3], **kwargs)if pretrained:model.load_state_dict(model_zoo.load_url(model_urls['resnet34']), strict=False)return modeldef resnet50(pretrained=False, **kwargs):"""Constructs a ResNet-50 model.Args:pretrained (bool): If True, returns a model pre-trained on ImageNet"""model = ResNet(Bottleneck, [3, 4, 6, 3], **kwargs)if pretrained:model.load_state_dict(model_zoo.load_url(model_urls['resnet50']), strict=False)return modeldef resnet101(pretrained=False, **kwargs):"""Constructs a ResNet-101 model.Args:pretrained (bool): If True, returns a model pre-trained on ImageNet"""model = ResNet(Bottleneck, [3, 4, 23, 3], **kwargs)if pretrained:model.load_state_dict(model_zoo.load_url(model_urls['resnet101']), strict=False)return modeldef resnet152(pretrained=False, **kwargs):"""Constructs a ResNet-152 model.Args:pretrained (bool): If True, returns a model pre-trained on ImageNet"""model = ResNet(Bottleneck, [3, 8, 36, 3], **kwargs)if pretrained:model.load_state_dict(model_zoo.load_url(model_urls['resnet152']))return model## build resnet
def build_backbone(model_name='resnet18', pretrained=False):if model_name == 'resnet18':model = resnet18(pretrained)feat_dim = 512elif model_name == 'resnet34':model = resnet34(pretrained)feat_dim = 512elif model_name == 'resnet50':model = resnet34(pretrained)feat_dim = 2048elif model_name == 'resnet101':model = resnet34(pretrained)feat_dim = 2048return model, feat_dimif __name__=='__main__':model, feat_dim = build_backbone(model_name='resnet18', pretrained=True)print(model)input = torch.randn(1, 3, 416, 416)print(model(input).shape)

2、添加neck

• 在原版的YOLOv1中，是没有Neck网络的概念的，但随着目标检测领域的发展，相关框架的成熟，一个通用目标检测网络结构可以被划分为Backbone、Neck、Head三大部分。当前的YOLO工作也符合这一设计。

• 因此，我们遵循当前主流的设计理念，为我们的YOLOv1添加一个Neck网络。这里，我们选择YOLOV5版本中所用的SPPF模块。

SPPF主要的代码如下:

# RT-ODLab/models/detectors/yolov1/yolov1_neck.pyimport torch
import torch.nn as nn
from .yolov1_basic import Conv# Spatial Pyramid Pooling - Fast (SPPF) layer for YOLOv5 by Glenn Jocher
class SPPF(nn.Module):"""This code referenced to https://github.com/ultralytics/yolov5"""def __init__(self, in_dim, out_dim, expand_ratio=0.5, pooling_size=5, act_type='lrelu', norm_type='BN'):super().__init__()inter_dim = int(in_dim * expand_ratio)self.out_dim = out_dim# 1×1卷积，通道降维self.cv1 = Conv(in_dim, inter_dim, k=1, act_type=act_type, norm_type=norm_type)# 1×1卷积，通道降维self.cv2 = Conv(inter_dim * 4, out_dim, k=1, act_type=act_type, norm_type=norm_type)# 5×5最大池化self.m = nn.MaxPool2d(kernel_size=pooling_size, stride=1, padding=pooling_size // 2)def forward(self, x):# 1、经过Conv-BN-LeakyReLUx = self.cv1(x)# 2、串行经过第1个5×5的最大池化y1 = self.m(x)# 3、串行经过第2个5×5的最大池化y2 = self.m(y1)# 4、串行经过第3个5×5的最大池化y3 = self.m(y2)# 5、将上面4个得到的结果concatconcat_y = torch.cat((x, y1, y2, y3), 1)# 6、再经过一个Conv-BN-LeakyReLUreturn self.cv2(concat_y)def build_neck(cfg, in_dim, out_dim):model = cfg['neck']print('==============================')print('Neck: {}'.format(model))# build neckif model == 'sppf':neck = SPPF(in_dim=in_dim,out_dim=out_dim,expand_ratio=cfg['expand_ratio'], pooling_size=cfg['pooling_size'],act_type=cfg['neck_act'],norm_type=cfg['neck_norm'])return neck     

3、Detection Head网络

• 原版的YOLOv1采用了全连接层来完成最终的处理，即将此前卷积输出的二维 H×W 特征图拉平（flatten操作）成一维 HW 向量，然后接全连接层得到4096维的一维向量。
• flatten操作会破坏特征的空间结构信息，因此，我们抛掉这里的flatten操作，改用当下主流的基于卷积的检测头。
• 具体来说，我们使用普遍用在RetinaNet、FCOS、YOLOX等通用目标检测网络中的解耦检测头（Decoupled head），即使用两条并行的分支，分别用于提取类别特征和位置特征，两条分支都由卷积层构成。
• 下图展示了我们所采用的Decoupled head以及后面的预测层结构。

# RT-ODLab/models/detectors/yolov1/yolov1_head.pyimport torch
import torch.nn as nn
try:from .yolov1_basic import Conv
except:from yolov1_basic import Convclass DecoupledHead(nn.Module):def __init__(self, cfg, in_dim, out_dim, num_classes=80):super().__init__()print('==============================')print('Head: Decoupled Head')self.in_dim = in_dimself.num_cls_head=cfg['num_cls_head']self.num_reg_head=cfg['num_reg_head']self.act_type=cfg['head_act']self.norm_type=cfg['head_norm']# cls headcls_feats = []self.cls_out_dim = max(out_dim, num_classes)for i in range(cfg['num_cls_head']):if i == 0:cls_feats.append(Conv(in_dim, self.cls_out_dim, k=3, p=1, s=1, act_type=self.act_type,norm_type=self.norm_type,depthwise=cfg['head_depthwise']))else:cls_feats.append(Conv(self.cls_out_dim, self.cls_out_dim, k=3, p=1, s=1, act_type=self.act_type,norm_type=self.norm_type,depthwise=cfg['head_depthwise']))# reg headreg_feats = []self.reg_out_dim = max(out_dim, 64)for i in range(cfg['num_reg_head']):if i == 0:reg_feats.append(Conv(in_dim, self.reg_out_dim, k=3, p=1, s=1, act_type=self.act_type,norm_type=self.norm_type,depthwise=cfg['head_depthwise']))else:reg_feats.append(Conv(self.reg_out_dim, self.reg_out_dim, k=3, p=1, s=1, act_type=self.act_type,norm_type=self.norm_type,depthwise=cfg['head_depthwise']))self.cls_feats = nn.Sequential(*cls_feats)self.reg_feats = nn.Sequential(*reg_feats)def forward(self, x):"""in_feats: (Tensor) [B, C, H, W]"""cls_feats = self.cls_feats(x)reg_feats = self.reg_feats(x)return cls_feats, reg_feats# build detection head
def build_head(cfg, in_dim, out_dim, num_classes=80):head = DecoupledHead(cfg, in_dim, out_dim, num_classes) return head

4、预测层

检测头最后部分的预测层（Prediction layer）也要做相应的修改。

• 我们只要求YOLOv1在每一个网格只需要预测一个bbox，而非两个甚至更多的bbox。尽管原版的YOLOv1在每个网格预测两个bbox，但在推理阶段，每个网格最终只输出一个bbox，从结果上来看，这和每个网格只预测一个bbox是一样的。
• objectness的预测。我们在Decoupled head的类别分支的输出后面接一层1x1卷积去做objectness的预测，并在最后使用Sigmoid函数来输出objectness预测值。
• classificaton预测。遵循当下YOLO系列常用的方法，我们同样采用Sigmoid函数来输出classification预测值，即每个类别的置信度都是0～1。classificaton预测和objectness预测都采用Decoupled的同一分支。
• bbox regression预测。另一分支的位置特征就被用于预测边界框的偏移量，即我们使用另一层1x1卷积去处理检测头的位置特征，得到每个网格的边界框的偏移量预测。

## 预测层
self.obj_pred = nn.Conv2d(head_dim, 1, kernel_size=1)
self.cls_pred = nn.Conv2d(head_dim, num_classes, kernel_size=1)
self.reg_pred = nn.Conv2d(head_dim, 4, kernel_size=1)

5、改进YOLO的详细网络图

• 不同于原版的YOLOv1，我们在不偏离原版的大部分核心理念的前提下，做了更加符合当下的设计理念的修改，包括使用更好的Backbone、添加Neck模块、修改检测头等。
• 改进YOLO的详细网络图如下

# RT-ODLab/models/detectors/yolov1/yolov1.pyimport torch
import torch.nn as nn
import numpy as npfrom utils.misc import multiclass_nmsfrom .yolov1_backbone import build_backbone
from .yolov1_neck import build_neck
from .yolov1_head import build_head# YOLOv1
class YOLOv1(nn.Module):def __init__(self,cfg,device,img_size=None,num_classes=20,conf_thresh=0.01,nms_thresh=0.5,trainable=False,deploy=False,nms_class_agnostic :bool = False):super(YOLOv1, self).__init__()# ------------------------- 基础参数 ---------------------------self.cfg = cfg                                 # 模型配置文件self.img_size = img_size                       # 输入图像大小self.device = device                           # cuda或者是cpuself.num_classes = num_classes                 # 类别的数量self.trainable = trainable                     # 训练的标记self.conf_thresh = conf_thresh                 # 得分阈值self.nms_thresh = nms_thresh                   # NMS阈值self.stride = 32                               # 网络的最大步长self.deploy = deployself.nms_class_agnostic = nms_class_agnostic# ----------------------- 模型网络结构 -------------------------## 主干网络self.backbone, feat_dim = build_backbone(cfg['backbone'],trainable & cfg['pretrained'])## 颈部网络self.neck = build_neck(cfg, feat_dim, out_dim=512)head_dim = self.neck.out_dim## 检测头self.head = build_head(cfg, head_dim, head_dim, num_classes)## 预测层self.obj_pred = nn.Conv2d(head_dim, 1, kernel_size=1)self.cls_pred = nn.Conv2d(head_dim, num_classes, kernel_size=1)self.reg_pred = nn.Conv2d(head_dim, 4, kernel_size=1)def create_grid(self, fmp_size):""" 用于生成G矩阵，其中每个元素都是特征图上的像素坐标。"""passdef decode_boxes(self, pred, fmp_size):"""将txtytwth转换为常用的x1y1x2y2形式。pred：回归预测参数fmp_size：特征图宽和高"""passdef postprocess(self, bboxes, scores):"""后处理代码，包括阈值筛选和非极大值抑制1、滤掉低得分（边界框的score低于给定的阈值）的预测边界框；2、滤掉那些针对同一目标的冗余检测。Input:bboxes: [HxW, 4]scores: [HxW, num_classes]Output:bboxes: [N, 4]score:  [N,]labels: [N,]"""[e-mail]_grad()def inference(self, x):# 测试阶段的前向推理代码# 主干网络feat = self.backbone(x)# 颈部网络feat = self.neck(feat)# 检测头cls_feat, reg_feat = self.head(feat)# 预测层obj_pred = self.obj_pred(cls_feat)cls_pred = self.cls_pred(cls_feat)reg_pred = self.reg_pred(reg_feat)fmp_size = obj_pred.shape[-2:]# 对 pred 的size做一些view调整，便于后续的处理# [B, C, H, W] -> [B, H, W, C] -> [B, H*W, C]obj_pred = obj_pred.permute(0, 2, 3, 1).contiguous().flatten(1, 2)cls_pred = cls_pred.permute(0, 2, 3, 1).contiguous().flatten(1, 2)reg_pred = reg_pred.permute(0, 2, 3, 1).contiguous().flatten(1, 2)# 测试时，笔者默认batch是1，# 因此，我们不需要用batch这个维度，用[0]将其取走。obj_pred = obj_pred[0]       # [H*W, 1]cls_pred = cls_pred[0]       # [H*W, NC]reg_pred = reg_pred[0]       # [H*W, 4]# 每个边界框的得分scores = torch.sqrt(obj_pred.sigmoid() * cls_pred.sigmoid())# 解算边界框, 并归一化边界框: [H*W, 4]bboxes = self.decode_boxes(reg_pred, fmp_size)if self.deploy:# [n_anchors_all, 4 + C]outputs = torch.cat([bboxes, scores], dim=-1)return outputselse:# 将预测放在cpu处理上，以便进行后处理scores = scores.cpu().numpy()bboxes = bboxes.cpu().numpy()# 后处理bboxes, scores, labels = self.postprocess(bboxes, scores)return bboxes, scores, labelsdef forward(self, x):# 训练阶段的前向推理代码if not self.trainable:return self.inference(x)else:# 主干网络feat = self.backbone(x)# 颈部网络feat = self.neck(feat)# 检测头cls_feat, reg_feat = self.head(feat)# 预测层obj_pred = self.obj_pred(cls_feat)cls_pred = self.cls_pred(cls_feat)reg_pred = self.reg_pred(reg_feat)fmp_size = obj_pred.shape[-2:]# 对 pred 的size做一些view调整，便于后续的处理# [B, C, H, W] -> [B, H, W, C] -> [B, H*W, C]obj_pred = obj_pred.permute(0, 2, 3, 1).contiguous().flatten(1, 2)cls_pred = cls_pred.permute(0, 2, 3, 1).contiguous().flatten(1, 2)reg_pred = reg_pred.permute(0, 2, 3, 1).contiguous().flatten(1, 2)# decode bboxbox_pred = self.decode_boxes(reg_pred, fmp_size)# 网络输出outputs = {"pred_obj": obj_pred,                  # (Tensor) [B, M, 1]"pred_cls": cls_pred,                   # (Tensor) [B, M, C]"pred_box": box_pred,                   # (Tensor) [B, M, 4]"stride": self.stride,                  # (Int)"fmp_size": fmp_size                    # (List) [fmp_h, fmp_w]}           return outputs

到此，我们完成了YOLOv1网络的搭建，并且实现了前向推理。但是，在推理的代码中还遗留了几个重要的问题尚待处理：

1. 如何从边界框偏移量reg_pred解耦出边界框坐标box_pred？
2. 如何实现后处理操作？
3. 如何计算训练阶段的损失？

当然还有数据集的加载，数据集如何增强，如何选择正样本进行训练等内容。