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.py

import 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 = 1

    def __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 = 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)

        if self.downsample is not None:
            identity = self.downsample(x)

        out += identity
        out = self.relu(out)

        return out

class Bottleneck(nn.Module):
    expansion = 4

    def __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 = 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


# --------------------- ResNet -----------------------
class ResNet(nn.Module):

    def __init__(self, block, layers, zero_init_residual=False):
        super(ResNet, self).__init__()
        self.inplanes = 64
        self.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/1706.02677
        if 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 = None
        if 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.expansion
        for _ 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 model

def 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 model

def 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 model

def 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 model

def 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 = 512
    elif model_name == 'resnet34':
        model = resnet34(pretrained)
        feat_dim = 512
    elif model_name == 'resnet50':
        model = resnet34(pretrained)
        feat_dim = 2048
    elif model_name == 'resnet101':
        model = resnet34(pretrained)
        feat_dim = 2048

    return model, feat_dim


if __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.py

import 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-LeakyReLU
        x = 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concat
        concat_y = torch.cat((x, y1, y2, y3), 1)
        # 6、再经过一个Conv-BN-LeakyReLU
        return self.cv2(concat_y)


def build_neck(cfg, in_dim, out_dim):
    model = cfg['neck']
    print('==============================')
    print('Neck: {}'.format(model))
    # build neck
    if 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.py

import torch
import torch.nn as nn
try:
    from .yolov1_basic import Conv
except:
    from yolov1_basic import Conv


class 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_dim
        self.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 head
        cls_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 head
        reg_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.py

import torch
import torch.nn as nn
import numpy as np

from utils.misc import multiclass_nms

from .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cpu
        self.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 = deploy
        self.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。
        """
        pass


    def 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
        """
        pass


    def 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,]
        """
       pass


    @torch.no_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 outputs
        else:
            # 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, labels


    def 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 bbox
            box_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。

经典目标检测YOLO系列(一)复现YOLOV1(2)反解边界框及后处理

经典目标检测YOLO系列(一)复现YOLOV1(3)正样本的匹配及损失函数的实现

经典目标检测YOLO系列(一)复现YOLOV1(4)VOC2007数据集的读取及预处理

经典目标检测YOLO系列(一)复现YOLOV1(5)模型的训练及验证文章来源地址https://www.toymoban.com/news/detail-769913.html

到了这里，关于经典目标检测YOLO系列(一)YOLOV1的复现(1)总体架构的文章就介绍完了。如果您还想了解更多内容，请在右上角搜索TOY模板网以前的文章或继续浏览下面的相关文章，希望大家以后多多支持TOY模板网！