经典目标检测YOLO系列(二)YOLOV2的复现(1)总体网络架构及前向推理过程
和之前实现的YOLOv1一样,根据《YOLO目标检测》(ISBN:9787115627094)
一书,在不脱离YOLOv2的大部分核心理念的前提下,重构一款较新的YOLOv2检测器,来对YOLOV2有更加深刻的认识。
书中源码连接: RT-ODLab: YOLO Tutorial
对比原始YOLOV2网络,主要改进点如下:
-
添加了后续YOLO中使用的neck,即SPPF模块
-
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)
-
修改损失函数,分类分支替换为BCE loss,回归分支替换为GIou loss。
-
由基于边界框的正样本匹配策略,改为基于先验框的正样本匹配策略。
对比之前实现的YOLOV1网络,主要改进点:
-
主干网络由ResNet18改为DarkNet19
-
添加先验框机制
-
正样本匹配策略改为:基于先验框的正样本匹配策略
-
YOLOv2代码和之前实现的YOLOv1相比,修改之处不多,建议先看之前实现的YOLOv1的相关文章。
1、YOLOv2网络架构
1.1 DarkNet19主干网络
- 使用原版YOLOv2中提出的DarkNet19作为主干网络(backbone)。
- 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。DarkNet19网络的下采样倍数为32,一张图片(416×416×3)经过主干网络,得到13×13×1024的特征图。
- 根据官方的做法,DarkNet19需要现在ImageNet数据集上进行预训练。不过,作者提供了DarkNet19在ImageNet数据集上的预训练权重,因此,我们只需要直接加载即可。
- 这里我们不去实现原版YOLOv2中的passthrough层,仅仅输出一个尺度,即c5层。
# RT-ODLab/models/detectors/yolov2/yolov2_backbone.py
import torch
import torch.nn as nn
model_urls = {
"darknet19": "https://github.com/yjh0410/image_classification_pytorch/releases/download/weight/darknet19.pth",
}
__all__ = ['DarkNet19']
# --------------------- Basic Module -----------------------
class Conv_BN_LeakyReLU(nn.Module):
def __init__(self, in_channels, out_channels, ksize, padding=0, stride=1, dilation=1):
super(Conv_BN_LeakyReLU, self).__init__()
self.convs = nn.Sequential(
nn.Conv2d(in_channels, out_channels, ksize, padding=padding, stride=stride, dilation=dilation),
nn.BatchNorm2d(out_channels),
nn.LeakyReLU(0.1, inplace=True)
)
def forward(self, x):
return self.convs(x)
# --------------------- DarkNet-19 -----------------------
class DarkNet19(nn.Module):
def __init__(self):
super(DarkNet19, self).__init__()
# backbone network : DarkNet-19
# output : stride = 2, c = 32
self.conv_1 = nn.Sequential(
Conv_BN_LeakyReLU(3, 32, 3, 1),
nn.MaxPool2d((2,2), 2),
)
# output : stride = 4, c = 64
self.conv_2 = nn.Sequential(
Conv_BN_LeakyReLU(32, 64, 3, 1),
nn.MaxPool2d((2,2), 2)
)
# output : stride = 8, c = 128
self.conv_3 = nn.Sequential(
Conv_BN_LeakyReLU(64, 128, 3, 1),
Conv_BN_LeakyReLU(128, 64, 1),
Conv_BN_LeakyReLU(64, 128, 3, 1),
nn.MaxPool2d((2,2), 2)
)
# output : stride = 8, c = 256
self.conv_4 = nn.Sequential(
Conv_BN_LeakyReLU(128, 256, 3, 1),
Conv_BN_LeakyReLU(256, 128, 1),
Conv_BN_LeakyReLU(128, 256, 3, 1),
)
# output : stride = 16, c = 512
self.maxpool_4 = nn.MaxPool2d((2, 2), 2)
self.conv_5 = nn.Sequential(
Conv_BN_LeakyReLU(256, 512, 3, 1),
Conv_BN_LeakyReLU(512, 256, 1),
Conv_BN_LeakyReLU(256, 512, 3, 1),
Conv_BN_LeakyReLU(512, 256, 1),
Conv_BN_LeakyReLU(256, 512, 3, 1),
)
# output : stride = 32, c = 1024
self.maxpool_5 = nn.MaxPool2d((2, 2), 2)
self.conv_6 = nn.Sequential(
Conv_BN_LeakyReLU(512, 1024, 3, 1),
Conv_BN_LeakyReLU(1024, 512, 1),
Conv_BN_LeakyReLU(512, 1024, 3, 1),
Conv_BN_LeakyReLU(1024, 512, 1),
Conv_BN_LeakyReLU(512, 1024, 3, 1)
)
def forward(self, x):
c1 = self.conv_1(x) # c1
c2 = self.conv_2(c1) # c2
c3 = self.conv_3(c2) # c3
c3 = self.conv_4(c3) # c3
c4 = self.conv_5(self.maxpool_4(c3)) # c4
c5 = self.conv_6(self.maxpool_5(c4)) # c5
return c5
# --------------------- Fsnctions -----------------------
def build_backbone(model_name='darknet19', pretrained=False):
if model_name == 'darknet19':
# model
model = DarkNet19()
feat_dim = 1024
# load weight
if pretrained:
print('Loading pretrained weight ...')
url = model_urls['darknet19']
# checkpoint state dict
checkpoint_state_dict = torch.hub.load_state_dict_from_url(
url=url, map_location="cpu", check_hash=True)
# model state dict
model_state_dict = model.state_dict()
# check
for k in list(checkpoint_state_dict.keys()):
if k in model_state_dict:
shape_model = tuple(model_state_dict[k].shape)
shape_checkpoint = tuple(checkpoint_state_dict[k].shape)
if shape_model != shape_checkpoint:
checkpoint_state_dict.pop(k)
else:
checkpoint_state_dict.pop(k)
print(k)
model.load_state_dict(checkpoint_state_dict)
return model, feat_dim
if __name__ == '__main__':
import time
model, feat_dim = build_backbone(pretrained=True)
x = torch.randn(1, 3, 416, 416)
t0 = time.time()
for layer in model.children():
x = layer(x)
print(layer.__class__.__name__, 'output shape:', x.shape)
# y = model(x)
t1 = time.time()
print('Time: ', t1 - t0)
1.2 添加neck
- 和之前实现的YOLOv1一致,选择YOLOV5版本中所用的SPPF模块。
- 代码在RT-ODLab/models/detectors/yolov2/yolov2_neck.py文件中,不在赘述。
1.3 Detection Head网络
- 和之前实现的YOLOv1一致,即使用解耦检测头(Decoupled head)。
- 代码在RT-ODLab/models/detectors/yolov2/yolov1_head.py文件中,不在赘述。
1.4 预测层
- 如下图,由于预测层多了先验框,因此预测层的输出通道的数量略有变化。
## 预测层
# 与YoloV1相比,YoloV2每个网格会预测5个框(VOC数据集),因此需×5
self.obj_pred = nn.Conv2d(head_dim, 1 * self.num_anchors, kernel_size=1)
self.cls_pred = nn.Conv2d(head_dim, num_classes * self.num_anchors, kernel_size=1)
self.reg_pred = nn.Conv2d(head_dim, 4 * self.num_anchors, kernel_size=1)
1.5 改进YOLOv2的详细网络图
- 与之前实现的YOLOv1相比,主干网络由ResNet18变为DarkNet19,每个网格预测5个anchor box,其他方面一致。
- 与原版的YOLOv2相比,做了更加符合当下的设计理念的修改,包括添加Neck模块、修改检测头等,但是没有引入passthrough层。
- 尽管和原版的YOLOv2有所差别,但内核思想是一致的,均是在YOLOv1的单级检测架构上引入了先验框。
# RT-ODLab/models/detectors/yolov2/yolov2.py
import torch
import torch.nn as nn
import numpy as np
from utils.misc import multiclass_nms
from .yolov2_backbone import build_backbone
from .yolov2_neck import build_neck
from .yolov2_head import build_head
# YOLOv2
class YOLOv2(nn.Module):
def __init__(self,
cfg,
device,
num_classes=20,
conf_thresh=0.01,
nms_thresh=0.5,
topk=100,
trainable=False,
deploy=False,
nms_class_agnostic=False):
super(YOLOv2, self).__init__()
# ------------------- Basic parameters -------------------
self.cfg = cfg # 模型配置文件
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.topk = topk # topk
self.stride = 32 # 网络的最大步长
self.deploy = deploy
self.nms_class_agnostic = nms_class_agnostic
# ------------------- Anchor box -------------------
self.anchor_size = torch.as_tensor(cfg['anchor_size']).float().view(-1, 2) # [A, 2]
self.num_anchors = self.anchor_size.shape[0]
# ------------------- Network Structure -------------------
## 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
# 与YoloV1相比,YoloV2每个网格会预测5个框(VOC数据集),因此需×5
self.obj_pred = nn.Conv2d(head_dim, 1 * self.num_anchors, kernel_size=1)
self.cls_pred = nn.Conv2d(head_dim, num_classes * self.num_anchors, kernel_size=1)
self.reg_pred = nn.Conv2d(head_dim, 4 * self.num_anchors, kernel_size=1)
if self.trainable:
self.init_bias()
def init_bias(self):
# init bias
init_prob = 0.01
bias_value = -torch.log(torch.tensor((1. - init_prob) / init_prob))
nn.init.constant_(self.obj_pred.bias, bias_value)
nn.init.constant_(self.cls_pred.bias, bias_value)
def generate_anchors(self, fmp_size):
pass
def decode_boxes(self, anchors, reg_pred):
pass
def postprocess(self, obj_pred, cls_pred, reg_pred, anchors):
"""
后处理代码,包括topk操作、阈值筛选和非极大值抑制
"""
pass
@torch.no_grad()
def inference(self, x):
bs = x.shape[0]
# 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(reg_feat)
cls_pred = self.cls_pred(cls_feat)
reg_pred = self.reg_pred(reg_feat)
fmp_size = obj_pred.shape[-2:]
# anchors: [M, 2]
anchors = self.generate_anchors(fmp_size)
# 对 pred 的size做一些view调整,便于后续的处理
# [B, A*C, H, W] -> [B, H, W, A*C] -> [B, H*W*A, C]
obj_pred = obj_pred.permute(0, 2, 3, 1).contiguous().view(bs, -1, 1) # [1, 845=13×13×5, 1]
cls_pred = cls_pred.permute(0, 2, 3, 1).contiguous().view(bs, -1, self.num_classes)
reg_pred = reg_pred.permute(0, 2, 3, 1).contiguous().view(bs, -1, 4)
# 测试时,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*A, 1]
cls_pred = cls_pred[0] # [H*W*A, NC]
reg_pred = reg_pred[0] # [H*W*A, 4]
if self.deploy:
scores = torch.sqrt(obj_pred.sigmoid() * cls_pred.sigmoid())
bboxes = self.decode_boxes(anchors, reg_pred)
# [n_anchors_all, 4 + C]
outputs = torch.cat([bboxes, scores], dim=-1)
return outputs
else:
# post process
bboxes, scores, labels = self.postprocess(
obj_pred, cls_pred, reg_pred, anchors)
return bboxes, scores, labels
def forward(self, x):
if not self.trainable:
return self.inference(x)
else:
bs = x.shape[0]
# 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(reg_feat)
cls_pred = self.cls_pred(cls_feat)
reg_pred = self.reg_pred(reg_feat)
fmp_size = obj_pred.shape[-2:]
# A就是Anchor的数量,VOC数据集上设置为5
# anchors: [M, 2], M = H*W*A
anchors = self.generate_anchors(fmp_size)
# 对 pred 的size做一些view调整,便于后续的处理
# [B, A*C, H, W] -> [B, H, W, A*C] -> [B, H*W*A, C]
obj_pred = obj_pred.permute(0, 2, 3, 1).contiguous().view(bs, -1, 1)
cls_pred = cls_pred.permute(0, 2, 3, 1).contiguous().view(bs, -1, self.num_classes)
reg_pred = reg_pred.permute(0, 2, 3, 1).contiguous().view(bs, -1, 4)
# decode bbox
box_pred = self.decode_boxes(anchors, reg_pred)
# 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
2、YOLOV2的前向推理
在1.5代码中,还遗留几个问题:
- 如何从边界框偏移量reg_pred解耦出边界框坐标box_pred?
- 如何实现后处理操作?
- 如何计算训练阶段的损失?
2.1 解耦边界框坐标
2.1.1 先验框矩阵的生成
YOLOv2网络配置参数如下,我们从中能看到anchor_size变量。这是基于kmeans聚类,在COCO数据集上聚类出的先验框,由于COCO数据集更大、图片更加丰富,因此我们将这几个先验框用在VOC数据集上。
# RT-ODLab/config/model_config/yolov2_config.py
# YOLOv2 Config
yolov2_cfg = {
# input
'trans_type': 'ssd',
'multi_scale': [0.5, 1.5],
# model
'backbone': 'darknet19',
'pretrained': True,
'stride': 32, # P5
'max_stride': 32,
# neck
'neck': 'sppf',
'expand_ratio': 0.5,
'pooling_size': 5,
'neck_act': 'lrelu',
'neck_norm': 'BN',
'neck_depthwise': False,
# head
'head': 'decoupled_head',
'head_act': 'lrelu',
'head_norm': 'BN',
'num_cls_head': 2,
'num_reg_head': 2,
'head_depthwise': False,
'anchor_size': [[17, 25],
[55, 75],
[92, 206],
[202, 21],
[289, 311]], # 416
# matcher
'iou_thresh': 0.5,
# loss weight
'loss_obj_weight': 1.0,
'loss_cls_weight': 1.0,
'loss_box_weight': 5.0,
# training configuration
'trainer_type': 'yolov8',
}
-
回想一下,在之前实现的YOLOv1中,我们通过构造矩阵G,得到了每一个网格(grid_x,grid_y)的坐标。
-
由于我们在YOLOv2中引入了先验框,因此,我们不仅需要每一个网格(grid_x,grid_y)的坐标,还要包含先验框(5个)的尺寸信息。
-
先验框矩阵生成代码如下
# RT-ODLab/models/detectors/yolov2/yolov2.py
def generate_anchors(self, fmp_size):
"""
fmp_size: (List) [H, W]
默认缩放后的图像为416×416,那么经过32倍下采样后,fmp_size为13×13
"""
# 1、特征图的宽和高
fmp_w, fmp_h = fmp_size
# 2、生成网格的x坐标和y坐标
anchor_y, anchor_x = torch.meshgrid([torch.arange(fmp_h), torch.arange(fmp_w)])
# 3、将xy两部分的坐标拼接起来,shape为[H, W, 2]
# 再转换下, shape变为[HW, 2]
anchor_xy = torch.stack([anchor_x, anchor_y], dim=-1).float().view(-1, 2)
# 4、引入了anchor box机制,每个网格包含A个anchor,因此每个(grid_x, grid_y)的坐标需要复制A(Anchor nums)份
# 相当于 每个网格左上角的坐标点复制5份 作为5个不同宽高anchor box的中心点
# [HW, 2] -> [HW, A, 2] -> [M, 2]
anchor_xy = anchor_xy.unsqueeze(1).repeat(1, self.num_anchors, 1)
anchor_xy = anchor_xy.view(-1, 2).to(self.device)
# 5、将kmeans聚类得出的5组anchor box的宽高复制13×13份
# [A, 2] -> [1, A, 2] -> [HW, A, 2] -> [M, 2]
anchor_wh = self.anchor_size.unsqueeze(0).repeat(fmp_h*fmp_w, 1, 1)
anchor_wh = anchor_wh.view(-1, 2).to(self.device)
# 6、将中心点和宽高cat起来,得到的shape为[M, 4]
# 其中M=13×13×5 表示feature map为13×13,每个网格有5组anchor box
# 4代表anchor box的位置(x_center, y_center, w, h)
# 需要注意:
# x_center, y_center是feature map上的坐标位置,需要×stride 才能得到缩放后原始图像上的中心点
# w, h是针对缩放后原始图像
anchors = torch.cat([anchor_xy, anchor_wh], dim=-1)
return anchors
2.1.2 解耦边界框
- 生成先验框矩阵后,我们就能通过边界框偏移量reg_pred解耦出边界框坐标box_pred》。
- 计算预测边界框的中心点坐标与之前计算YOLOv1是一致的,但是计算宽高发生了变化。这是因为YOLOv2中,我们引入了先验框,而且我们先验框的尺寸设定是相对于resize后图像大小,因此不需要乘stride。
def decode_boxes(self, anchors, reg_pred):
"""
1、依据预测值reg_pred(t_x,t_y,t_w,t_h)结算出边界框中心点坐标c_x, c_y和宽高b_w, b_h
c_x = ( grid_x + sigmoid(t_x) ) × stride
c_y = ( grid_y + sigmoid(t_y) ) × stride
b_w = p_w × exp(t_w)
b_h = p_h × exp(t_h)
其中 grid_x,grid_y,p_w,p_h为先验框的结果,即anchors结果
2、转换为常用的x1y1x2y2形式。
注意:
预测的宽高不是相对于feature map的,而是相对于resize后图像大小,因此不需要×stride
"""
# 1、计算预测边界框的中心点坐标和宽高
pred_ctr = (anchors[..., :2] + torch.sigmoid(reg_pred[..., :2])) * self.stride
pred_wh = anchors[..., 2:] * torch.exp(reg_pred[..., 2:]) # 不需要×stride
# 2、将所有bbox的中心点坐标和宽高换算成x1y1x2y2形式
pred_x1y1 = pred_ctr - pred_wh * 0.5
pred_x2y2 = pred_ctr + pred_wh * 0.5
pred_box = torch.cat([pred_x1y1, pred_x2y2], dim=-1)
return pred_box
2.2 后处理操作
- 之前YOLOv1的后处理操作,仅仅包含了阈值筛选和非极大值抑制NMS,这里由于引入了先验框,因此我们后处理的框的数量由之前的13×13变成了13×13×5(845)个。
- 这845个框不都是高质量的,因此我们先做一个topk,依据得分从高到低取前k个。对于COCO数据集来说,一张图片的目标数量不超过100,因此一般只需要设定topk=100。这里,作者为了提高测试的mAP,默认设置topk=1000。
- topk操作后,继续进行阈值筛选和非极大值抑制。
# RT-ODLab/models/detectors/yolov2/yolov2.py
def postprocess(self, obj_pred, cls_pred, reg_pred, anchors):
"""
后处理代码,包括topk操作、阈值筛选和非极大值抑制
1、topk操作:
在coco数据集中,检测对象的数量一半不会超过100,因此先选择得分最高的k个边界框,这里为了取得更高的mAP,取k=1000
在实际的场景中,不需要把k值取这么大
2、滤掉低得分(边界框的score低于给定的阈值)的预测边界框;
3、滤掉那些针对同一目标的冗余检测。
Input:
obj_pred: (Tensor) [H*W*A, 1]
cls_pred: (Tensor) [H*W*A, C]
reg_pred: (Tensor) [H*W*A, 4]
anchors: (Tensor) [H*W*A, 4]
其中,H*W*A = 13×13×5 = 845
"""
# (H x W x A x C,)
# 13×13×5×20 = 16900
scores = torch.sqrt(obj_pred.sigmoid() * cls_pred.sigmoid()).flatten()
# 1、topk操作
# Keep top k top scoring indices only.
num_topk = min(self.topk, reg_pred.size(0))
# torch.sort is actually faster than .topk (at least on GPUs)
predicted_prob, topk_idxs = scores.sort(descending=True)
topk_scores = predicted_prob[:num_topk]
topk_idxs = topk_idxs[:num_topk]
# 2、滤掉低得分(边界框的score低于给定的阈值)的预测边界框
# filter out the proposals with low confidence score
keep_idxs = topk_scores > self.conf_thresh
scores = topk_scores[keep_idxs]
topk_idxs = topk_idxs[keep_idxs]
# 获取flatten之前topk_scores所在的idx以及相应的label
anchor_idxs = torch.div(topk_idxs, self.num_classes, rounding_mode='floor') # 获取
labels = topk_idxs % self.num_classes
reg_pred = reg_pred[anchor_idxs]
anchors = anchors[anchor_idxs]
# 解算边界框, 并归一化边界框: [H*W*A, 4]
bboxes = self.decode_boxes(anchors, reg_pred)
# to cpu & numpy
scores = scores.cpu().numpy()
labels = labels.cpu().numpy()
bboxes = bboxes.cpu().numpy()
# 3、滤掉那些针对同一目标的冗余检测。
# nms
scores, labels, bboxes = multiclass_nms(
scores, labels, bboxes, self.nms_thresh, self.num_classes, self.nms_class_agnostic)
return bboxes, scores, labels
接下来,就到了正样本的匹配和损失函数计算了。
- 原版YOLOv2会先解耦出边界框,
计算边界框和目标框的IoU,只有IoU最大的才被标记为正样本
,用来计算置信度损失、类别损失以及边界框位置损失,其他预测的边界框均为负样本,仅仅计算置信度损失。 - 这样,先验框没有为正样本匹配带来直接影响,仅仅被用于解算边界框的坐标。
- 既然先验框有边界框的先验尺寸信息,那么它可以直接参与正样本的匹配,因此我们接下来采用当下更加常用的策略来发挥先验框在标签匹配中的作用,
即基于先验框的正样本匹配策略
。