Yolov8详解与实战

摘要

YOLOv8 是 ultralytics 公司在 2023 年 1月 10 号开源的 YOLOv5 的下一个重大更新版本，目前支持图像分类、物体检测和实例分割任务，鉴于Yolov5的良好表现，Yolov8在还没有开源时就收到了用户的广泛关注。yolov8的整体架构如下：

Yolov8的改进之处有以下几个地方：

• Backbone：使用的依旧是CSP的思想，将YOLOv5中的C3模块被替换成了C2f模块，实现了进一步的轻量化，同时YOLOv8依旧使用了YOLOv5等架构中使用的SPPF模块；
• PAN-FPN：YOLOv8依旧使用了PAN的思想，不同的是YOLOv8将YOLOv5中PAN-FPN上采样阶段中的卷积结构删除了，同时也将C3模块替换为了C2f模块；
• Decoupled-Head：这一点源自YOLOX；分类和回归两个任务的head不再共享参数，YoloV8也借鉴了这样的head设计。
• Anchor-Free：YOLOv8抛弃了以往的Anchor-Base，使用了Anchor-Free的思想；
• 损失函数：YOLOv8使用VFL Loss作为分类损失，使用DFL Loss+CIOU Loss作为分类损失；
• 样本匹配：YOLOv8抛弃了以往的IOU匹配或者单边比例的分配方式，而是使用了Task-Aligned Assigner匹配方式。

yolov8是个模型簇，从小到大包括：yolov8n、yolov8s、yolov8m、yolov8l、yolov8x等。模型参数、运行速度、参数量等详见下表：

对比yolov5，如下表：

mAP和参数量都上升了不少，具体的感受还是要亲自实践一番。

这篇文章首先对YoloV8做详细的讲解，然后实现对COCO数据集的训练和测试，最后，实现自定义数据集的训练和测试。
希望能帮助到朋友们！

分割的结果

分类的结果

模型详解

C2F模块

yolov8将yolov5中的C3模块换成了C2F模型，我们先了解一下C3模块，如图：

C3模块，其主要是借助CSPNet提取分流的思想，同时结合残差结构的思想，设计了所谓的C3 Block，这里的CSP主分支梯度模块为BottleNeck模块，堆叠的个数由参数n来进行控制，不同的模型，n的个数也不相同。C3的pytorch代码如下：

class C3(nn.Module):# CSP Bottleneck with 3 convolutionsdef __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):  # ch_in, ch_out, number, shortcut, groups, expansionsuper().__init__()c_ = int(c2 * e)  # hidden channelsself.cv1 = Conv(c1, c_, 1, 1)self.cv2 = Conv(c1, c_, 1, 1)self.cv3 = Conv(2 * c_, c2, 1)  # optional act=FReLU(c2)self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, k=((1, 1), (3, 3)), e=1.0) for _ in range(n)))def forward(self, x):return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), 1))

接下来，我们一起学习C2F模块，先经过一个Conv，然后使用chunk函数将out平均拆分成两个向量，然后保存到list中，将后半部分输入到Bottleneck Block里面，Bottleneck Block里面有n个Bottleneck，将每个Bottleneck的输出都追加list中，有n个，所以拼接之后的out等于0.5✖(n+2)。然后经过一个Conv输出，所以输出为h×w×c_out。如下图：

如果还是比较难懂，我将具体的数据代入图中，得出下图：

Loss

对于YOLOv8，其分类损失为VFL Loss，其回归损失为CIOU Loss+DFL的形式，这里Reg_max默认为16。

VFL主要改进是提出了非对称的加权操作，FL和QFL都是对称的。而非对称加权的思想来源于论文PISA，该论文指出首先正负样本有不平衡问题，即使在正样本中也存在不等权问题，因为mAP的计算是主正样本。

q是label，正样本时候q为bbox和gt的IoU，负样本时候q=0，当为正样本时候其实没有采用FL，而是普通的BCE，只不过多了一个自适应IoU加权，用于突出主样本。而为负样本时候就是标准的FL了。可以明显发现VFL比QFL更加简单，主要特点是正负样本非对称加权、突出正样本为主样本。

针对这里的DFL（Distribution Focal Loss），其主要是将框的位置建模成一个 general distribution，让网络快速的聚焦于和目标位置距离近的位置的分布。

DFL 能够让网络更快地聚焦于目标 y 附近的值，增大它们的概率；

DFL的含义是以交叉熵的形式去优化与标签y最接近的一左一右2个位置的概率，从而让网络更快的聚焦到目标位置的邻近区域的分布；也就是说学出来的分布理论上是在真实浮点坐标的附近，并且以线性插值的模式得到距离左右整数坐标的权重。

head部分

相对于YOLOv5，YOLOv8将Head里面C3模块替换为了C2f，将上采样之前的1×1卷积去除了，将Backbone不同阶段输出的特征直接送入了上采样操作。通过下图对比可以看出差别：

YOLOv8则是使用了Decoupled-Head，同时由于使用了DFL 的思想，因此回归头的通道数也变成了4*reg_max的形式：

模型实战

训练COCO数据集

本次使用2017版本的COCO数据集作为例子，演示如何使用YoloV8训练和预测。

下载数据集

Images:

• 2017 Train images [118K/18GB] ：http://images.cocodataset.org/zips/train2017.zip
• 2017 Val images [5K/1GB]：http://images.cocodataset.org/zips/val2017.zip
• 2017 Test images [41K/6GB]：http://images.cocodataset.org/zips/unlabeled2017.zip

Annotations:

• 2017 annotations_trainval2017 [241MB]：http://images.cocodataset.org/annotations/annotations_trainval2017.zip

COCO转yolo格式数据集（适用V4，V5，V6，V7，V8）

最初的研究论文中，COCO中有91个对象类别。然而，在2014年的第一次发布中，仅发布了80个标记和分割图像的对象类别。2014年发布之后，2017年发布了后续版本。详细的类别如下：

ID	OBJECT (PAPER)	OBJECT (2014 REL.)	OBJECT (2017 REL.)	SUPER CATEGORY
1	person	person	person	person
2	bicycle	bicycle	bicycle	vehicle
3	car	car	car	vehicle
4	motorcycle	motorcycle	motorcycle	vehicle
5	airplane	airplane	airplane	vehicle
6	bus	bus	bus	vehicle
7	train	train	train	vehicle
8	truck	truck	truck	vehicle
9	boat	boat	boat	vehicle
10	trafficlight	traffic light	traffic light	outdoor
11	fire hydrant	fire hydrant	fire hydrant	outdoor
12	street	sign	-	-
13	stop sign	stop sign	stop sign	outdoor
14	parking meter	parking meter	parking meter	outdoor
15	bench	bench	bench	outdoor
16	bird	bird	bird	animal
17	cat	cat	cat	animal
18	dog	dog	dog	animal
19	horse	horse	horse	animal
20	sheep	sheep	sheep	animal
21	cow	cow	cow	animal
22	elephant	elephant	elephant	animal
23	bear	bear	bear	animal
24	zebra	zebra	zebra	animal
25	giraffe	giraffe	giraffe	animal
26	hat	-	-	accessory
27	backpack	backpack	backpack	accessory
28	umbrella	umbrella	umbrella	accessory
29	shoe	-	-	accessory
30	eye glasses	-	-	accessory
31	handbag	handbag	handbag	accessory
32	tie	tie	tie	accessory
33	suitcase	suitcase	suitcase	accessory
34	frisbee	frisbee	frisbee	sports
35	skis	skis	skis	sports
36	snowboard	snowboard	snowboard	sports
37	sports ball	sports ball	sports ball	sports
38	kite	kite	kite	sports
39	baseball bat	baseball bat	baseball bat	sports
40	baseball glove	baseball glove	baseball glove	sports
41	skateboard	skateboard	skateboard	sports
42	surfboard	surfboard	surfboard	sports
43	tennis racket	tennis racket	tennis racket	sports
44	bottle	bottle	bottle	kitchen
45	plate	-	-	kitchen
46	wine glass	wine glass	wine glass	kitchen
47	cup	cup	cup	kitchen
48	fork	fork	fork	kitchen
49	knife	knife	knife	kitchen
50	spoon	spoon	spoon	kitchen
51	bowl	bowl	bowl	kitchen
52	banana	banana	banana	food
53	apple	apple	apple	food
54	sandwich	sandwich	sandwich	food
55	orange	orange	orange	food
56	broccoli	broccoli	broccoli	food
57	carrot	carrot	carrot	food
58	hot dog	hot dog	hot dog	food
59	pizza	pizza	pizza	food
60	donut	donut	donut	food
61	cake	cake	cake	food
62	chair	chair	chair	furniture
63	couch	couch	couch	furniture
64	potted plant	potted plant	potted plant	furniture
65	bed	bed	bed	furniture
66	mirror	-	-	furniture
67	dining table	dining table	dining table	furniture
68	window	-	-	furniture
69	desk	-	-	furniture
70	toilet	toilet	toilet	furniture
71	door	-	-	furniture
72	tv	tv	tv	electronic
73	laptop	laptop	laptop	electronic
74	mouse	mouse	mouse	electronic
75	remote	remote	remote	electronic
76	keyboard	keyboard	keyboard	electronic
77	cell phone	cell phone	cell phone	electronic
78	microwave	microwave	microwave	appliance
79	oven	oven	oven	appliance
80	toaster	toaster	toaster	appliance
81	sink	sink	sink	appliance
82	refrigerator	refrigerator	refrigerator	appliance
83	blender	-	-	appliance
84	book	book	book	indoor
85	clock	clock	clock	indoor
86	vase	vase	vase	indoor
87	scissors	scissors	scissors	indoor
88	teddy bear	teddy bear	teddy bear	indoor
89	hair drier	hair drier	hair drier	indoor
90	toothbrush	toothbrush	toothbrush	indoor
91	hair brush	-	-	indoor

可以看到，2014年和2017年发布的对象列表是相同的，它们是论文中最初91个对象类别中的80个对象。所以在转换的时候，要重新对类别做映射，映射函数如下：

def coco91_to_coco80_class():  # converts 80-index (val2014) to 91-index (paper)# https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/# a = np.loadtxt('data/coco.names', dtype='str', delimiter='\n')# b = np.loadtxt('data/coco_paper.names', dtype='str', delimiter='\n')# x1 = [list(a[i] == b).index(True) + 1 for i in range(80)]  # darknet to coco# x2 = [list(b[i] == a).index(True) if any(b[i] == a) else None for i in range(91)]  # coco to darknetx = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, None, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, None, 24, 25, None,None, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, None, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50,51, 52, 53, 54, 55, 56, 57, 58, 59, None, 60, None, None, 61, None, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72,None, 73, 74, 75, 76, 77, 78, 79, None]return x

接下来，开始格式转换，工程的目录如下：

• coco：存放解压后的数据集。
  -out：保存输出结果。
  -coco2yolo.py：转换脚本。

转换代码如下：

import json
import glob
import os
import shutil
from pathlib import Path
import numpy as np
from tqdm import tqdmdef make_folders(path='../out/'):# Create foldersif os.path.exists(path):shutil.rmtree(path)  # delete output folderos.makedirs(path)  # make new output folderos.makedirs(path + os.sep + 'labels')  # make new labels folderos.makedirs(path + os.sep + 'images')  # make new labels folderreturn pathdef convert_coco_json(json_dir='./coco/annotations_trainval2017/annotations/'):jsons = glob.glob(json_dir + '*.json')coco80 = coco91_to_coco80_class()# Import jsonfor json_file in sorted(jsons):fn = 'out/labels/%s/' % Path(json_file).stem.replace('instances_', '')  # folder namefn_images = 'out/images/%s/' % Path(json_file).stem.replace('instances_', '')  # folder nameos.makedirs(fn,exist_ok=True)os.makedirs(fn_images,exist_ok=True)with open(json_file) as f:data = json.load(f)print(fn)# Create image dictimages = {'%g' % x['id']: x for x in data['images']}# Write labels filefor x in tqdm(data['annotations'], desc='Annotations %s' % json_file):if x['iscrowd']:continueimg = images['%g' % x['image_id']]h, w, f = img['height'], img['width'], img['file_name']file_path='coco/'+fn.split('/')[-2]+"/"+f# The Labelbox bounding box format is [top left x, top left y, width, height]box = np.array(x['bbox'], dtype=np.float64)box[:2] += box[2:] / 2  # xy top-left corner to centerbox[[0, 2]] /= w  # normalize xbox[[1, 3]] /= h  # normalize yif (box[2] > 0.) and (box[3] > 0.):  # if w > 0 and h > 0with open(fn + Path(f).stem + '.txt', 'a') as file:file.write('%g %.6f %.6f %.6f %.6f\n' % (coco80[x['category_id'] - 1], *box))file_path_t=fn_images+fprint(file_path,file_path_t)shutil.copy(file_path,file_path_t)def coco91_to_coco80_class():  # converts 80-index (val2014) to 91-index (paper)# https://tech.amikelive.com/node-718/what-object-categories-labels-are-in-coco-dataset/# a = np.loadtxt('data/coco.names', dtype='str', delimiter='\n')# b = np.loadtxt('data/coco_paper.names', dtype='str', delimiter='\n')# x1 = [list(a[i] == b).index(True) + 1 for i in range(80)]  # darknet to coco# x2 = [list(b[i] == a).index(True) if any(b[i] == a) else None for i in range(91)]  # coco to darknetx = [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, None, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, None, 24, 25, None,None, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, None, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50,51, 52, 53, 54, 55, 56, 57, 58, 59, None, 60, None, None, 61, None, 62, 63, 64, 65, 66, 67, 68, 69, 70, 71, 72,None, 73, 74, 75, 76, 77, 78, 79, None]return xconvert_coco_json()

开始运行：

转换完成后，验证转换的结果：

import cv2
import osdef draw_box_in_single_image(image_path, txt_path):# 读取图像image = cv2.imread(image_path)# 读取txt文件信息def read_list(txt_path):pos = []with open(txt_path, 'r') as file_to_read:while True:lines = file_to_read.readline()  # 整行读取数据if not lines:break# 将整行数据分割处理，如果分割符是空格，括号里就不用传入参数，如果是逗号， 则传入‘，'字符。p_tmp = [float(i) for i in lines.split(' ')]pos.append(p_tmp)  # 添加新读取的数据# Efield.append(E_tmp)passreturn pos# txt转换为boxdef convert(size, box):xmin = (box[1]-box[3]/2.)*size[1]xmax = (box[1]+box[3]/2.)*size[1]ymin = (box[2]-box[4]/2.)*size[0]ymax = (box[2]+box[4]/2.)*size[0]box = (int(xmin), int(ymin), int(xmax), int(ymax))return boxpos = read_list(txt_path)print(pos)tl = int((image.shape[0]+image.shape[1])/2)lf = max(tl-1,1)for i in range(len(pos)):label = str(int(pos[i][0]))print('label is '+label)box = convert(image.shape, pos[i])image = cv2.rectangle(image,(box[0], box[1]),(box[2],box[3]),(0,0,255),2)cv2.putText(image,label,(box[0],box[1]-2), 0, 1, [0,0,255], thickness=2, lineType=cv2.LINE_AA)passif pos:cv2.imwrite('./Data/see_images/{}.png'.format(image_path.split('\\')[-1][:-4]), image)else:print('None')img_folder = "./out/images/val2017"
img_list = os.listdir(img_folder)
img_list.sort()label_folder = "./out/labels/val2017"
label_list = os.listdir(label_folder)
label_list.sort()
if not os.path.exists('./Data/see_images'):os.makedirs('./Data/see_images')
for i in range(len(img_list)):image_path = img_folder + "\\" + img_list[i]txt_path = label_folder + "\\" + label_list[i]draw_box_in_single_image(image_path, txt_path)

结果展示：

配置yolov8环境

可以直接安装requirements.txt里面所有的库文件，执行安装命令：

pip install -r requirements.txt

如果不想安装这么多库文件，在运行的时候，查看缺少哪个库，就安装哪个库，比如我的环境：

pip install thop

我的本地只缺少了这个库文件。

训练

下载代码：https://github.com/ultralytics/ultralytics，通过下载的方式可以下载到源码，这样方便修改。
也可以使用命令：

pip install ultralytics

如果仅仅是为了使用yolov8，可以使用这种方式安装。

yolov8还支持使用命令的方式，例如：

yolo predict model=yolov8n.pt source="https://ultralytics.com/images/bus.jpg"

接下来，创建训练脚本，可以使用yaml文件创建，例如：

from ultralytics import YOLO# Load a model
model = YOLO("yolov8n.yaml")  # build a new model from scratch

模型文件在ultralytics/models/v8下面，如图：

也可以使用预训练模型创建。例如：

model = YOLO("yolov8n.pt")  # load a pretrained model (recommended for training)

然后开启训练。

# Use the model
model.train(data="coco128.yaml", epochs=3)  # train the model

数据集的配置文件在：ultralytics/datasets/下面，如图：

是不是很简单！！！！

接下来，我们配置自己的环境。
第一步 找到ultralytics/datasets/coco.yaml文件。

然后将其复制到根目录

将里面的路径修改为：

# Ultralytics YOLO 🚀, GPL-3.0 license
# COCO 2017 dataset http://cocodataset.org by Microsoft
# Example usage: yolo train data=coco.yaml
# parent
# ├── ultralytics
# └── datasets
#     └── coco  ← downloads here (20.1 GB)# Train/val/test sets as 1) dir: path/to/imgs, 2) file: path/to/imgs.txt, or 3) list: [path/to/imgs1, path/to/imgs2, ..]train: ./coco/images/train2017  # train images (relative to 'path') 118287 images
val: ./coco/images/val2017  # val images (relative to 'path') 5000 images
test: test-dev2017.txt  # 20288 of 40670 images, submit to https://competitions.codalab.org/competitions/20794

关于数据集的路径，大家可以自行尝试，我经过多次尝试发现，YoloV8会自行添加datasets这个文件，所以设置./coco/images/train2017,则实际路径是datasets/coco/images/train2017。
第二步 新建train.py脚本。

from ultralytics import YOLO# 加载模型
model = YOLO("ultralytics/models/v8/yolov8n.yaml")  # 从头开始构建新模型# Use the model
results = model.train(data="coco.yaml", epochs=3,device='3')  # 训练模型

然后，点击train.py可以运行了。
如果设置多卡，可以在device中设置，例如我使用四张卡，可以设置为：

results = model.train(data="coco.yaml", epochs=3,device='0,1,2,3')  # 训练模型

第三步 修改参数，在ultralytics/yolo/cfg/default.yaml文件中查看。例如：

# Train settings -------------------------------------------------------------------------------------------------------
model:  # path to model file, i.e. yolov8n.pt, yolov8n.yaml
data:  # path to data file, i.e. coco128.yaml
epochs: 100  # number of epochs to train for
patience: 50  # epochs to wait for no observable improvement for early stopping of training
batch: 16  # number of images per batch (-1 for AutoBatch)
imgsz: 640  # size of input images as integer or w,h
save: True  # save train checkpoints and predict results
save_period: -1 # Save checkpoint every x epochs (disabled if < 1)
cache: False  # True/ram, disk or False. Use cache for data loading
device:  # device to run on, i.e. cuda device=0 or device=0,1,2,3 or device=cpu
workers: 8  # number of worker threads for data loading (per RANK if DDP)
project:  # project name
name:  # experiment name, results saved to 'project/name' directory
exist_ok: False  # whether to overwrite existing experiment
pretrained: False  # whether to use a pretrained model
optimizer: SGD  # optimizer to use, choices=['SGD', 'Adam', 'AdamW', 'RMSProp']
verbose: True  # whether to print verbose output
seed: 0  # random seed for reproducibility
deterministic: True  # whether to enable deterministic mode
single_cls: False  # train multi-class data as single-class
image_weights: False  # use weighted image selection for training
rect: False  # support rectangular training if mode='train', support rectangular evaluation if mode='val'
cos_lr: False  # use cosine learning rate scheduler
close_mosaic: 10  # disable mosaic augmentation for final 10 epochs
resume: False  # resume training from last checkpoint

上面是训练过程中常用的参数，我们调用yolo函数可以自行修改。
等待测试完成后，就可以看到结果，如下图：

测试

新建测试脚本test.py.

from ultralytics import YOLO# Load a model
model = YOLO("runs/detect/train11/weights/best.pt")  # load a pretrained model (recommended for training)results = model.predict(source="ultralytics/assets",device='3')  # predict on an image
print(results)

这个results保存了所有的结果。如下图：

predict的参数也可以在ultralytics/yolo/cfg/default.yaml文件中查看。例如：

# Prediction settings --------------------------------------------------------------------------------------------------
source:  # source directory for images or videos
show: False  # show results if possible
save_txt: False  # save results as .txt file
save_conf: False  # save results with confidence scores
save_crop: False  # save cropped images with results
hide_labels: False  # hide labels
hide_conf: False  # hide confidence scores
vid_stride: 1  # video frame-rate stride
line_thickness: 3  # bounding box thickness (pixels)
visualize: False  # visualize model features
augment: False  # apply image augmentation to prediction sources
agnostic_nms: False  # class-agnostic NMS
classes:  # filter results by class, i.e. class=0, or class=[0,2,3]
retina_masks: False  # use high-resolution segmentation masks
boxes: True  # Show boxes in segmentation predictions

训练自定义数据集

Labelme数据集