• Ubuntu 18.04
• CUDA 11.0.3
• Python 3.8.5
• Pytorch 1.7.1
• torchvision 0.8.2

进入虚拟环境，安装YOLOv8，官方给了两种安装方式：

pip install

 pip install ultralytics

git clone

git clone https://github.com/ultralytics/ultralytics
cd ultralytics
pip install -e .

有四个文件夹dataset, images, labels, xml
dataset 数据集划分train.txt, val.txt, trainval.txt, test.txt
images 存放图片
labels 存放标注txt文件
xml 存放标注xml文件

└── one
    ├── dataset
    ├── images
    ├── labels
    └── xml

例如.img, .txt, .xml

import os
import shutil

path_xml = "/home/wanglu/detection/one/1jiwei"
filelist = os.listdir(path_xml)
path1 = "/home/wanglu/detection/one/1jiweiimages/"
path2 = "/home/wanglu/detection/one/1jiweilabels/"
path3 = "/home/wanglu/detection/one/1jiweixml/"


for files in filelist:
    filename1 = os.path.splitext(files)[1]  # 读取文件后缀名
    filename0 = os.path.splitext(files)[0]  #读取文件名
    # print(filename1)
    # m = filename1 == '.txt'
    # print(m)
    if filename1 == '.txt' :
        full_path = os.path.join(path_xml, files)
        despath = path2 + filename0 +'.txt' #.txt为你的文件类型，即后缀名，读者自行修改
        shutil.move(full_path, despath)

    if filename1 == '.jpg':
        full_path = os.path.join(path_xml, files)
        despath = path1 + filename0 + '.jpg'  #.jpg为你的文件类型，即后缀名，读者自行修改
        shutil.move(full_path, despath)

    if filename1 == '.xml':
        full_path = os.path.join(path_xml, files)
        despath = path3 + filename0 + '.xml'  # .xml为你的文件类型，即后缀名，读者自行修改
        shutil.move(full_path, despath)

新建split_train_val.py文件，生成上述dataset文件夹中train.txt, val.txt, trainval.txt, test.txt文件

# coding:utf-8
 
import os
import random
import argparse
 
parser = argparse.ArgumentParser()
#xml文件的地址，根据自己的数据进行修改 xml一般存放在Annotations下
parser.add_argument('--xml_path', default='/home/wanglu/detection/one/one/xml', type=str, help='input xml label path')
#数据集的划分，地址选择自己数据下的ImageSets/Main
parser.add_argument('--txt_path', default='/home/wanglu/detection/one/one/dataset', type=str, help='output txt label path')
opt = parser.parse_args()
 
trainval_percent = 0.9
train_percent = 0.8
xmlfilepath = opt.xml_path
txtsavepath = opt.txt_path
total_xml = os.listdir(xmlfilepath)
if not os.path.exists(txtsavepath):
    os.makedirs(txtsavepath)
 
num = len(total_xml)
list_index = range(num)
tv = int(num * trainval_percent)
tr = int(tv * train_percent)
trainval = random.sample(list_index, tv)
train = random.sample(trainval, tr)
 
file_trainval = open(txtsavepath + '/trainval.txt', 'w')
file_test = open(txtsavepath + '/test.txt', 'w')
file_train = open(txtsavepath + '/train.txt', 'w')
file_val = open(txtsavepath + '/val.txt', 'w')
 
for i in list_index:
    name ='/home/wanglu/detection/one/one/images/' + total_xml[i][:-4] + '.jpg\n'
    if i in trainval:
        file_trainval.write(name)
        if i in train:
            file_train.write(name)
        else:
            file_val.write(name)
    else:
        file_test.write(name)
 
file_trainval.close()
file_train.close()
file_val.close()
file_test.close()

在ultralytics/ultralytics/datasets/文件夹下新建dengjianji.yaml
文件路径
ultralytics/ultralytics/datasets/dengjianji.yaml

train: # train images (relative to 'path')  
  /home/wanglu/detection/one/one/dataset/trainval.txt
val: # val images (relative to 'path')  
  /home/wanglu/detection/one/one/dataset/test.txt
test: # test images (optional)
  /home/wanglu/detection/one/one/dataset/test.txt

# number of classes 
nc: 20

# Classes
names:
  0: ganjing
  1: guahua
  2: liewen_pingshen
  3: heidian
  4: liangshao
  5: jingshangfen
  6: gouya
  7: juzui
  8: gehen
  9: lvgai
  10: zhabutie
  11: gaisong
  12: quegai
  13: pingzi
  14: liewen_pingdi
  15: bianxing
  16: liang
  17: aoxian
  18: xianglianhuan
  19: ganjing_fx

至此数据集的准备已经完成，可以开始训练啦！

直接指定device=‘0, 1, 2, 3’，即可实现多卡训练

多卡训练错误可能会抛出错误：barrier() got an unexpected keyword argument ‘device_ids’

更改/root/miniconda3/lib/python3.8/site-packages/ultralytics/yolo/utils/torch_utils.py

@contextmanager
def torch_distributed_zero_first(local_rank: int):
    """Decorator to make all processes in distributed training wait for each local_master to do something."""
    initialized = torch.distributed.is_available() and torch.distributed.is_initialized()
    if initialized and local_rank not in (-1, 0):
        # dist.barrier(device_ids=[local_rank])
        dist.barrier()
    yield
    if initialized and local_rank == 0:
        # dist.barrier(device_ids=[0])
        dist.barrier()

pretrained=True即可

训练参数见下表：

source 为图像路径，或者使用摄像头source=0

新建export.py，终端运行
注意指定opset, 本人指定opset=11

from ultralytics import YOLO

model = YOLO('/home/wanglu/runs/detect/train3/weights/best.pt')
model.export(format='engine', dynamic=True, opset=11, device=0)

参数见下表：

可用的YOLOv8导出格式如下表所示。您可以使用format参数导出为任何格式，即format='onnx’或format=‘engine’。