yolov8

开始安装

• https://zhuanlan.zhihu.com/p/603272140
• YOLOv8 教程系列！

注意到安装比较简单，克隆项目，准备好 pytorch 环境，安装一个依赖即可。

尝试运行的命令？

• https://blog.csdn.net/tingwhere/article/details/130394227

• try1

yolo detect mode=predict model=yolov8l.pt source=ultralytics/assets/bus.jpg

注意到该命令是相对于项目根目录下的绝对路径。执行完毕后，项目根目录下会出现一个名为 yolov8l.pt 的文件。

• try2

yolo detect mode=predict

该命令不会去调用 cfg 配置声明的额外覆盖配置。没有找到合适的方式，实现复杂命令参数的统一配置。

根据教程开始搭建环境

• YOLOv8 教程系列！

笔记本电脑可用的命令行：

训练

yolo task=detect mode=train model=models/yolov8n.pt data=business/-temp-try/conf.yaml project=business/-temp-try/res batch=8 epochs=1 workers=8 imgsz=640

验证

yolo task=detect mode=val model=business/-temp-try/res/train2/weights/best.pt data=business/-temp-try/conf.yaml project=business/-temp-try/res

预测

yolo task=detect mode=predict model=business/-temp-try/res/train2/weights/best.pt source=business/-temp-try/images project=business/-temp-try/res

目前的工作流

• 打标。不管用哪种打标工具，只要结果用 VOC，xml 格式导出即可。这远比单纯的 yolo，text 格式要好得多。
• 数据处理。拆分数据集和格式转换。事实上可以随便更改其次序。
  • splitDataset.py 拆分数据
  • xml-2-txt-chatgpt-v1.py 数据格式转换 同时录入打标的标签。

linux 训练

使用同一套运行命令。

linux 环境下常用的命令行

conda activate pytorch
cd /home/ai_rzn/code/ultralytics/
nohup 运行命令 &

查看指定用户的进程

ps -u root

在指定目录内，查看日志。这里默认特指 /home/ai_rzn/code/yolov5-master/ 目录

cat nohup.out

常用的路径

数据集存储在 business 文件夹内

/home/ai_rzn/code/ultralytics/business/

/home/ai_rzn/code/ultralytics/

yolo task=detect mode=train model=models/yolov8n.pt data=business/person-drop-litter/conf.yaml project=business/person-drop-litter/res batch=8 epochs=1 workers=8 imgsz=640

训练：

nohup yolo task=detect mode=train model=models/yolov8n.pt data=business/person-drop-litter/conf.yaml project=business/person-drop-litter/res batch=8 epochs=350 workers=8 imgsz=640 &

预测：

yolo task=detect mode=predict model=business/person-drop-litter/res/train2/weights/best.pt source=business/person-drop-litter/images project=business/person-drop-litter/res

ImportError: Cannot load backend 'TkAgg' which requires the 'tk' interactive framework, as 'headless' is currently running

在 linux 内出现此问题。在训练到最后时出现该错误：

Traceback (most recent call last):
  File "/home/anaconda3/envs/pytorch/bin/yolo", line 33, in <module>
    sys.exit(load_entry_point('ultralytics', 'console_scripts', 'yolo')())
  File "/home/ai_rzn/code/ultralytics/ultralytics/yolo/cfg/__init__.py", line 398, in entrypoint
    getattr(model, mode)(**overrides)  # default args from model
  File "/home/ai_rzn/code/ultralytics/ultralytics/yolo/engine/model.py", line 371, in train
    self.trainer.train()
  File "/home/ai_rzn/code/ultralytics/ultralytics/yolo/engine/trainer.py", line 192, in train
    self._do_train(world_size)
  File "/home/ai_rzn/code/ultralytics/ultralytics/yolo/engine/trainer.py", line 370, in _do_train
    self.metrics, self.fitness = self.validate()
  File "/home/ai_rzn/code/ultralytics/ultralytics/yolo/engine/trainer.py", line 476, in validate
    metrics = self.validator(self)
  File "/home/anaconda3/envs/pytorch/lib/python3.9/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
  File "/home/ai_rzn/code/ultralytics/ultralytics/yolo/engine/validator.py", line 177, in __call__
    stats = self.get_stats()
  File "/home/ai_rzn/code/ultralytics/ultralytics/yolo/v8/detect/val.py", line 128, in get_stats
    self.metrics.process(*stats)
  File "/home/ai_rzn/code/ultralytics/ultralytics/yolo/utils/metrics.py", line 710, in process
    results = ap_per_class(tp,
  File "/home/ai_rzn/code/ultralytics/ultralytics/yolo/utils/metrics.py", line 523, in ap_per_class
    plot_pr_curve(px, py, ap, save_dir / f'{prefix}PR_curve.png', names, on_plot=on_plot)
  File "/home/ai_rzn/code/ultralytics/ultralytics/yolo/utils/__init__.py", line 197, in wrapper
    plt.switch_backend(original_backend)
  File "/home/anaconda3/envs/pytorch/lib/python3.9/site-packages/matplotlib/pyplot.py", line 279, in switch_backend
    raise ImportError(
ImportError: Cannot load backend 'TkAgg' which requires the 'tk' interactive framework, as 'headless' is currently running

• https://www.coder.work/article/6322766
• https://www.cnblogs.com/gy77/p/15740857.html
• https://github.com/open-mmlab/mmdetection/issues/5885

# 显示包版本
pip show matplotlib
# 移除
pip uninstall matplotlib
# 指定安装定版本
pip install matplotlib==3.2.1

经过一系列的配置后，无论是 yolov5、slowfast、还是 yolov8。这些模型的多平台部署都让人非常恼火。现在打算学习 docker 来一次性完成环境部署了。累了。

其他同事在 matplotlib3.2.1 版本下，成功运行，其运行命令如下：

yolo task=detect mode=train model=yolov8n.pt data=business/person-drop-litter/conf.yaml batch=4 epochs=200 imgsz=640 workers=0

YOLOv5 和 YOLOv8 使用体验

YOLOv8 用起来更加轻松，不用动脑子。

• 项目运行根目录
  /home/ai_rzn/code/ultralytics/

• 图片文件路径
  /home/ai_rzn/code/ultralytics/business/person-drop-litter/images/

• 训练配置文件路径
  /home/ai_rzn/code/ultralytics/business/person-drop-litter/conf.yaml

• 实际执行的命令
  nohup yolo task=detect mode=train model=models/yolov8n.pt data=business/person-drop-litter/conf.yaml project=business/person-drop-litter/res batch=8 epochs=350 workers=8 imgsz=640 &