YOLOv8测试（1）

• 1. 训练最简流程
• • 1.1 安装包
  • 1.2 训练demo
  • 1.3 验证
• 参考文献资料

1. 训练最简流程

对于多数用户来说，都是面向边缘端的项目，实际上，日常监测任务也不要求特别高的训练精度，为了能够快速训练，这里选择Github中提到的YOLOv8n模型。

1.1 安装包

pip install ultralytics

1.2 训练demo

官方给的demo似乎对于Pycharm用户会有小报错，使用那个网页版的jupyter notebook似乎不会，应该就是线程的问题

• 官方代码

from ultralytics import YOLO# Load a model
model = YOLO("yolov8n.yaml")  # build a new model from scratch
model = YOLO("yolov8n.pt")  # load a pretrained model (recommended for training)# Use the model
results = model.train(data="coco128.yaml", epochs=3)  # train the model

• 修改的，增加了main函数就行

from ultralytics import YOLO# Load a model
def main():model = YOLO("yolov8n.yaml")  # build a new model from scratchmodel = YOLO("yolov8n.pt")  # load a pretrained model (recommended for training)# model.train(data="coco128.yaml", epochs=5)if __name__ == '__main__':main()

1.3 验证

• 完整代码

from ultralytics import YOLO# Load a model
def main():model = YOLO("yolov8n.yaml")  # build a new model from scratchmodel = YOLO("yolov8n.pt")  # load a pretrained model (recommended for training)# model.train(data="coco128.yaml", epochs=5)results = model.val()  # evaluate model performance on the validation setif __name__ == '__main__':main()

• 日志

Image sizes 640 train, 640 val
Using 8 dataloader workers
Logging results to runs\detect\train9
Starting training for 5 epochs...Epoch    GPU_mem   box_loss   cls_loss   dfl_loss  Instances       Size1/5      5.85G      1.213      1.429      1.258        215        640: 100%|██████████| 8/8 [00:05<00:00,  1.58it/s]Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 4/4 [00:36<00:00,  9.12s/it]all        128        929      0.668       0.54      0.624      0.461Epoch    GPU_mem   box_loss   cls_loss   dfl_loss  Instances       Size2/5      6.87G      1.156      1.327      1.243        163        640: 100%|██████████| 8/8 [00:04<00:00,  1.64it/s]Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 4/4 [00:35<00:00,  8.91s/it]all        128        929      0.667      0.589      0.651      0.487Epoch    GPU_mem   box_loss   cls_loss   dfl_loss  Instances       Size...

• 预测结果
  返回的信息也挺全面的，包括Boxes 原始图像 尺寸之类的

results:[{ '_keys': . at 0x0000023FA2FC0350>,'boxes': ultralytics.yolo.engine.results.Boxes
type:  torch.Tensor
shape: torch.Size([6, 6])
dtype: torch.float32
tensor([[2.40000e+01, 2.26000e+02, 8.02000e+02, 7.58000e+02, 8.75480e-01, 5.00000e+00],[4.80000e+01, 3.97000e+02, 2.46000e+02, 9.06000e+02, 8.74487e-01, 0.00000e+00],[6.70000e+02, 3.79000e+02, 8.10000e+02, 8.77000e+02, 8.53311e-01, 0.00000e+00],[2.19000e+02, 4.06000e+02, 3.44000e+02, 8.59000e+02, 8.16101e-01, 0.00000e+00],[0.00000e+00, 2.54000e+02, 3.20000e+01, 3.25000e+02, 4.91605e-01, 1.10000e+01],[0.00000e+00, 5.50000e+02, 6.40000e+01, 8.76000e+02, 3.76493e-01, 0.00000e+00]], device='cuda:0'),'masks': None,'names': { 0: 'person',1: 'bicycle',...79: 'toothbrush'},'orig_img': array([[[122, 148, 172],[120, 146, 170],[125, 153, 177],...,...,[ 99,  89,  95],[ 96,  86,  92],[102,  92,  98]]], dtype=uint8),'orig_shape': (1080, 810),'path': '..\\bus.jpg','probs': None,'speed': {'inference': 30.916452407836914, 'postprocess': 2.992391586303711, 'preprocess': 3.988981246948242}}]

• 可视化结果
  

小结：

1. 避免多线程导致的程序报错

错误：This probably means that you are not using fork to start your
child processes and you have forgotten to use the proper idiom
in the main module:

        if __name__ == '__main__':freeze_support()...The "freeze_support()" line can be omitted if the programis not going to be frozen to produce an executable.

2. 训练注意关闭KX上网工具，否则模型文件可能下载失败
   错误：requests.exceptions.ProxyError: HTTPSConnectionPool(host=‘pypi.org’, port=443): Max retries exceeded with url: /pypi/ultralytics/json (Caused by ProxyError(‘Cannot connect to proxy.’, FileNotFoundError(2, ‘No such file or directory’)))
3. opencv包报错：具体可参考这篇【2】
   The function is not implemented. Rebuild the library with Windows, GTK+ 2.x or Cocoa support.

参考文献资料

【1】YOLOv8官方Github
【2】https://stackoverflow.com/questions/74035760/opencv-waitkey-throws-assertion-rebuild-the-library-with-windows-gtk-2-x-or