笔记本训练的YOLOv5模型部署树莓派4B失败求解决方案

问题场景 #

在Windows笔记本上训练YOLOv5自定义数据集后，将训练好的crosswalk_detection_best.pt模型复制到树莓派4B，执行以下检测命令时出现报错：

python3 detect.py --weights /home/hasha/project/yolov5/runs/train/crosswalk_detection/weights/crosswalk_detection_best.pt --source /home/hasha/project/1.jpg

报错信息 #

detect: weights=['/home/hasha/project/yolov5/runs/train/crosswalk_detection/weights/crosswalk_detection_best.pt'], source=/home/hasha/project/1.jpg, data=data/coco128.yaml, imgsz=[640, 640], conf_thres=0.25, iou_thres=0.45, max_det=1000, device=, view_img=False, save_txt=False, save_csv=False, save_conf=False, save_crop=False, nosave=False, classes=None, agnostic_nms=False, augment=False, visualize=False, update=False, project=runs/detect, name=exp, exist_ok=False, line_thickness=3, hide_labels=False, hide_conf=False, half=False, dnn=False, vid_stride=1
YOLOv5 🚀 v7.0-335-g40f490d9 Python-3.11.2 torch-2.3.1 CPU

Traceback (most recent call last):
  File "/home/hasha/project/yolov5/detect.py", line 313, in <module>
    main(opt)
  File "/home/hasha/project/yolov5/detect.py", line 308, in main
    run(**vars(opt))
  File "/home/hasha/myenv/lib/python3.11/site-packages/torch/utils/_contextlib.py", line 115, in decorate_context
    return func(*args, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/hasha/project/yolov5/detect.py", line 116, in run
    model = DetectMultiBackend(weights, device=device, dnn=dnn, data=data, fp16=half)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/hasha/project/yolov5/models/common.py", line 467, in __init__
    model = attempt_load(weights if isinstance(weights, list) else w, device=device, inplace=True, fuse=fuse)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/hasha/project/yolov5/models/experimental.py", line 98, in attempt_load
    ckpt = torch.load(attempt_download(w), map_location="cpu")  # load
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/hasha/myenv/lib/python3.11/site-packages/torch/serialization.py", line 1025, in load
    return _load(opened_zipfile,
           ^^^^^^^^^^^^^^^^^^^^^
  File "/home/hasha/myenv/lib/python3.11/site-packages/torch/serialization.py", line 1446, in _load
    result = unpickler.load()
             ^^^^^^^^^^^^^^^^
  File "/usr/lib/python3.11/pathlib.py", line 874, in __new__
    raise NotImplementedError("cannot instantiate %r on your system"
NotImplementedError: 无法在您的系统上实例化'WindowsPath'

错误原因 #

模型在Windows系统训练时，训练日志和配置中保存了Windows格式的WindowsPath路径对象，当在Linux系统（树莓派）加载模型时，Python无法反序列化Windows专属的路径类型，导致报错。

解决方法 #

方法1：导出为ONNX格式跨平台部署 #

在Windows的YOLOv5环境中将模型导出为ONNX格式，该格式不受系统路径影响：

1. 执行导出命令：

   python export.py --weights crosswalk_detection_best.pt --include onnx
   

2. 将导出的crosswalk_detection_best.onnx复制到树莓派指定目录
3. 在树莓派上用ONNX模型执行检测：

   python3 detect.py --weights crosswalk_detection_best.onnx --source /home/hasha/project/1.jpg
   

方法2：修改模型加载代码兼容路径类型 #

直接修改树莓派上YOLOv5的模型加载逻辑，将WindowsPath映射为Linux的PosixPath：

1. 打开models/experimental.py文件
2. 找到attempt_load函数里的ckpt = torch.load(attempt_download(w), map_location="cpu")代码块，替换为以下内容：

   import pickle
   from pathlib import PosixPath
   
   class PathUnpickler(pickle.Unpickler):
       def find_class(self, module, name):
           if module == 'pathlib' and name == 'WindowsPath':
               return PosixPath
           return super().find_class(module, name)
   
   with open(attempt_download(w), 'rb') as f:
       ckpt = PathUnpickler(f).load()
   

3. 保存文件后重新执行原检测命令

方法3：在Linux/WSL环境重新训练模型 #

如果条件允许，直接在Linux系统或Windows的WSL子系统中训练模型，训练生成的模型会保存Linux兼容的PosixPath路径信息，复制到树莓派后可直接加载使用。

内容的提问来源于stack exchange，提问作者Hashan Senarathna