Laughing

Results 196 comments of Laughing

> > 我认为是cuda和cudnn版本问题,16系列显卡没有半精度,但模型训练时候使用了半精度,导致无法计算 > > 应该不是,跑v5就没问题 我们这里复现不了这个问题,如果可以的话,你可以发我一部分数据复现一下么,十分感谢。邮箱:[email protected]

@BugOOM 有试过单独用val测试么

@BugOOM 这个预测是正确的吗? 也麻烦请用下面的命令测试一下 ```python model = YOLO("your_model.pt") model.val(data="your_data_yaml") ```

@BugOOM 可能是因为你的目标检测太小,在训练的时候过滤掉了,过滤的代码在这: [https://github.com/ultralytics/ultralytics/blob/ac628c0d3e4204236e370e399d45f6a15c2966dd/ultralytics/yolo/data/augment.py#L361](https://github.com/ultralytics/ultralytics/blob/ac628c0d3e4204236e370e399d45f6a15c2966dd/ultralytics/yolo/data/augment.py#L361) 你可以手动调整一下,或者说你关闭数据增强再试试

@BugOOM 额有可能,但是我觉得应该还是能匹配的。主要目前我们这无法复现。。如果之后你有任何发现请及时告诉我们。

@BugOOM > 有没有可能和样本匹配方式有关,对于比较细长的物体很难匹配正样本 之后我会用一些这样的数据测试一下,有任何发现我会同步给你

@BugOOM hi, may I ask if there're corrupt labels in your datasets? ![oRFtMgd2vI](https://user-images.githubusercontent.com/61612323/213076139-3ec3d352-42ca-4b8a-82ed-3834b40364c2.jpg) actually I just found that we got a mismatch issue if there's corrupt label in datasets, so...

@SkalskiP oh I think that's because there's a `clear()` operation: [https://github.com/ultralytics/ultralytics/blob/0cb87f7dd340a2611148fbf2a0af59b544bd7b1b/ultralytics/yolo/engine/predictor.py#L200](https://github.com/ultralytics/ultralytics/blob/0cb87f7dd340a2611148fbf2a0af59b544bd7b1b/ultralytics/yolo/engine/predictor.py#L200) BTW we're working on upgrading `Result`, we'll release it soon. :)

@joelwebb this happens after training right? @AyushExel I'll try to reproduce later today.

guys, I really can't reproduce the same error(can't find best.pt error) with python interface and DDP. tested command: ```python from ultralytics.yolo.engine.model import YOLO import torch model = YOLO("yolov8n.yaml") model.train(data="runs/balloon.yaml", epochs=10,...