Laughing

@Tsardoz so you edited the source code? which file did you edit ? actually there're two dataloaders v8 loader and v5 loader(v5loader is for debug for now), I think you...

@Tsardoz oh sorry I didn't see you've mentioned the file. yes the `yolo/data/augment.py` is the correct path for v8loader. And I just did what you tried, edited `T=[A.ImageCompression(quality_lower=75, p=0.5),]` before...

@Tsardoz it works well to me...  output:  command: ```python yolo task=detect mode=train model=yolov8s.pt data=coco.yaml device=\'0,1\' ```

@Tsardoz yes, maybe you can also install the latest version and try again. :)

hi @mzhen77, yeah we treat model.predict() as a generator now, we added a `return_outputs` arg to make sure users can get outputs and do whatever they want with it. Sorry...

@maciejkr no worries! that's how tracker algorithms work, matching predictions of current frame and predictions from previous frame, hence the predictions from first frame don't have IDs yet.

@initdebugs looks like a dataloader issue, please try it again with `v5loader=True`.

@initdebugs could you please tell me more about this issue? like is there negative labels or empty labels in you custom dataset? I'd like to reproduce your issue and solve...

@initdebugs thanks!

@initdebugs hi, I reproduced your error and checked your dataset. Most of your labels are detection labels which include only five columns but some of your labels are segmentation labels...