fix nan/inf loss

[Laughing-q]

@AyushExel @glenn-jocher this PR fixed the nan loss issue. And the cause is that the target_scores_sum we're using in loss calculation could be 0 if there're no objects in targets(empty labels, background).

test command:

yolo detect train model=yolov8n.pt data=custom.yaml imgsz=640 rect=True

before fix(losses are nan and mAP stays 0.15):

after fix:

🛠️ PR Summary

_{Made with ❤️ by Ultralytics Actions}

🌟 Summary

Enhanced stability in loss calculation during model training.

📊 Key Changes

• Modified the calculation of target_scores_sum by ensuring it's never less than 1.

🎯 Purpose & Impact

• 🛠 Prevents division by zero errors when calculating loss, increasing the robustness of training.
• ⚖️ Contributors and users of the Ultralytics framework can expect more stable training sessions, particularly in edge cases with sparse detection targets.

[hdnh2006]

@Laughing-q I am still getting nan in my training. It seems for validation is solved:

After running pip install --upgrade ultralytics I get the following:

Looking in indexes: https://pypi.org/simple, https://pypi.ngc.nvidia.com
Requirement already satisfied: ultralytics in ********/.virtualenvs/ultralytics/lib/python3.8/site-packages (8.0.10)

[AyushExel]

@hdnh2006 wait for a while till this gets merged. We'll release the updated package later today

[CoderYiFei]

well done @Laughing-q
I always get nan value in training dataset , and valid dataset.
thanks very much.
I am waiting for the master' s code, that fixed the nan value bug.

[glenn-jocher]

@Laughing-q do you know what a typical value of target_scores_sum is? Could this change affect COCO training results?

Should we add a smaller value for a protected divide, i.e. x / (target_scores_sum + eps) where eps might be something like 1e-6 or 1e-3?

[glenn-jocher]

@Laughing-q I debugged this value on COCO128 and it's very large, i.e. target_scores_sum = 1000 at batch-size 16, so I think this is fine to max at 1.

[Laughing-q]

@glenn-jocher oh yes, I just want to tell you this.