Previous trained model giving different predictions in newer ultralytics

[lastmarchoftheents]

Hi,
I have trained a custom YOLOv8 model in Ultralytics v8.0.43 (in a custom dataset).
These are the predictions for a specific image:

Now, I upgraded the ultralytics version to v8.2.74 and loaded the same model.
These are the predictions for the same image:

Those differences are making our pipeline fail. Is there any reason that explains this behaviour? Is there any way to fix it, i.e., to obtain the same predictions in both ultralytics?

EDIT: I have tested, and the different results come from v8.0.53

[UltralyticsAssistant]

👋 Hello @lastmarchoftheents, thank you for bringing this to our attention 🚀! This is an automated response, and an Ultralytics engineer will assist you soon.

To better understand the issue with your predictions, please ensure you've checked the following:

• Upgrade your package to the latest version: Execute pip install -U ultralytics to ensure all new fixes and improvements are applied.
• Provide a minimum reproducible example: This will greatly assist us in diagnosing the problem.

🔄 If you've encountered this while custom training, double-check your setup against our Tips for Best Training Results.

Additionally, feel free to join our growing Ultralytics community for more interactive support:

• Chat in real time on Discord 🎧
• Engage in discussions on Discourse
• Explore threads with other users on our Subreddit

Upgrade Instructions

Ensure you're running an updated environment. Upgrade all requirements in a Python>=3.8 environment with PyTorch>=1.8:

pip install -U ultralytics

Environments

YOLO can be executed in various verified environments:

• Notebooks:
• Google Cloud: Check the GCP Quickstart Guide
• AWS: Follow the AWS Quickstart Guide
• Docker: Review the Docker Quickstart Guide

Status

Always ensure the Ultralytics CI badge is green for verified correct operations.

Thank you for your patience and support! 😊

[glenn-jocher]

Version updates can introduce changes affecting model predictions. To maintain consistency, consider using the same version for both training and inference. You can also review the release notes for any changes that might impact your results. For further guidance, visit our common issues page: https://docs.ultralytics.com/guides/yolo-common-issues/.

[lastmarchoftheents]

Found a solution. Just if anybody needs to obtain the original prediction (v8.0.43) you should change 2 files:

ultralytics/models/yolo/detect/predict.py: add round() to preds:

ultralytics/utils/ops.py: modify scale_boxes function from:

to:

Regards

[glenn-jocher]

Thank you for sharing your solution. If you have any further questions or need assistance, feel free to ask.