Detached inputs to the one2one head for Yolo v10

[matpalm]

Was surprised to see that the input to the one2one is detached.
https://github.com/ultralytics/ultralytics/blob/main/ultralytics/nn/modules/head.py#L74-L77

Am assuming that is due to instability with the conflicting gradients w.r.t. the one2many head ?
( Can't see anything in the original paper about it... )

Cheers, Mat

[UltralyticsAssistant]

👋 Hello @matpalm, thank you for your engagement with Ultralytics 🚀!

For addressing such queries, you might find the relevant sections in our Docs particularly useful. They cover aspects of model architecture and implementation that might help clarify your questions.

If you're reporting a potential issue or bug, please provide a minimum reproducible example to aid our debugging process 🐛.

For further insights into model intricacies or training inquiries, do share as much context as possible, including snippets of your implementation and any logs if relevant. Additionally, ensure you're following our Tips for Best Training Results.

While this is an automated response to guide you proactively, rest assured that an Ultralytics engineer will review your query shortly and provide more specific assistance.

Feel free to tap into our community for broader discussions: join us on Discord for real-time conversations, explore our Discourse for in-depth threads, or catch up with peers on our Subreddit. 🎧

Upgrade

Ensure you're using the latest version of the ultralytics package and all its requirements in a Python>=3.8 environment. This helps confirm whether your issue is already addressed in the newest release:

pip install -U ultralytics

Environments

YOLOv8 and beyond can be operated within these verified, up-to-date environments:

• Notebooks with free GPU:
• Google Cloud Deep Learning VM. See GCP Quickstart Guide
• Amazon Deep Learning AMI. See AWS Quickstart Guide
• Docker Image. See Docker Quickstart Guide

Status

When this badge is green, all Ultralytics CI tests are passing. CI tests cover the functionality of YOLOv8 Modes and Tasks across different OS environments every 24 hours and on each commit.

[glenn-jocher]

@matpalm hi Mat, the detachment in the one2one head is indeed to address gradient instability issues. This approach helps manage conflicting gradients between the one2one and one2many heads. If you have further questions, feel free to ask!