New Release: Ultralytics v8.2.65

Discussion
,
1

:rocket: Announcing Ultralytics v8.2.65 Release!

Hello Ultralytics Community!

We are thrilled to announce the release of Ultralytics v8.2.65! This update brings a host of new features, improvements, and bug fixes designed to enhance your experience and the performance of your models. Here’s a quick overview of what’s new:

:star2: Key Features and Improvements

  • New Activation Module: We’ve introduced the AGLU activation function within the neural network modules. This new function allows for more sophisticated activation handling, potentially boosting model performance on specific tasks. Check out the PR #123 by @glenn-jocher.

  • Update to cache_ram Feature: The cache_ram feature has been disabled due to a known memory leak issue. This change ensures smoother classifier training sessions by preventing potential training disruptions. More details can be found in PR #124 by @RizwanMunawar.

  • Dependency Update for pytubefix: We now require version 6.5.2 or later of pytubefix. This update enhances compatibility and fixes bugs related to YouTube video processing. See PR #125 by @AyushExel.

  • NMS (Non-Maximum Suppression) Logic Update: Improvements have been made to the handling of single class and agnostic NMS, resulting in more accurate detections. For more information, refer to PR #126 by @glenn-jocher.

  • Documentation Enhancement: We’ve added detailed documentation for activation functions, making it easier for both new and existing users to understand and utilize the tools available for model customization. Check out the updates in PR #127 by @RizwanMunawar.

:dart: Purpose & Impact

  • Enhanced Flexibility in NN Customization: The AGLU activation function provides more options for activation handling, which can improve model performance on specific tasks.
  • Operational Stability: Disabling cache_ram prevents memory leaks, ensuring smoother training sessions.
  • Improved Video Handling: The pytubefix update enhances compatibility and fixes bugs related to YouTube video processing.
  • Better NMS Accuracy: The updated NMS logic results in more accurate detections by properly handling edge cases.
  • User-Friendly Documentation: New guides and references help users understand and utilize neural network modules more effectively.

We encourage you to try out the new release and share your feedback with us. Your insights are invaluable in helping us improve Ultralytics for everyone.

For more details, visit the release page.

Happy coding!

The Ultralytics Team

2

New AGLU activation seems interesting, but also profiled slower also on my system that SiLU (both on CPU and TRT10 engine exported speeds).

Anyone have any good results using it?