New Release: Ultralytics v8.3.58

Discussion
,
1

:rocket: Ultralytics v8.3.58 Release Announcement

We’re excited to announce the release of Ultralytics v8.3.58, packed with updates to help you optimize performance, streamline workflows, and leverage new features for improved usability. With updates in model benchmarking, documentation, and training, this release is for developers and users looking to refine and accelerate their AI implementations. :tada:

:star2: Summary

The v8.3.58 release brings:

  • TensorRT benchmarking improvements for classification models.
  • New training options for dynamic image sizes.
  • Documentation upgrades with embedded videos and refined integration guides.
  • Optimized Docker builds for faster and leaner environments.

Let’s dive in!

:bar_chart: Key Changes

TensorRT Model Benchmarking Update

Benchmarking for TensorRT models now uses uint8 (integer) input data instead of float32 (decimal) for classification tasks. This change ensures alignment with real-world input formats, providing more accurate performance metrics for TensorRT users. :racing_car:

Documentation Enhancements

  1. Embedded Video Tutorials:

    • Videos added in the object counting and model exporting guides improve clarity and learning. :movie_camera:
    • Check them out in the documentation!

  2. YOLO11 Integration Updates:

    • References to YOLOv8 replaced to reflect the latest YOLO11 architecture, ensuring accuracy in tutorials and usage guides.

  3. Multi-Scale Training Argument:

    • Added in the docs for dynamic training resolutions. This feature helps train models across varying image sizes to improve adaptability and performance. :rainbow:

Docker Optimization

A .dockerignore file has been introduced to reduce build size by excluding unnecessary files. This streamlines Docker image builds, improving both build times and security. :whale2:

:dart: Purpose & Impact

Purpose:

  • Optimize TensorRT benchmarking for realistic input data.
  • Enhance user experience with better instructional clarity and precise documentation.
  • Enable new training flexibility with dynamic image size options.
  • Secure and streamline Docker deployments for efficiency.

Impact:

  • Faster and more accurate benchmarks for TensorRT users.
  • A smoother learning curve for both beginners and advanced users with improved documentation.
  • Flexible training with multi-scale options to handle diverse datasets effectively.
  • Leaner Docker images for clean and secure deployments.

:fire: What’s Changed

Here’s a breakdown of the key changes included in this release:

  • Instructional Video Addition: Added a YouTube tutorial to the docs by @RizwanMunawar in #18507.
  • TensorRT Guide Update: Replaced YOLOv8 references with YOLO11 in the TensorRT integration guide by @RizwanMunawar in #18513.
  • New Training Argument: Added the multi_scale argument for training models on dynamic image sizes by @Y-T-G in #18531.
  • Docker Improvement: Introduced .dockerignore to reduce Docker image clutter by @glenn-jocher in #18534.
  • TensorRT Profiling Update: Updated TensorRT classification profiling to use uint8 input by @Laughing-q in #18327.

:memo: Full Changelog

For a detailed overview of all code changes, please visit the full changelog.

:link: Release URL: v8.3.58

We encourage you to explore this new release, test the updates, and provide feedback to help us shape future improvements! Your insights are invaluable to the YOLO community and the Ultralytics team. :bulb:

Happy experimenting, and thank you for being part of the YOLO journey! :star2: