New Release: Ultralytics v8.3.75

:star2: Announcing Ultralytics v8.3.75: Reliable Improvements at Scale!


:rocket: Summary

We’re thrilled to introduce version v8.3.75, a feature-packed release that enhances model export compatibility, platform reliability, and user experience while refining documentation and adding practical new solutions. Dive into the details below, and don’t hesitate to try it out and share your thoughts with us!


:bar_chart: Key Changes

Enhanced CometML Integration

  • Transitioned to the new comet_ml.start() API for more reliable and streamlined experiment tracking.
  • Deprecated the COMET_MODE variable, introducing COMET_START_ONLINE for better consistency and usability.
    Contributed by @yaricom

Export Function Updates

  • Protobuf Dependency: Enforced protobuf>=5 for TensorFlow and TFLite exports, addressing prior compatibility issues.
    Contributed by @Y-T-G
  • Edge TPU and TF.js: Addressed Linux (ARM64) export issues to detect unsupported configurations upfront.
    Contributed by @lakshanthad

Documentation Enhancements

New CLI Solutions

  • Added practical use cases such as object counting, workout monitoring, queue analysis, and browser-based inference using Streamlit.
    Contributed by @RizwanMunawar

Benchmarking Models

  • Introduced performance metrics for selecting object detection models like Gold-YOLO, YOLO-NAS, and RTDETRv3.
    Contributed by @Laughing-q

Windows-Specific Fix

Improved Timing Precision

  • Adopted time.perf_counter() for more accurate latency measurements during training and benchmarking.
    Contributed by @Y-T-G

:dart: Purpose & Impact

  • Streamlined Integrations: Improved CometML tracking for smoother experiments and logging.
  • Reliable Exports: Ensuring TensorFlow and TFLite workflows are future-proof and platform-safe.
  • Optimized Usability: Enhanced CLI solutions and improved documentation simplify onboarding for all users.
  • Cross-Platform Support: Bug fixes ensure consistency, whether you’re on Windows, Linux, or ARM64.
  • Informed Decisions: New benchmarking metrics empower users to identify the best model for their needs without guesswork.

:star2: Contributions

A huge thanks to everyone who contributed to this release!

  • New Contributors:

    • @vfcosta resolved the issue of best epoch tracking during early stopping. (PR)
    • @eric80739 fixed async file writing on Windows. (PR)
  • Highlighted PRs:

    • Documentation and export updates by @Y-T-G, @RizwanMunawar, @Buligon, and @LexBarou.
    • Docker QEMU fixes for JetPack by @lakshanthad (PR).

You can learn more about the changes in the full changelog.


:link: Try It Out

  • Installation/Update: Simply use pip install ultralytics --upgrade to get started.
  • Feedback: Have suggestions or encountered an issue? Join the discussion or open an issue right here.

We’re eager to hear how these updates improve your workflows and projects! Together, let’s continue pushing the boundaries of computer vision. :rocket: