New Release: Ultralytics v8.3.44

Discussion
, ,
1

:rocket: Ultralytics v8.3.44 Release Notes

:glowing_star: Summary

We are excited to announce the release of v8.3.44. This update focuses on enhancing Triton Inference Server integration with improved metadata support, optimizing model export functionality, and delivering significant usability and stability enhancements to streamline your workflows. :rocket:

:bar_chart: Key Changes

Triton Inference Enhancements

  • :magnifying_glass_tilted_left: Added functionality to retrieve and store model metadata during export (on_export_end callback).
  • :card_index_dividers: Triton Model Repository now dynamically includes metadata (config.pbtxt) for seamless configurations.
  • :rocket: Enhanced TritonRemoteModel to support enriched customizability and traceability through metadata handling.
  • :white_check_mark: Set a default task (task=detect) for Triton Server models to assist users and minimise configuration errors.

Core Features & Fixes

  • Dependency Adjustment: Switched dependency back to lap for better stability and compatibility.
  • :hammer_and_wrench: Dynamic ONNX: Adjusted the dynamic setting to only trigger based on image size (not batch size) for robust and efficient ONNX model handling.
  • :light_bulb: Custom PyTorch Model Acceptance: Expanded AutoBackend to accept in-memory PyTorch models, improving workflow flexibility.
  • :high_voltage: AMP Compatibility: Hardcoded failing GPUs (e.g., GTX 16 & Quadro T series) to prevent NaN losses during AMP training.
  • :sparkles: New Utility - empty_like: Added streamlined tensor creation to enhance code reuse and maintainability.
  • :puzzle_piece: Segment Resampling Fix: Resolved segment-affecting issues to retain original points during resampling.

:bullseye: Purpose & Impact

:light_bulb: Enriched Triton Support

  • Improved metadata handling for smooth export and flexible deployment with Triton Inference Server.
  • Reduces user errors by introducing intelligent default task settings for Triton integrations.

:rocket: User Enhancements

  • In-memory PyTorch model acceptance simplifies integration in unique workflows.
  • Eliminates dependency issues for compatibility-sensitive workflows by reverting to lap.

:flexed_biceps: Stability & Performance Gains

  • Addressed AMP-related issues for specific GPUs, ensuring stable training experiences.
  • Dynamic ONNX tweaks make model exports more consistent and reliable.

:wrench: Maintenance & Optimization

  • empty_like utility reduces redundancy in tensor operations across the codebase.
  • Resampled segments now maintain fidelity to original data for enhanced accuracy.

This release represents a step forward in usability, stability, and technical sophistication for Ultralytics users—reducing friction across Triton, model export, and core functionality! :glowing_star:

:magnifying_glass_tilted_left: What’s Changed

:tada: New Contributors

Explore the Full Changelog: Detail View

:speech_balloon: Get Involved!

We encourage you to try out v8.3.44 and share your feedback in our Discussions. Your input helps improve Ultralytics and its offerings for the entire community. :rocket:

Thank you to all contributors and users for pushing the boundaries of innovation together! :sparkles: