New Release: Ultralytics v8.3.82

Discussion
,
1

:rocket: Announcing Ultralytics v8.3.82 Release!

We’re excited to announce the release of Ultralytics v8.3.82! This update brings significant improvements to ONNX model export, preprocessing pipelines, and hardware compatibility, along with enhanced dataset management. Dive into the details below to learn what’s new in this version! :star2:

:star2: Summary

Ultralytics v8.3.82 introduces fixes and enhancements to provide a smoother and more efficient workflow for developers, including improved ONNX export functionality, preprocessing accuracy, and broader hardware support. From addressing PyTorch incompatibilities to refining dataset configurations, this update is designed to elevate your productivity. :rocket:

:bar_chart: Key Changes

ONNX FP16 Export Fix

  • What’s New: A workaround (arange_patch) resolves PyTorch’s torch.arange incompatibility during ONNX model export with dynamic and half options.
  • Why It Matters: Enables seamless high-performance ONNX exports, especially useful for GPUs.

Improved ONNXRuntime Preprocessing

  • What’s New: Enhanced image resizing, aspect ratio handling, and padding for ONNXRuntime examples.
  • Why It Matters: Delivers consistent and reliable object detection across PyTorch and ONNXRuntime inferences.

Broader MNN Test Coverage

  • What’s New: MNN export testing added for Raspberry Pi.
  • Why It Matters: Expands compatibility to low-cost systems, supporting diverse hardware environments.

Streamlined Open Images V7 Dataset Configuration

  • What’s New: Centralized dataset directory handling in open-images-v7.yaml.
  • Why It Matters: Simplifies dataset setup for developers working with large datasets.

:dart: Purpose & Impact

  • Optimized ONNX Export: Solves FP16 precision export challenges in dynamic mode, enhancing GPU workflows and removing prior limitations.
  • Preprocessing Consistency: Fixes bugs to align preprocessing results with PyTorch outputs, ensuring robust results for object detection tasks.
  • Expanded Hardware Support: Ensures flexibility and reliability of MNN format exports across platforms like Raspberry Pi.
  • Effortless Dataset Management: Simplifies setup and management of Open Images V7, allowing you to focus on your models without configuration hassles.

These updates ensure optimized model export, preprocessing reliability, and better platform support for a wide range of machine learning tasks. :globe_with_meridians:

:wrench: What’s Changed

Here’s the complete list of changes introduced in v8.3.82:

:tada: New Contributors

A special shoutout to our new contributor:

We’re incredibly grateful for the amazing community that helps make these updates possible. :muscle:

Try It Out!

We encourage you to explore v8.3.82 and experience the improvements yourself! Full release notes can be found in the Ultralytics v8.3.82 Release Page.

For a detailed comparison of changes, take a look at the v8.3.81…v8.3.82 changelog.

Your feedback drives our progress—please let us know how this release works for you and share any suggestions or issues you encounter. As always, thank you for being part of the Ultralytics community! :bulb: