ultralytics 8.3.72

v8.3.72 - `ultralytics 8.3.72` Fix NVIDIA Jetson DLA core support for DLA inference (#19078)

on Python PyPI

🌟 Summary

The v8.3.72 release focuses on enhancing NVIDIA Jetson DLA (Deep Learning Accelerator) core compatibility for inference, improving export documentation, and resolving minor inefficiencies and errors for broader usability and smoother performance. 🚀

📊 Key Changes

• Enhanced NVIDIA Jetson DLA Support:
  • Introduced explicit control of DLA core selection (dla:0/dla:1) during TensorRT export and inference.
  • Added detailed documentation of NVIDIA Jetson DLA device specifications (core count, frequency, etc.).
  • Fixed metadata handling for DLA-specific inference settings.
• Export Documentation Overhaul:
  • Added detailed argument tables for all model export formats (e.g., ONNX, TensorRT, CoreML), improving clarity on custom export options such as half-precision (FP16), INT8 quantization, and dynamic input sizes.
• Optimized seg_bbox Rendering:
  • Refined label handling logic in the plotting utility, reducing unnecessary operations if a label is absent, slightly improving performance.
• Bug Fixes:
  • Resolved an issue with missing nc attributes during NMS export, improving reliability in multi-GPU or custom training setups.
• Documentation Updates:
  • Enhanced Crack Segmentation Dataset resources with direct Colab integration, a tutorial notebook, and a demo video for easier onboarding.

🎯 Purpose & Impact

• Improved Compatibility: The NVIDIA Jetson DLA improvements ensure that edge devices benefit from seamless inference setups, enabling accelerated performance with reduced bottlenecks. Ideal for IoT and edge AI devices. 🖥️✨
• Simplified Export Process: The new export argument tables demystify complex configurations, empowering users to adapt models for their specific hardware or workflows more easily. 📦🔧
• Performance Benefits: Minor optimizations ensure faster runtime efficiency, especially for visualization and plotting tasks where unnecessary computations are avoided. ⚡
• Enhanced Reliability: Fixes like handling missing nc attributes and metadata improve model robustness, particularly in advanced user scenarios (e.g., multi-GPU setups, custom models). ✅
• Streamlined Learning Experience: The improved Crack Segmentation training resources lower the barrier to entry for researchers in infrastructure and transportation safety fields. 🛠️🚗

This release represents a strong push for enhanced edge device support, better export usability, and overall reliability improvements while empowering both beginners and advanced users. 🎉

What's Changed

• Optimize seg_bbox calculations by @RizwanMunawar in #19056
• Resolve warnings by @glenn-jocher in #19073
• Add https://youtu.be/C4mc40YKm-g and notebook badge in docs by @RizwanMunawar in #19086
• Add Export Arguments tables to all Export docs by @lakshanthad in #18952
• Fix missing nc attribute error on NMS export by @Y-T-G in #19083
• Replace beautifulsoup4 pin with mkdocs-ultralytics-plugin>=0.1.17 by @Laughing-q in #19085
• ultralytics 8.3.72 Fix NVIDIA Jetson DLA core support for DLA inference by @Laughing-q in #19078

Full Changelog: v8.3.71...v8.3.72