v8.3.82 - `ultralytics 8.3.82` ONNX `dynamic` and `half` export (#19464)

🌟 Summary

Ultralytics v8.3.82 enhances ONNX model export functionality, improves preprocessing for ONNXRuntime examples, and streamlines compatibility across various hardware setups. 🚀

📊 Key Changes

• ONNX FP16 Export Fix: Introduced a workaround (arange_patch) for exporting ONNX models with both dynamic and half options, avoiding incompatibilities in the PyTorch torch.arange function.
• Preprocessing Accuracy Boost: Improved image handling (aspect ratio, resizing, and padding) for ONNXRuntime examples to ensure more precise object detection.
• Broader MNN Test Coverage: Enabled MNN export testing on Raspberry Pi to ensure compatibility.
• Dataset Configuration Streamlining: Refined open-images-v7.yaml to centralize dataset directory management for clarity and maintainability.

🎯 Purpose & Impact

• ONNX Export Enhancements:

  • Purpose: Solve PyTorch incompatibility issues during ONNX model export with FP16 precision in dynamic mode.
  • Impact: Developers can now seamlessly export high-performance ONNX models while avoiding device-specific restrictions, improving workflows on GPUs.

• Accurate Preprocessing:

  • Purpose: Fix preprocessing bugs in ONNXRuntime inference to ensure image resizing, aspect ratio, and padding align with expected outputs.
  • Impact: Results in more consistent and reliable object detection performance between PyTorch and ONNX models.

• Compatibility Extension:

  • Purpose: Enable MNN export tests on Raspberry Pi to ensure cross-platform functionality.
  • Impact: Broader hardware compatibility for MNN format users, improving reliability on low-cost systems like Raspberry Pi.

• Dataset Management Improvements:

  • Purpose: Streamline configuration for the Open Images V7 dataset.
  • Impact: Easier and more consistent dataset setup, particularly useful for developers managing large datasets.

This update improves model export workflows, ensures consistent inference results across platforms, and expands compatibility for developers using diverse environments. 🌐

What's Changed

• Enable mnn in Raspberry Pi Tests by @lakshanthad in #19483
• Fix ONNX Example letterboxing by @quangdungluong in #19486
• Fix dataset_dir error with open-images-v7.yaml by @Y-T-G in #19475
• ultralytics 8.3.82 ONNX dynamic and half export by @Y-T-G in #19464

New Contributors

• @quangdungluong made their first contribution in #19486

Full Changelog: v8.3.81...v8.3.82

v8.3.81 - `ultralytics 8.3.81` Fix Metrics `on_plot` circular references (#19318)

🌟 Summary

The Ultralytics v8.3.81 release resolves a critical memory management issue in model validation while introducing key updates to documentation, testing workflows, and system reporting for enhanced usability and efficiency. 🚀✨

---

📊 Key Changes

• 🧹 Memory Leak Fix in Validation Modules:
  Resolved circular references in metrics (on_plot) across validators (DetectionValidator, PoseValidator, etc.) to avoid CPU memory leaks during repeated evaluations.
• 📚 Documentation Enhancements:
  Added examples for annotators in the library, and fixed/corrected metadata instructions for Triton guides and links in SAM 2 documentation.
• 🐧 Raspberry Pi CI Improvements:
  Re-enabled Raspberry Pi CI workflow with improved benchmarks and added cleanup for self-hosted CI runners.
• 🔍 Installation Diagnostics Update:
  Added project root installation path in the system environment summary for easier debugging.
• 📊 Better Table Sorting in Docs:
  Improved table sorting functionality for file sizes, numbers, and dot-separated values.

---

🎯 Purpose & Impact

• Memory Management Improvement:
  Eliminates residual memory accumulation when evaluating models repeatedly in scripts, preventing CPU overloading and Out-Of-Memory (OOM) errors. This fix improves the efficiency and reliability of long-running evaluation pipelines. 🧠🔥
• User-Friendly Documentation:
  New usage examples and enhanced instructions make it easier for developers to understand and implement key features. Correcting SAM 2 links ensures access to the right research resources. 📚💡
• Expanded Testing Coverage:
  Re-enabling Raspberry Pi CI and adding benchmarks enhances testing on diverse hardware, ensuring broader reliability across platforms. Improved CI cleanup reduces risks of resource conflicts. 🌐🔧
• Debugging Simplification:
  Including installation paths in diagnostics facilitates resolving Python environment-related issues for developers. 🔍🛠️
• Enhanced Sorting for Visual Data:
  Improved table sorting in docs boosts usability, especially when navigating large datasets or results with complex numbering or file size values. 📈✨

Overall, this update strengthens performance, stability, and developer experience, while prioritizing clarity and usability for the broader community. 🎉

What's Changed

• Update Raspberry Pi 5 CIs with 16GB model by @lakshanthad in #19306
• New: add SAM and SAM-2 notebook in docs by @RizwanMunawar in #19461
• Add installation path to yolo checks output by @Y-T-G in #19463
• Clarify metadata instructions in Triton docs by @Y-T-G in #19457
• Fix Docs table number sorting by @Y-T-G in #19455
• Add examples for plotting reference by @RizwanMunawar in #19439
• Update SAM 2 research paper links by @joshua-dean in #19465
• Clean up Raspberry Pi CI runners by @lakshanthad in #19478
• ultralytics 8.3.81 Fix Metrics on_plot circular references by @RemiPT in #19318

New Contributors

• @joshua-dean made their first contribution in #19465
• @RemiPT made their first contribution in #19318

Full Changelog: v8.3.80...v8.3.81

v8.3.80 - `ultralytics 8.3.80` Fix YOLO-NAS export (#19426)

🌟 Summary (single-line synopsis)

Version 8.3.80 features crucial upgrades to YOLO-NAS handling, improved configuration merging for smoother exports, enhanced documentation interactivity, and compatibility refinements for key frameworks. 🚀

---

📊 Key Changes (bullet points highlighting any major changes)

• 🔄 Model Export Enhancements: Fixed YOLO-NAS export by integrating default configs (DEFAULT_CFG_DICT) into model attributes for higher flexibility.
• 🧠 RBOX Regularization: Refined bounding box angle calculations for consistency in the OpenCV definition.
• 📋 Documentation Upgrades: Introduced sortable tables for performance data to improve clarity and usability across pages.
• 🔧 Framework Compatibility: Constrained OpenVINO dependency to version >=2024.0.0,<2025.0.0 and updated outdated function calls for compatibility.
• 🐳 Docker Workflow Fix: Removed deprecated numpy dependency to resolve CI errors and streamlined build conditions.

---

🎯 Purpose & Impact (bullet points explaining any benefits and potential impact to users)

• 🌟 Improved Export Reliability: The configuration merging for YOLO-NAS ensures a more robust process, minimizing errors during model deployment and export.
• 🧮 Better Prediction Accuracy: Regularizing rboxes improves detections by addressing angle inconsistencies in bounding boxes.
• 🖱️ Enhanced User Experience: Sortable tables in documentation allow users to interactively organize performance metrics, making comparisons easier.
• ✅ Smoother OpenVINO Exports: Version constraints and function updates enhance stability and prepare the platform for future releases without breaking changes.
• 🔄 Cleaner CI Pipelines: Removing deprecated dependencies ensures more reliable builds, benefitting developers contributing to the codebase.

This update underscores Ultralytics' focus on improving usability, compatibility, and export-related functionalities for smoother workflows and precise model performance. 🚀

What's Changed

• Optimize Sony IMX500 doc by @lakshanthad in #19421
• Constrain OpenVINO versions to >=2024.0.0,<2025.0.0 by @ambitious-octopus in #19122
• New Docs sortable tables by @Y-T-G in #19376
• Fixing tflite export CI error by @ambitious-octopus in #19422
• Fix rbox regularization by @Y-T-G in #19429
• ultralytics 8.3.80 Fix YOLO-NAS export by @Y-T-G in #19426

Full Changelog: v8.3.79...v8.3.80

v8.3.79 - `ultralytics 8.3.79` Fix shift in HSV augmentation (#19311)

🌟 Summary

The v8.3.79 release of the Ultralytics YOLO framework introduces crucial bug fixes, enhancements to performance, and documentation updates. The primary focus is on correcting HSV augmentation mechanics and refining various code components for better reliability and usability. 🛠️✨

---

📊 Key Changes

• HSV Augmentation Fix: Resolved incorrect hue, saturation, and value shifts during data augmentation, ensuring consistent color transformations. 🎨
• YOLO12 Benchmark Refinement: Fixed performance metrics for YOLO12 models in documentation, updating speed and mAP comparisons ensuring accurate representations. 📈
• Docker Streamlining: Removed redundant tensorrt-cu12 dependency and added environment checks for better CI validation and Docker compatibility. 🐳
• Comet Integration Fix: Corrected class mapping index misalignment in Comet logging to avoid errors and misinterpretations in visualized data. 📋
• Single-Class Model Consistency: Enforced single-class model output alignment by adjusting nc and names parameters during training. 🚀
• Documentation Updates: Improved accuracy in multiple areas, embedding a YouTube interactive guide for YOLO12, and updating outdated citation links for research credibility. 📚
• General Code Cleanup: Enhanced readability by adhering to PEP8 conventions (snake_case), simplifying code for maintainability. 🖋️

---

🎯 Purpose & Impact

• Improved Data Augmentation: Provides users with accurate HSV augmentations, leading to better model robustness and performance during training.
• Benchmark Accuracy: Ensures users receive correct performance metrics allowing better model selection and easier comparison across versions.
• Enhanced Docker Usability: Smoother image builds and reduced dependency issues result in a simplified development experience. Perfect for CI pipelines!
• Better Logging: Developers benefit from accurate class mappings in Comet analytics, avoiding confusion when examining predictions.
• Single-Class Models Alignment: Offers consistency in model behavior, improving usability for tasks like binary classification.
• Improved Documentation: Users gain easier access to tutorials, accurate technical references, and a better learning experience overall.
• Cleaner Codebase: Enhances long-term maintainability and aligns code with modern Python standards, beneficial for both contributors and end-users.

This release makes strides in improving accuracy, user experience, and usability for developers and researchers alike. 🚀🌟

What's Changed

• Fix YOLO12 cfg links by @glenn-jocher in #19332
• Fix mkdocs_github_authors.yaml author LexBarou by @glenn-jocher in #19334
• Fix keypoints data display in pose.md by @RizwanMunawar in #19337
• Fix YOLO12 performance table by @Laughing-q in #19333
• Fix benchmark.js by @glenn-jocher in #19351
• Fix TensorRT export issue by removing redundant package by @lakshanthad in #19343
• Add https://youtu.be/SMSJvjUG1ko to yolo12.md by @RizwanMunawar in #19358
• Update model_data.py authors by @glenn-jocher in #19365
• Assign variable dt before reference in exception block by @sjhpark in #19349
• Use chart.js@latest by @glenn-jocher in #19372
• YOLOv3 paper URL update by @glenn-jocher in #19378
• Add imx export tip by @RizwanMunawar in #19404
• Det models with nc=1 when single_cls=True by @kevinconka in #19381
• Fixed Comet integration to use class map aligned index when trying to get class name by @yaricom in #19408
• Simplify auto_annotate function by @Burhan-Q in #19400
• Updates imgToAnns variable to snake_case by @Burhan-Q in #19402
• Fix Grad strides do not match bucket view strides warning for YOLO12 DDP training by @Laughing-q in #19395
• Updates scaleFill argument to snake_case by @Burhan-Q in #19401
• Check environment in docker.yml by @glenn-jocher in #19418
• Disable Dockerfile tests temporarily by @glenn-jocher in #19420
• Fix incorrect metrics when nothing is detected by @Y-T-G in #19362
• ultralytics 8.3.79 Fix shift in HSV augmentation by @picsalex in #19311

New Contributors

• @picsalex made their first contribution in #19311
• @kevinconka made their first contribution in #19381
• @sjhpark made their first contribution in #19349

Full Changelog: v8.3.78...v8.3.79

v8.3.78 - `ultralytics 8.3.78` new YOLO12 models (#19325)

🌟 Summary

This release, v8.3.78, brings an exciting new model to the family: YOLO12 🚀, featuring an attention-centric design for superior accuracy and efficiency across a variety of computer vision tasks.

---

📊 Key Changes

• Introduction of YOLO12 Models:

  • New Architecture: Incorporates attention mechanisms like Area Attention, R-ELAN, and FlashAttention for optimized performance.
  • Comprehensive Task Support:
    • Object detection, segmentation, pose estimation, classification, and oriented bounding box (OBB) detection.
  • Benchmark Improvements:
    • Demonstrates higher mAP (mean Average Precision) and efficiency compared to YOLO10/YOLO11 and competitors like RT-DETR.

• Model-Specific Enhancements:

  • Introduced multiple YOLO12 variants (n, s, m, l, x) catering to different computing environments such as cloud systems and edge devices.
  • Added new task-focused configurations for image classification, pose estimation, and segmentation.

• Documentation Updates:

  • YOLO12 now included in detailed model documentation with performance metrics and usage examples.
  • Extensive references, including benchmarks for comparison with leading global detection models.

• Code Simplifications and Bug Fixes:

  • ONNX Run-Time Fixes: Improved device handling and tensor reshaping for ONNX users.
  • TFLite Export Cleanup: Removed redundant parameters for simpler TensorFlow Lite export logic.
  • Code Refinement: Enhanced readability and maintainability across inference and export pipelines.

---

🎯 Purpose & Impact

• Purpose:

  • YOLO12 brings a paradigm shift in accuracy and efficiency by adopting attention mechanisms tailored for real-time object detection.
  • Streamlines codebase for easier maintenance and integration in diverse projects.

• Impact:

  • Developers gain access to cutting-edge state-of-the-art models excelling in versatility, speed, and precision.
  • Tasks like multi-object detection, segmentation, and pose estimation become more accessible for smaller devices (e.g., edge devices).
  • Improved user experience with easier model selection, robust export support, and refined prediction outputs.

🔮 This update is not only a leap forward in technological advancement but also a commitment to making intelligent vision accessible to all.

What's Changed

• Remove unused parameters from export_tflite by @Y-T-G in #19319
• Fix ONNX RuntimeError with dynamic WorldModel by @Y-T-G in #19322
• Add YOLO12 model info by @Laughing-q in #19328
• Add https://youtu.be/BPYkGt3odNk to docs by @RizwanMunawar in #19331
• Refactor with simplifications by @glenn-jocher in #19329
• update model_data.py by @RizwanMunawar in #19330
• ultralytics 8.3.78 new YOLO12 models by @Laughing-q in #19325

Full Changelog: v8.3.77...v8.3.78

v8.3.77 - `ultralytics 8.3.77` faster YOLOv8-Segment ONNX Runtime example (#19312)

🌟 Summary

The YOLOv8 v8.3.77 release introduces a significant performance optimization for ONNX Runtime segmentation models, as well as essential compatibility enhancements and minor fixes. 🚀

---

📊 Key Changes

• 🚀 YOLOv8-Segment Optimization (ONNX Runtime)

  • Post-processing was drastically improved for both GPU (91.7% faster) and CPU (58.9% faster), resulting in a total inference speedup of up to 32.4%.

• 🔧 Optional thop Dependency Support

  • Made the thop library optional by handling its absence gracefully to avoid errors in environments lacking it.

• 📂 ONNX Export Improvements

  • Streamlined ONNX dynamic model export logic by revising type handling, ensuring better reliability and maintainability.

---

🎯 Purpose & Impact

• Faster and More Efficient Inference ⚡

  • The optimization of YOLOv8-Segment ONNX Runtime reduces latency significantly, improving user experiences in deployment scenarios where real-time performance is essential.

• Increased Compatibility with Minimal Setups 🤝

  • By making thop optional, the release ensures broader support for systems, including lightweight environments like Conda setups, enhancing user flexibility.

• Improved Developer Experience 🛠️

  • Simplified export logic and code maintainability ensures a more robust development and debugging process for ONNX users.

These updates collectively enhance usability, efficiency, and reliability for YOLOv8 users across diverse applications. 🎉

What's Changed

• Fix dynamic export with YOLO World by @Y-T-G in #19304
• Revert "Fix dynamic export with YOLO World" by @Laughing-q in #19308
• Allow missing thop package by @glenn-jocher in #19314
• ultralytics 8.3.77 faster YOLOv8-Segment ONNX Runtime example by @AdnanEkici in #19312

New Contributors

• @AdnanEkici made their first contribution in #19312

Full Changelog: v8.3.76...v8.3.77

v8.3.76 - `ultralytics 8.3.76` fix `dynamic` batch inference with NMS export (#19249)

🌟 Summary

The release of Ultralytics v8.3.76 introduces improved dynamic batch inference for ONNX models with NMS export, a better object tracking experience, and various code and documentation enhancements. 🚀

---

📊 Key Changes

• Dynamic Batch Improvements:

  • Solved issues with dynamic=True and nms=True where the batch size was fixed at export.
  • Enabled dynamic input handling by introducing padding for unmatched batch sizes during ONNX exports.

• Tracking Enhancements:

  • Addressed errors when processing Torch tensors in model.track().
  • Improved integration of original input images with trackers for accuracy.

• Performance Accuracy:

  • Fixed GPU memory conversion errors for logging VRAM usage to ensure accurate reporting.

• Documentation Updates:

  • Standardized formatting in documentation for better consistency.
  • Enhanced examples for interpreting prediction results across detection, pose, segmentation, and more.

• Other Code Refinements:

  • Addressed layer miscount in logging by ensuring layers with no parameters are also displayed.
  • Enhanced GitHub issue templates for clearer differentiation between bug reports and feature requests.

---

🎯 Purpose & Impact

• 🛠 Improved Model Deployment: Dynamic padding during export ensures robust handling of varying batch sizes while maintaining compatibility with ONNX workflows.
• 🎥 Better Tracking: Smoother operation for streamed data and enhanced consistency in object tracking results benefit both developers and end-users.
• 📋 Accurate Logging: Correct VRAM usage metrics improve debugging and resource optimization.
• 📚 Developer Friendliness: Updated docs and examples make it easier for users to understand and utilize prediction results effectively.
• 🚀 Efficiency Boost: Code tweaks and fixes culminate in faster, more accurate model handling without disruptions.

This release addresses several community-reported issues, focusing on operational accuracy and usability across export, tracking, and development workflows! 🙌

What's Changed

• Initialize model_name attribute by @LoveAndHope-dev in #19224
• Update results.boxes docs: boxes.id may be None by @shankangke in #19227
• Fix bytes to GB conversion for logged VRAM usage by @Y-T-G in #19254
• Fix memory conversion by @Y-T-G in #19256
• Use new "Issue Type" in templates by @Y-T-G in #19257
• Add https://youtu.be/im7xBCnPURg to docs by @RizwanMunawar in #19265
• Support CoreML NMS export for Segment, Pose and OBB by @Y-T-G in #19173
• Fix layer count; show layers with no params in detailed log by @Y-T-G in #19202
• Standardize str formatting in docs by @lakshanthad in #19276
• Add detailed usage with demos to reCamera doc by @lakshanthad in #19275
• Add examples showing how to use result for all tasks by @Y-T-G in #19282
• Update model_data.py by @RizwanMunawar in #19267
• Revert "Support CoreML NMS export for Segment, Pose and OBB" by @glenn-jocher in #19273
• Fix error with torch tensor input in model.track() by @Y-T-G in #19278
• ultralytics 8.3.76 fix dynamic batch inference with NMS export by @Y-T-G in #19249

New Contributors

• @LoveAndHope-dev made their first contribution in #19224

Full Changelog: v8.3.75...v8.3.76

v8.3.75 - `ultralytics 8.3.75` Comet update to new `comet_ml.start()` API (#19187)

🌟 Summary

The v8.3.75 release includes robust updates for improved model export compatibility, user experience, and error handling across platforms, alongside enhanced documentation and integration refinements. 🚀

---

📊 Key Changes

• Enhanced CometML Integration:

  • Transitioned to the new comet_ml.start() API for smoother experiment handling.
  • Deprecated the COMET_MODE variable, introducing COMET_START_ONLINE for consistency.

• Export Function Updates:

  • Protobuf Dependency: Added support for protobuf>=5 for TensorFlow and TFLite exports, resolving compatibility issues.
  • Edge TPU and TF.js: Addressed platform-specific limitations for ARM64 and Linux exports to prevent unsupported configuration errors.

• Documentation Improvements:

  • Updated SAM auto-annotation, YOLOv8, and export format descriptions for clarity.
  • Redesigned inference examples to use accessible publicly hosted image URLs.

• New CLI Solutions:

  • Introduced practical use cases, including object counting, workout monitoring, queue analysis, and browser-based inference with Streamlit.

• Benchmarking Added:

  • Include new comparative performance metrics for popular object detection models like Gold-YOLO, YOLO-NAS, RTDETRv3, etc.

• Windows-Specific Fix:

  • Resolved an async file write bug to improve caching reliability on Windows systems.

• Improved Timing Precision:

  • Switched to time.perf_counter() for latency measurements, ensuring greater precision during benchmarking.

---

🎯 Purpose & Impact

• Improved Experiment Tracking:

  • Seamless CometML integration provides better environment consistency and logging during training processes.

• Enhanced Export Reliability:

  • Future-proofs TensorFlow and TFLite workflows while providing early error detection for ARM64/Linux users.

• Streamlined User Experience:

  • Updated documentation and example consistency ensure clarity, especially for beginners, minimizing friction during model setup and usage.

• Greater Platform Support:

  • Addressed critical Windows and platform-specific export edge cases, enhancing cross-platform usability.

• Better Model Insights:

  • Added benchmarks empower users to make informed decisions about which object detection models to implement based on accuracy, speed, and computational cost.

This release focuses heavily on improving reliability, usability, and documentation quality while resolving critical bugs and adding more tools for diverse real-world applications.

What's Changed

• Auto-annotate and SAM docs update by @Y-T-G in #19156
• Switch to perf_counter() for latency measurement by @Y-T-G in #19177
• Force protobuf>=5 for SavedModel export by @Y-T-G in #19206
• Fix Docker QEMU issues while building JetPack 6 Dockerfile by @lakshanthad in #19216
• Fix bus.jpg path in predict.md by @RizwanMunawar in #19203
• Update NMS description on export-args.md by @Buligon in #19211
• Add NMS related args to export-table.md by @Y-T-G in #19215
• Add Docs models benchmark by @Laughing-q in #19176
• Tasks Docs updates by @LexBarou in #19181
• YOLOv8 Docs updates by @LexBarou in #19182
• Add Solutions CLI usage in quickstart.md by @RizwanMunawar in #19160
• Fix edgetpu and tfjs exports for arm64 Linux by @lakshanthad in #19154
• Fix windows async np.save bug by @eric80739 in #19218
• Fix print() for ConfusionMatrix for Classify task by @Y-T-G in #19169
• Fix updating of best epoch during early stopping by @vfcosta in #19164
• ultralytics 8.3.75 Comet update to new comet_ml.start() API by @yaricom in #19187

New Contributors

• @vfcosta made their first contribution in #19164
• @eric80739 made their first contribution in #19218

Full Changelog: v8.3.74...v8.3.75

v8.3.74 - `ultralytics 8.3.74` Fix Ray Tune callback error (#19144)

🌟 Summary

Ultralytics v8.3.74 introduces updates to improve compatibility with modern tools (like Ray Tune), smooth errors, and enhance deterministic training and export flexibility. 🛠✨ Simplified workflows for developers with better usability.

---

📊 Key Changes

• 🔧 Fixed Ray Tune Callback Issues: Replaced deprecated ray.tune.is_session_enabled() with ray.train._internal.session.get_session() ensuring compatibility with latest Ray versions.
• 🛠 Enhanced Deterministic Training Control: Added unset_deterministic() to handle environment changes, and prevent unnecessary CUDA warnings.
• 🖼 PIL Image Support in plot(): Allowed direct return of PIL images with annotator.im, improving compatibility with PIL workflows.
• 🚀 Improved Export Flexibility: Adjusted model.export() to take a data parameter while simplifying predict() calls.
• 🐳 Optimized Docker Workflow: Improved Docker token authentication and switched to docker build for better stability and security.
• ✅ Streamlined Benchmarking: Cleaned dataset and metric assignments in benchmarking to avoid redundancy and improve reliability.

---

🎯 Purpose & Impact

• 🚀 Greater Compatibility: Seamless integration with the latest versions of Ray ensures that errors linked to deprecated methods are resolved.
• ⚡ Workflow Flexibility: Managing deterministic settings dynamically boosts training adaptability while cleaning up workflow logs.
• 📸 Visualization Improvements: Returning PIL images directly simplifies further processing in pipelines dependent on image outputs.
• 🛠️ User-Friendly Model Exports: Configurable export makes model usage and testing more straightforward for developers.
• 🔒 Stronger Security: Docker workflow improvements enhance authentication security, benefitting advanced build setups.
• ✅ Clarity in Development: Benchmark logic cleanup minimizes confusion and potential errors, improving developer experience.

This version is packed with incremental improvements, making model training, testing, and deployment smoother and more user-friendly while preparing Ultralytics for the future. 🎉

What's Changed

• Fix docker.yml by @glenn-jocher in #19125
• Fix missing data warning and undefined variables by @Y-T-G in #19124
• Fix missing data.yaml error on int8 export by @Y-T-G in #19128
• Return PIL image if pil=True by @Y-T-G in #19146
• Unset CUBLAS_WORKSPACE_CONFIG for non-deterministic training and inference by @Y-T-G in #19138
• ultralytics 8.3.74 Fix Ray Tune callback error by @Y-T-G in #19144

Full Changelog: v8.3.73...v8.3.74

v8.3.73 - `ultralytics 8.3.73` GitHub Container Registry Images at `ghcr.io` (#19114)

🌟 Summary

The Ultralytics v8.3.73 release focuses on enhancing containerization workflows, updating library dependencies, improving documentation, and refining the development process. 🚀

---

📊 Key Changes

• Containerization Enhancements:
  • Added support for publishing Docker images to both GitHub Container Registry (GHCR) and Docker Hub with rich metadata for better usability. 🐋
  • Removed Ubuntu 24.04 ARM in CI workflows for streamlined testing.
• Dependency and Platform Updates:
  • Updated NVIDIA Jetson support with PyTorch 2.2.0 and Torchvision 0.17.2 for improved compatibility and performance. 🤖
  • Removed the beautifulsoup4 dependency, cleaning up the development environment. 🧹
• Code Refactoring:
  • Improved SQL result export by simplifying insertion logic and fixing potential issues with empty results.
  • Enhanced type hinting for better code clarity.
• Documentation and Tutorial Updates:
  • Added an embedded YouTube tutorial on Package Segmentation in the documentation. 🎥✨

---

🎯 Purpose & Impact

• Containerization Accessibility:
  • Publishing images to both Docker Hub and GHCR ensures users have multiple options for pulling images, increasing global availability and reducing friction. 🌍
  • The inclusion of detailed metadata in Docker images improves clarity for end-users.
• Better Hardware and Development Support:
  • NVIDIA Jetson users benefit from newer library versions for seamless deployment and better model performance.
  • Leaner development environments reduce installation times and maintenance burdens.
• Improved Learning Resources:
  • The YouTube tutorial enriches the documentation and aids new and existing users in understanding package segmentation workflows visually. 📚👩‍💻

---

TL;DR: This version updates Docker container workflows, improves NVIDIA Jetson compatibility, cleans up dev dependencies, and enhances user education through new video tutorials. 🚀💡

What's Changed

• Remove beautifulsoup4<=4.12.3 pin by @Laughing-q in #19103
• Update JetPack 5 torch and torchvision packages by @lakshanthad in #19098
• Minor Results.to_sql cleanup by @Laughing-q in #19081
• Add https://youtu.be/im7xBCnPURg to docs by @RizwanMunawar in #19115
• ultralytics 8.3.73 GitHub Container Registry Images at ghcr.io by @glenn-jocher in #19114

Full Changelog: v8.3.72...v8.3.73