Ultralytics

🌟 Summary
The **v8.3.70** release brings feature enhancements with improved export functionalities, updated compatibility for PyTorch, and usability enhancements in benchmarking and documentation. 🚀

---

📊 Key Changes
- **Sony IMX500 Export Update**: Added support for the `data` argument, enabling dataset configuration during export for better control over quantization in formats like OpenVINO, TensorRT, and TF Lite. 📁
- **Torch 2.6 Compatibility**: Updated Torch-Torchvision mappings to ensure seamless functionality with the latest PyTorch update. 🔧
- **Format-Specific Benchmarking**: Introduced benchmarking support for individual formats (e.g., ONNX) to allow targeted performance evaluations. 📊
- **NVIDIA DLA Support**: Implemented support for running models on specific NVIDIA DLA cores—a key feature for specialized hardware optimization. 🖥️
- **Improved `numpy` Stability**: Pinned `numpy` version to prevent compatibility issues with OpenVINO and TFLite during CI tests. ✅
- **Documentation Enhancements**: Added tutorial videos and refined sections for clarity, aiding users and contributors. 📚

---

🎯 Purpose & Impact

- **Improved Export Workflows**:
- **Purpose**: The `data` argument helps users customize exports with specific dataset configurations, simplifying quantization and compatibility for edge and on-premise deployment.
- **Impact**: Makes exports more robust and adaptable to diverse workflows, ensuring higher-quality models with optimized performance.

- **Torch Compatibility**:
- **Purpose**: Keep the framework current with the latest PyTorch improvements.
- **Impact**: Allows users to leverage PyTorch 2.6's advancements without compatibility hiccups, maintaining a seamless experience.

- **More Granular Benchmarking**:
- **Purpose**: Enable granular analysis of models' efficiency in specific formats like ONNX or TensorFlow Lite.
- **Impact**: Helps developers fine-tune models for scenarios where particular formats are essential for deployment.

- **DLA Optimization**:
- **Purpose**: Ensure efficient inference on NVIDIA's specialized hardware.
- **Impact**: Reduces computational overhead and maximizes performance for users running models on NVIDIA DLA platforms.

- **CI Stability with `numpy`**:
- **Purpose**: Prevent runtime or testing errors due to incompatible `numpy` versions.
- **Impact**: Ensures reliable and predictable performance for developers and CI pipelines.

- **Accessible Documentation**:
- **Purpose**: Make it easier for new contributors and users to onboard through visual and detailed guides.
- **Impact**: Encourages community growth and simplifies the learning curve for both model and framework regulars.

---

🎉 This release is packed with features to empower smoother workflows, improve hardware compatibility, and promote user-friendly innovation! 🌟

What's Changed
* Update torchvision compatibility table for `torch 2.6` by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18935
* Add https://youtu.be/yMR7BgwHQ3g to docs by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/18936
* Add support for single `export` format `benchmark` by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/18740
* Remove YOLOv10 benchmarks from CI by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18937
* mkdocs-ultralytics-plugin>=0.1.16 by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18942
* Pin `numpy<=2.1.1` to resolve failing --slow CI by lakshanthad in https://github.com/ultralytics/ultralytics/pull/18943
* Add DLA specific core usage by AbelHaro in https://github.com/ultralytics/ultralytics/pull/18930
* Minor Docs edits by LexBarou in https://github.com/ultralytics/ultralytics/pull/18940
* Eliminate `numpy<2.0.0` pin for OpenVINO on macOS by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18945
* `ultralytics 8.3.70` add `data` argument to Sony IMX500 export by lakshanthad in https://github.com/ultralytics/ultralytics/pull/18852

New Contributors
* LexBarou made their first contribution in https://github.com/ultralytics/ultralytics/pull/18940
* AbelHaro made their first contribution in https://github.com/ultralytics/ultralytics/pull/18930

**Full Changelog**: https://github.com/ultralytics/ultralytics/compare/v8.3.69...v8.3.70

🌟 Summary
The Ultralytics `v8.3.69` release introduces enhanced integration for data export, including a new `to_sql()` method for saving model results directly into an SQL database. This version also continues refining the documentation, stability, and benchmarking experience to provide a smoother user workflow. 🚀

---

📊 Key Changes
- **New SQL Export Capability:** Users can now use the `to_sql()` method to store YOLO model inference results directly in an SQL database for organization and analysis. 🗄️
- **Generalized Export Options:** Expanded export methods for results, adding `to_df`, `to_csv`, `to_xml`, and `to_json` for improved compatibility with different formats.
- **Improved Documentation:**
- Added dynamic performance visualization charts to model documentation for more engaging and intuitive comparisons. 📈
- Simplified and clarified YOLOv3 documentation tables for better readability. 📚
- **Benchmark Enhancements:**
- Strengthened validation for input sizes, ensuring square images are required for benchmarking. 🖼️
- Modified logging to lessen verbosity and improve user-friendliness during prediction and validation tasks. 💡
- **Fixes and Stability:**
- Corrected edge cases in `AutoBatch` with better `RT-DETR` compatibility. ✅
- Implemented model deep copy for profiling tasks to ensure unmodified behavior during GFLOP measurements. 🔒
- **CI Pipeline Adjustments:**
- Temporarily disabled Windows CI and Raspberry Pi CI workflows for maintenance, ensuring smoother ongoing operations. 🛠️

---

🎯 Purpose & Impact
- **Purpose:**
- The `to_sql()` function provides seamless integration with relational databases, making it easier to organize, query, and analyze results within existing workflows.
- Enhanced export flexibility supports various use cases and workflows, from technical development to high-level reporting.
- Improvements in benchmarking and documentation provide clarity for researchers and developers determining model performance and deployment strategies.
- **Impact:**
- **For Developers:** Effortlessly manage results with SQL integration, while enjoying a more streamlined benchmarking process.
- **For Researchers:** Leverage clearer documentation and performance visualizations for easier evaluation of model trade-offs.
- **For General Users:** Reduced complexity and improved tools make interacting with the platform more intuitive and accessible. 🌟

This release continues to strengthen both backend functionality and user experience, paving the way for effective use of YOLO and supporting tools across diverse projects! 🎉

What's Changed
* Fix YOLOv3 table by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18902
* Add Docs models JS charts by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18905
* Simplify build_docs.py by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18910
* Fix `AutoBatch` when working with RT-DETR models by Laughing-q in https://github.com/ultralytics/ultralytics/pull/18912
* Add `PP-YOLOE+` params and flops data by Laughing-q in https://github.com/ultralytics/ultralytics/pull/18911
* Temporarily disable Raspberry Pi CI due to maintenance by lakshanthad in https://github.com/ultralytics/ultralytics/pull/18923
* Fix Docs edit button links by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18932
* Add imgsz check and improve logs for benchmarks by Y-T-G in https://github.com/ultralytics/ultralytics/pull/18917
* `ultralytics 8.3.69` New Results `to_sql()` method for SQL format by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/18921


**Full Changelog**: https://github.com/ultralytics/ultralytics/compare/v8.3.68...v8.3.69

🌟 Summary
This release (`v8.3.68`) delivers meticulous updates enhancing benchmarking workflows, export processes, documentation clarity, and model comparison tools for improved usability and precision. 🚀✨

---

📊 Key Changes
- **Benchmarking Model Path Fix**: Corrected model path handling in benchmarking to prioritize `pt_path`, fallback to `ckpt_path`, and then `model_name` for file identification. Improved log clarity.
- **EfficientDet Integration**: Added EfficientDet (d0-d3) benchmarking stats for performance evaluation with other models.
- **Enhanced Visualization**: Streamlined chart rendering for benchmarks, including refined dataset logic and active model configurations via page settings.
- **Export Adjustments**: Fixed issues with ONNX dynamic export, OpenVINO int8, and TFLite at edge cases (`imgsz=32`). Improved handling of classification models and adjusted NMS logic.
- **Documentation Updates**: Improved AzureML Python version recommendations and introduced a fallback mechanism for file minification during documentation builds.

---

🎯 Purpose & Impact
- 📋 **Clarity & Consistency**: Benchmarking logs now show clearer and more intuitive references to simplify debugging and analysis.
- 🚀 **Improved Model Evaluation**: Adding EfficientDet and chart enhancements helps users make better decisions when comparing models.
- ⚙️ **Robust Edge Case Handling**: Fixes to TFLite, ONNX, and OpenVINO exports safeguard against errors, particularly with smaller image sizes or specific benchmarks.
- 🧪 **Improved Testing & Usability**: Adjustments in export configuration reduce runtime errors during testing.
- 📝 **Developer-Friendly Documentation**: Clarified setup instructions in AzureML and optimized minification improve user experience, especially for new developers.

This release focuses on **greater flexibility, reliability, and usability** for users managing benchmarking, exporting, and evaluating models! 🌟

What's Changed
* Simplify chart legend by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18878
* Add EfficientDet to model comparisons by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18884
* Add Javascript active models argument by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18886
* Minify fallback on docs build by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18887
* Fix benchmark.js by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18890
* Fix export test matrices to exclude `nms` from Classify models by Y-T-G in https://github.com/ultralytics/ultralytics/pull/18880
* Fix Python Version Recommendation in AzureML Quickstart by Lucashygi in https://github.com/ultralytics/ultralytics/pull/18889
* Fix TFLite error and OpenVINO int8 error at imgsz=32 by Y-T-G in https://github.com/ultralytics/ultralytics/pull/18898
* `ultralytics 8.3.68` Benchmarking model path fix by Y-T-G in https://github.com/ultralytics/ultralytics/pull/18894

New Contributors
* Lucashygi made their first contribution in https://github.com/ultralytics/ultralytics/pull/18889

**Full Changelog**: https://github.com/ultralytics/ultralytics/compare/v8.3.67...v8.3.68

📊 Key Changes
- **Added NMS support** during export for multiple formats: ONNX, TensorRT, TFLite, TFJS, SavedModel, OpenVINO, and TorchScript 🧩.
- Enabled export-specific NMS for `detect`, `segment`, `pose`, and `obb` tasks with enhanced options like `nms=True`.
- Expanded NMS-related functionality in models and exporters, including support for more complex configurations like agnostic or rotated NMS.
- Streamlined model APIs to support embedded NMS using an updated `NMSModel` wrapper.

🎯 Purpose & Impact
- **Purpose**:
- Simplifies deployment pipelines by embedding NMS directly into exported models, removing the need for custom post-processing 🔗.
- Enhances usability across deployment platforms (e.g., TensorFlow, ONNX, OpenVINO) by integrating NMS into the export pipeline.
- **Impact**:
- Significantly improves portability and ease of deployment for real-time applications 🎯.
- Makes YOLO models more accessible for hardware-accelerated environments like TensorRT and Edge TPU 🚀.
- Reduces errors and complexity in downstream pipelines by unifying pre/post-processing across tasks.

Overall, this update empowers developers to deploy YOLO models with integrated NMS across a wide variety of frameworks and platforms, making the process **faster**, **more robust**, and **less error-prone**. 🌟

What's Changed
* Update HUB Shared Inference API limits by sergiuwaxmann in https://github.com/ultralytics/ultralytics/pull/18850
* Add `YOLO_TQDM_RICH` environment variable by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18854
* `ultralytics 8.3.67` NMS Export for Detect, Segment, Pose and OBB YOLO models by Y-T-G in https://github.com/ultralytics/ultralytics/pull/18484


**Full Changelog**: https://github.com/ultralytics/ultralytics/compare/v8.3.66...v8.3.67

🌟 Summary
The **v8.3.66** release introduces support for Rockchip RKNN export, enhances hardware compatibility, refines documentation, and fixes several bugs, marking a significant step for developers working on edge AI and cross-platform deployments.

---

📊 Key Changes
- ✨ **Rockchip RKNN Support**: Added the ability to export YOLO models to the RKNN format for deployment on Rockchip devices. Includes support for key parameters such as `imgsz`, `batch`, and `name`.
- 📄 **Integration Documentation**:
- **Rockchip RKNN**: Expanded instructions, performance benchmarks, and FAQs for smoother deployment.
- **Seeed Studio reCamera**: Introduced documentation for using YOLO models with the reCamera for edge AI, including ONNX and cvimodel exports.
- 🚀 **Optimizations and Fixes**:
- Streamlined RKNN export code for better clarity and reliability.
- Fixed ONNX model path issue to resolve export naming conflicts.
- Enhanced debugging during ONNXRuntime CUDA initialization.
- Improved label class validation logic to prevent dataset misconfigurations.
- Updated Albumentations' `ImageCompression` augmentation range for higher realism.
- 📦 **Testing Enhancements**:
- Added CI support for **Ubuntu ARM64** builds, enhancing platform compatibility for ARM-based environments.
- 🔧 **Code Improvements**:
- Introduced a custom `TQDM` class for consistent progress bar functionality.
- Refactored unused arguments in modules like `TorchVision` and `Index`.
- Adjusted optimizer logic during training for better performance in DDP setups.

---

🎯 Purpose & Impact
- **🚀 Expanded Hardware Reach**: Rockchip RKNN and Seeed Studio reCamera integration allow effortless deployment on specialized hardware, facilitating edge AI applications like real-time object detection and energy-efficient designs.
- **🔗 Enhanced Usability**: Rich documentation, benchmarks, and FAQs guide developers through complex setups, broadening accessibility for newcomers.
- **✅ More Robust Exports**: RKNN and ONNX updates improve compatibility and prevent export errors, reducing troubleshooting time for developers.
- **⚡ Performance Gains**: Augmentation and label validation improve model robustness and reduce errors during training and deployment across datasets and hardware.
- **🛠 Streamlined Development**: Refactors simplify code maintenance while maintaining compatibility, fostering a cleaner codebase.

What's Changed
* Update thumbnail for Rockchip RKNN integration page by lakshanthad in https://github.com/ultralytics/ultralytics/pull/18787
* Cleanup TorchVision related functions by Y-T-G in https://github.com/ultralytics/ultralytics/pull/18790
* Fix IMX onnx model path by Laughing-q in https://github.com/ultralytics/ultralytics/pull/18813
* New Seeedstudio reCamera Docs page by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/18801
* Add https://youtu.be/8gePl_Zcs5c to docs by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/18824
* Fix missing IMX500 export decorator by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18823
* Fix spelling by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18827
* Fix dataset category indexes check by Laughing-q in https://github.com/ultralytics/ultralytics/pull/18840
* Ubuntu ARM GitHub CI runners by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18762
* Fix automatic optimizer LR with DDP training by Laughing-q in https://github.com/ultralytics/ultralytics/pull/18842
* Update inference.cpp [bug in case of cudaEnable = false] by pmermigkas in https://github.com/ultralytics/ultralytics/pull/18831
* RKNN export clean up by Laughing-q in https://github.com/ultralytics/ultralytics/pull/18841
* Move `task=classify` with `mode=track` warning to tracker `on_predict_start` by Laughing-q in https://github.com/ultralytics/ultralytics/pull/18837
* Use TQDM class by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18846
* Fix Albumentations ImageCompression `quality_range` arg by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18847
* `ultralytics 8.3.66` add Rockchip RKNN export in `tutorial.ipynb` by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/18848

New Contributors
* pmermigkas made their first contribution in https://github.com/ultralytics/ultralytics/pull/18831

**Full Changelog**: https://github.com/ultralytics/ultralytics/compare/v8.3.65...v8.3.66

📊 Key Changes
- **Rockchip RKNN Integration**:
- Added RKNN export for YOLO models optimized for Rockchip hardware (e.g., RK3588, RK3566).
- Simplified deployment with enhanced documentation and tools for RKNN models.
- Supported hardware inference via `rknn-toolkit2` with assisted device compatibility checks.

- **Stability and Performance Improvements**:
- Enhanced data loader robustness by handling worker termination safely under edge cases. ✅
- Updated CI workflows to support macOS 15, ensuring compatibility with the latest macOS environments.

- **Compatibility Fixes**:
- Dynamic handling of `numpy` dependencies for NVIDIA Jetson devices to improve TensorRT functionality, reducing rigid constraints for all other users. 🌍

- **Refactoring**:
- Replaced mutable Python `set` with immutable `frozenset` across codebase to improve performance, ensure thread safety, and prevent unintended data modifications. 🚀

- **Documentation Cleanup and Maintenance**:
- Updated regex for consistent link conversion in documentation (plaintext to HTML), simplifying maintenance and improving reliability. ✍️

🎯 Purpose & Impact
- **Purpose**:
- Simplify AI deployment for edge devices, particularly Rockchip-based hardware, using RKNN format.
- Improve the user experience by addressing edge-case errors in data loaders and ensuring compatibility with macOS and NVIDIA-specific scenarios.
- Modernize internal code structure for faster performance and better reliability.

- **Impact**:
- 🧠 **RKNN Support**: Developers now have a streamlined process to export and deploy YOLO models on Rockchip's NPU-enabled devices, unlocking high-performance AI functionality for embedded systems.
- ✅ **Enhanced Stability**: Reduced chances of crashes by safely handling resource cleanup issues (e.g., in data loaders).
- 📈 **Optimized Performance**: Better immutability within system configurations creates a stable baseline for developers working in multi-threaded environments.
- 📚 **Improved Documentation**: Cleaner formatting and precise integrations make it easier for users to implement new features and understand their capabilities.

This release empowers developers with new deployment options while improving the robustness and maintainability of the toolset. 🚀

What's Changed
* Catch and ignore exception in dataloader cleanup by Y-T-G in https://github.com/ultralytics/ultralytics/pull/18772
* Pin `numpy` 1.23.5 for JetPack 4 on NVIDIA Jetson Nano by lakshanthad in https://github.com/ultralytics/ultralytics/pull/18783
* Use frozenset() by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18785
* Adopt `macos-15` GitHub CI runners by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18763
* Update convert_plaintext_links_to_html by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18786
* `ultralytics 8.3.65` Rockchip RKNN Integration for Ultralytics YOLO models by IvorZhu331 in https://github.com/ultralytics/ultralytics/pull/16308


**Full Changelog**: https://github.com/ultralytics/ultralytics/compare/v8.3.64...v8.3.65