Ultralytics

🌟 Summary
The `v8.3.47` release focuses on enhancing YOLO's classification capabilities by improving model export, prediction handling, and loss computation for smoother training and deployment. 🚀

📊 Key Changes
- **Improved YOLO Classification Module**:
- Added export functionality (`export=True`) to the classification head, enabling easier deployment. 📤
- Enhanced post-processing to handle predictions more efficiently, including better support for tuple-based outputs. ⚙️
- Updated loss function to gracefully manage tuple-based predictions for accurate classification loss calculation. 📊
- Updated logic in the classification head to switch seamlessly between training and inference, including integrated softmax during inference. 🔄

- **Documentation Improvements**:
- Clarified functionality for `copy_paste` augmentation, specifying it requires segmentation labels. ✍️
- Fixed typos and improved YAML configurations for smoother developer workflows. 🛠
- Removed redundant documentation entries to enhance clarity. 📚

🎯 Purpose & Impact
- **For End Users**:
- **Enhanced Deployment Options**: You can now export YOLO classification models more effectively, unlocking advanced deployment use cases in various environments. 🌐
- **Streamlined Workflows**: Updated logic ensures classification workflows are more intuitive and robust—ideal for developers juggling both training and inference. 🌟

- **For Developers**:
- **Clearer Documentation**: Clarifications and typo fixes reduce confusion, saving time while implementing advanced features like `copy_paste`. ✨
- **Improved Configuration**: Fixes to YAML workflows ensure consistent deployments and smoother integration during the development lifecycle. ✅

Overall, this release significantly improves YOLO's flexibility and usability for classification tasks while enhancing the clarity of supporting documentation. 💡

What's Changed
* Fix Docs YAML boolean by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18072
* Eliminate duplicate bullet points in docs by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/18074
* Mention `copy_paste` applies only when using segmentation labels by Y-T-G in https://github.com/ultralytics/ultralytics/pull/18077
* `ultralytics 8.3.47` fix `softmax` and `Classify` head commonality by Laughing-q in https://github.com/ultralytics/ultralytics/pull/18085


**Full Changelog**: https://github.com/ultralytics/ultralytics/compare/v8.3.44...v8.3.47

---

📊 Key Changes
Triton Inference Enhancements:
- Added ability to **retrieve and store model metadata** during export (`on_export_end` callback).
- Triton Model Repository configurations now dynamically include metadata (`config.pbtxt`).
- Enhanced `TritonRemoteModel` to handle metadata for better customization and traceability.
- Set a default task (`task=detect`) for Triton Server models when unspecified.

Core Features & Fixes:
- **Dependency Change**: Reverted from `lapx` back to the original, widely-used `lap` package for improved compatibility and stability.
- **Model Dynamics**: Refined handling of dynamic ONNX models by intelligently setting the `dynamic` property based on input shape rather than batch size.
- **Custom Backend Flexibility**: Expanded `AutoBackend` to directly accept PyTorch models in memory, in addition to weight file paths.
- **AMP Checks**: Hardcoded failing GPUs (e.g., GTX 16 series, Quadro T series) to prevent NaN losses in training when using AMP (Automatic Mixed Precision).
- **New Utility**: Introduced `empty_like` function to streamline operations and improve maintainability for tensors and arrays.
- **Segment Resampling Fix**: Resolved issues with preserving original segment points during resampling.

---

🎯 Purpose & Impact
💡 **Triton Support Improvements**:
- Enables smooth deployment configurations with enriched metadata providing better traceability and error prevention. Streamlines the export process for Triton Inference Server users.

🧠 **User-Friendly Adjustments**:
- Setting a default task reduces errors for users unfamiliar with Triton configurations.
- Allowing in-memory PyTorch models boosts integration convenience in customized workflows.

🚀 **Stability and Performance Gains**:
- Utilizing `lap` simplifies dependency management and ensures better compatibility in tracking workflows.
- Fixed AMP-related compatibility issues for specific GPUs, preventing unstable training outcomes.

🔧 **Code Optimization**:
- Introduced the `empty_like` utility to centralize tensor creation logic, reducing repetitive code and improving maintainability.

This update bridges gaps in stability, scalability, and user experience for Triton deployment and core Ultralytics functionalities, ensuring a smoother and more reliable platform for all users! 🌟

What's Changed
* Revert `lapx` to `lap` by Laughing-q in https://github.com/ultralytics/ultralytics/pull/17908
* Preserve original points in `resample_segments` by Y-T-G in https://github.com/ultralytics/ultralytics/pull/18051
* Hardcode failing GPUs in AMP checks by Y-T-G in https://github.com/ultralytics/ultralytics/pull/17977
* Set `dynamic` to True only if imgsz is dynamic for ONNX by Y-T-G in https://github.com/ultralytics/ultralytics/pull/17872
* Set `task` for Triton inferenceTriton by Laughing-q in https://github.com/ultralytics/ultralytics/pull/18064
* Modified the parameter description of 'weights' in the AutoBackend class by ye-yangshuo in https://github.com/ultralytics/ultralytics/pull/18059
* Fix `np.empty_like` and `torch.empty_like` input type by Laughing-q in https://github.com/ultralytics/ultralytics/pull/18062
* `ultralytics 8.3.44` improve Triton Inference Server metadata by Y-T-G in https://github.com/ultralytics/ultralytics/pull/17921

New Contributors
* ye-yangshuo made their first contribution in https://github.com/ultralytics/ultralytics/pull/18059

**Full Changelog**: https://github.com/ultralytics/ultralytics/compare/v8.3.43...v8.3.44

📊 Key Changes
- **PyPI Security Fix**: Ensures secure publication workflow for the Ultralytics package. 🔒
- **Batch Processing**: Added a `batch` parameter to `predict()` for faster processing of large datasets or videos. 🏎️
- **Segmentation Enhancement**: Adjusted output handling for segmentation tasks to address class name issues and improve robustness. 🛠️
- **Improved Examples**: Simplified heatmap and queue management examples for clearer usage. 📖
- **Dynamic Benchmarking**: Made version control for Raspberry Pi benchmarks easier in docs. 🔄
- **YouTube Tutorial Link**: Added a direct video link in `tutorial.ipynb` for enhanced learning. 🎥
- **Documentation Updates**: Fixed and updated links, added a `homepage` footer link, and improved accuracy in various docs. 🌐
- **Bug Fix**: Resolved an integer conversion issue in `xyxy2xywhn` for accurate results. 🐛

🎯 Purpose & Impact
- **Secure Publishing**: Protects the integrity of updates published on PyPI, safeguarding against potential threats. ✔️
- **Faster Inference**: The new `batch` parameter enables higher throughput, particularly valuable for power users with demanding workloads. 🚀
- **Better Segmentation**: Improved handling of segmentation workflows reduces errors, enhancing reliability for users working on complex datasets. ✅
- **User-Friendly**: Clearer examples and accurate docs reduce entry barriers and enhance the learning experience for beginners. 📘
- **Streamlined Workflow**: Innovations in dependency and workflow management make development and contribution processes smoother for maintainers. 🤝
- **Learning Resources**: A new tutorial video provides a practical guide to using Ultralytics YOLO models with Colab, bridging knowledge gaps for users. 🎬

This release balances technical fixes and documentation improvements, making it more robust, user-friendly, and secure! 🎉

What's Changed
* Update mkdocs.yml homepage by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17924
* Add `batch` to list of `predict()` arguments by Y-T-G in https://github.com/ultralytics/ultralytics/pull/17979
* Add https://youtu.be/ZN3nRZT7b24 to tutorial.ipynb by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/17875
* Improve docs for Ultralytics version in benchmarks by lakshanthad in https://github.com/ultralytics/ultralytics/pull/17925
* `ultralytics 8.3.42` update AutoBackend `names` placeholder by Laughing-q in https://github.com/ultralytics/ultralytics/pull/17970
* Update CLI message by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/18041
* Update functions description by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/18050
* Fix `np.empty_like` function input type by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/18049
* `ultralytics 8.3.43` PyPI publishing security fix by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/18052


**Full Changelog**: https://github.com/ultralytics/ultralytics/compare/v8.3.40...v8.3.43

📊 Key Changes
- **TrackZone Added**: A new solution for zone-based tracking, allowing users to monitor objects in custom-defined regions.
- **Enhanced Documentation**: Detailed guidance on `TrackZone` usage, its arguments, and real-world applications added. 📝
- **Framework Updates**: Improvements to tracking arguments, CI dependency handling, and updated Raspberry Pi benchmarks.

🎯 Purpose & Impact
- **More Precise Analytics**: By confining tracking to user-defined zones, the solution optimizes resource usage and allows fine-grained insights for scenarios like surveillance, crowd management, and industrial monitoring. 🚨
- **Simplifies Complex Applications**: Users can now easily define and analyze specific areas of interest without needing to process unnecessary parts of a video feed, reducing computational overhead. 🚀
- **Improved Documentation and Benchmarks**: Helps users navigate with ease while accessing expanded Raspberry Pi benchmarks for better framework comparison. 💡

Example Use Case
For instance, in a security application, you can define a "restricted area" within a camera feed and monitor only that zone for intrusions, improving both performance and practicality. 🛡️

What's Changed
* GPU CI Fix wrong Ultralytics Installation by Skillnoob in https://github.com/ultralytics/ultralytics/pull/17883
* Fix typo in Sony IMX500 doc by lakshanthad in https://github.com/ultralytics/ultralytics/pull/17871
* wrong expression fix Update README.md by ArtificialZeng in https://github.com/ultralytics/ultralytics/pull/17868
* Add more tracking args for solutions by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/17878
* `uv pip install` for Benchmarks by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17749
* Add `MNN` benchmarks to Raspberry Pi doc by lakshanthad in https://github.com/ultralytics/ultralytics/pull/17910
* Add arguments for solutions in `usage/cfg` docs page by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/17920
* `ultralytics 8.3.40` new `TrackZone` Solution by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/17918

New Contributors
* ArtificialZeng made their first contribution in https://github.com/ultralytics/ultralytics/pull/17868

**Full Changelog**: https://github.com/ultralytics/ultralytics/compare/v8.3.39...v8.3.40

🌟 Summary
The Ultralytics `v8.3.39` release focuses on improving model behavior, functionality, and user experience across multiple aspects, including classification validation, documentation enhancements, and tool usability. It introduces critical fixes and new features to improve the overall quality of the platform. 🚀

---

📊 Key Changes

- 🧠 **Fixed Classification Validation Loss:**
- Adjusted classification model's loss scaling during validation to improve output consistency and accuracy.
- Introduced a refined approach to apply `softmax` only in necessary scenarios for clarity and precision.

- 🎯 **"Classes" Filter in Training:**
- Added a new `classes` argument to the training configuration, enabling model training on specific class IDs selectively.

- 🎥 **Enhanced Video Annotation Tool:**
- Introduced a "Sweep Annotation" utility for dynamic video annotation. Users can now visualize objects based on an interactive sweep line that tracks their positions.

- 🎨 **Improved Color Handling in LibTorch Example**:
- Addressed a key issue by adding a **BGR to RGB conversion** step in the C++ LibTorch inference example, ensuring color compatibility for accurate YOLO results.

- 🗂️ **Documentation Updates:**
- Significant improvements in README files:
- Clickable YOLO11 performance plot images now redirect to documentation.
- Enhanced clarity about model auto-download behavior and training details.
- Added new high-quality tutorial videos across docs for better onboarding and understanding.
- Fixed `YOLOv11` references to the correct term `YOLO11` for consistency.

- ⚙️ **Code Improvements and Maintenance:**
- Simplified segmentation handling with better clipping (`clip()`) for out-of-bounds coordinates in segmentation tasks.
- Added an elegant `__getattr__` method making model attributes (e.g., `stride` or `task`) directly accessible from the `Model` class.
- Refined model logging for better debugging and developer experience.

---

🎯 Purpose & Impact

- **Enhanced Accuracy and Model Behavior**: The classification loss scaling fix addresses a crucial inconsistency, delivering more reliable results during validation phases.
- **Increased Flexibility**: The "classes" argument empowers users with precise control, making training workflows more tailored and efficient by focusing on specific class IDs. 💡
- **Better Video Annotation**: The "Sweep Annotation" tool adds an intuitive way to annotate video data interactively, offering new possibilities for detection and tracking tasks.
- **Improved Inference Quality**: The BGR to RGB fix ensures accurate detections for users operating in C++ environments with LibTorch inference.
- **Streamlined User Education**: Updated and accessible documentation alongside engaging video tutorials helps onboard new users quickly while enhancing knowledge for experienced developers. 📚
- **Consistency**: Terminology such as `YOLO11` aligned across documentation ensures clarity and avoids user confusion.

This release keeps refining both functionality and usability, advancing the YOLO ecosystem for a diverse range of practical applications. 🎉

What's Changed
* Add YOLO11 docs page redirect in `README.md` and `README.zh-CN.md` by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/17806
* Fix missing labels when all segment points are out of bounds by Y-T-G in https://github.com/ultralytics/ultralytics/pull/17810
* New Solutions sweep counting annotator by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/17742
* Improved Docs minify by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17816
* Revert Docs minify attempt by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17831
* Update format.yml Discord and Kaggle links by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17814
* Update contributing.md with open-sourcing guide by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17832
* Fix YOLO11 usage by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17834
* Update Tasks Banner in README.md by pderrenger in https://github.com/ultralytics/ultralytics/pull/17833
* Fix region-counting indents by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17835
* Update Tasks banner spacing by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17843
* Add functions descriptions in `plotting.py` by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/17813
* Remove Docs Tasks banner linebreak by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17844
* Add https://youtu.be/-aYO-6VaDrw and https://youtu.be/M7xWw4Iodhg to docs by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/17863
* Add `classes` to train arguments by Y-T-G in https://github.com/ultralytics/ultralytics/pull/17856
* Add BGR to RGB conversion in LibTorch example by Y-T-G in https://github.com/ultralytics/ultralytics/pull/17864
* `__getattr__` support to access YOLO attributes via Model class by WYYAHYT in https://github.com/ultralytics/ultralytics/pull/17805
* `ultralytics 8.3.39` fix classification validation loss scaling by Y-T-G in https://github.com/ultralytics/ultralytics/pull/17851

New Contributors
* WYYAHYT made their first contribution in https://github.com/ultralytics/ultralytics/pull/17805

**Full Changelog**: https://github.com/ultralytics/ultralytics/compare/v8.3.38...v8.3.39

🌟 Summary
The release of 'v8.3.38' introduces significant enhancements, particularly emphasizing video interaction capabilities through the new `SAM2VideoPredictor` class for object segmentation and tracking in videos. This update also includes general improvements and optimizations across various modules.

📊 Key Changes
- **SAM2VideoPredictor**: A new class aimed at enhancing video segmentation and object tracking, supporting advanced interactions such as prompts for segment modifications.
- **Improved Video Segmentation**: Features non-overlapping masks, better memory management, and support for interactive user prompts.
- **Configuration Clean-Up**: Removal of obsolete parameters such as `label_smoothing`.
- **Platform Compatibility**: Extended detection for NVIDIA Jetson devices, accommodating more models.
- **Documentation and Code Updates**: Adjustments for improved clarity and accuracy in both code and documentation.

🎯 Purpose & Impact
- 📽️ **Enhanced Video Interaction**: The `SAM2VideoPredictor` allows users to fine-tune video processing outputs dynamically, making video segmentation more precise and interactive.
- 🚀 **Efficiency & Resource Management**: Optimized memory use during video segmentation leads to faster inferencing and resource savings, beneficial for running on resource-constrained devices.
- 🛠️ **Code Simplification**: Removing unnecessary parameters like `label_smoothing` helps streamline configuration settings, reducing potential user confusion.
- 📱 **Broader Device Support**: Updating device compatibility ensures the software is functional across a wider range of hardware, improving the user experience for those utilizing NVIDIA Jetson platforms.
- 📚 **Improved User Documentation**: Enhanced documentation aids both beginners and advanced users by making it easier to understand and implement model configurations and changes efficiently.

What's Changed
* Delete .github/workflows/codeql.yaml by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17772
* Fix `RepC3` module for RT-DETR models by Andrewymd in https://github.com/ultralytics/ultralytics/pull/17086
* Removes unused argument `label_smoothing` by Burhan-Q in https://github.com/ultralytics/ultralytics/pull/16014
* Concat all segments by default for multi-part masks by Y-T-G in https://github.com/ultralytics/ultralytics/pull/16826
* Improve `is_jetson` to support more Jetson devices by lakshanthad in https://github.com/ultralytics/ultralytics/pull/17770
* Fix DLA export by Laughing-q in https://github.com/ultralytics/ultralytics/pull/17765
* Fix CI.md CodeQL badges by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17779
* Fix Prettier docs issues by glenn-jocher in https://github.com/ultralytics/ultralytics/pull/17798
* Convert frames to RGB before SAHI inference by Y-T-G in https://github.com/ultralytics/ultralytics/pull/17795
* Fix parking points annotator for macOS by RizwanMunawar in https://github.com/ultralytics/ultralytics/pull/17797
* Fix `forward_fuse` comment by arun477 in https://github.com/ultralytics/ultralytics/pull/17714
* `ultralytics 8.3.38` SAM 2 video inference by Laughing-q in https://github.com/ultralytics/ultralytics/pull/14851

New Contributors
* Andrewymd made their first contribution in https://github.com/ultralytics/ultralytics/pull/17086
* arun477 made their first contribution in https://github.com/ultralytics/ultralytics/pull/17714

**Full Changelog**: https://github.com/ultralytics/ultralytics/compare/v8.3.37...v8.3.38