v8.2.61 - `ultralytics 8.2.61` fix `num_threads` for CPU training (#14544)

🌟 Summary

Ultralytics v8.2.61 improves documentation and code understandability alongside some performance updates.

📊 Key Changes

• Documentation Script Improvement: Enhanced the build_docs.py for better documentation processing.
  • Added BeautifulSoup for the HTML processing.
  • Converts plaintext links to HTML hyperlinks.
• YOLOv9 Benchmarks: Added detailed benchmarking information for YOLOv9.
• Documentation Refinements: Adjusted references and docstrings across multiple files to enhance readability and clarity.
• Class Descriptions Added: More comprehensive docstrings added to classes for better understanding.
• Configuration File Update: Adjustments made to support new configurations and improve user guidance.

🎯 Purpose & Impact

• Better Documentation Navigation: 📚 Enhances user experience by ensuring documentation links are properly formatted, allowing seamless navigation through web-based documents.
• Benchmarking Insights: 📊 Provides users with a clear understanding of how to benchmark YOLOv9 models, improving model evaluation processes.
• Improved Readability: ✍️ Refined documentation and class descriptions make the codebase more approachable for both new and existing users.
• Enhanced Configuration Management: 🛠️ Better guiding users in configuring and using the YOLO models effectively by offering detailed configuration examples and handling deprecated keys more elegantly.

What's Changed

• Fix Multi-GPU trained model export by @Y-T-G in #14551
• Fix Docs pretty __init__.py URLs by @glenn-jocher in #14550
• Add https://youtu.be/ziJR01lKnio to docs by @RizwanMunawar in #14554
• Add Custom CLIP Model Download Path by @wangsrGit119 in #14517
• ultralytics 8.2.61 fix num_threads for CPU training by @Laughing-q in #14544

New Contributors

• @wangsrGit119 made their first contribution in #14517

Full Changelog: v8.2.60...v8.2.61

v8.2.60 - `ultralytics 8.2.60` refactor `process_mask_upsample` (#14474)

🌟 Summary

The Ultralytics v8.2.60 release includes several updates that enhance functionality and improve performance, along with documentation improvements and bug fixes.

📊 Key Changes

• 🖼️ Enhanced Documentation for Datasets: Added an embedded YouTube video tutorial for the Caltech-256 dataset.
• 🔍 Improved Validation Scripts: Refined the detection and segmentation validation scripts with more precise checks for various datasets.
• 🛠️ Code Cleanup: Removed the process_mask_upsample function to streamline code and improve readability.
• 📝 Version Update: Incremented the version number from 8.2.59 to 8.2.60.
• 🏗️ Export Fix: Updated the file-moving operation in the NCNN exporter to use shutil.move() for improved reliability and clarity.

🎯 Purpose & Impact

• 🎓 Educational Resource: The new tutorial video in the Caltech-256 dataset documentation helps users quickly get up to speed with training image classification models using the dataset.
• 🧩 Validation Accuracy: Updates to the validation scripts ensure more accurate dataset handling, especially for COCO and LVIS datasets, which is crucial for reliable model evaluation.
• 🚀 Performance Enhancement: Removal of unnecessary functions and streamline of code can lead to more efficient operations, reducing potential confusion and simplifying maintenance.
• 📦 Export Reliability: The updated file-moving operation in the exporter script minimizes errors and improves the robustness of model deployment workflows.

What's Changed

• Allow test-dev2017.txt for val by @yang-0201 in #14519
• Add https://youtu.be/isc06_9qnM0 to docs by @RizwanMunawar in #14525
• Fix NCNN multiple-volumes PNNX download bug by @akukulanski in #14533
• ultralytics 8.2.60 refactor process_mask_upsample by @Laughing-q in #14474

New Contributors

• @yang-0201 made their first contribution in #14519
• @akukulanski made their first contribution in #14533

Full Changelog: v8.2.59...v8.2.60

v8.2.59 - `ultralytics 8.2.59` use `Results.save_txt` for validation (#14496)

🌟 Summary

The 'v8.2.59' release of Ultralytics makes several important enhancements and fixes to improve user experience and model performance.

📊 Key Changes

• Update training command references to use more specific model types (yolov8n-pose.pt, yolov8n-seg.pt).
• Added a detailed integration guide for using Kaggle, including new features and troubleshooting sections.
• Enhanced annotation and result management in the documentation for FastSAM and SAM models.
• Added support for txt saving with better management of predicted results for various YOLO model types (detection, OBB, pose, and segmentation).
• Introduced warning for incorrect mAP calculations when save_hybrid=True is enabled.

🎯 Purpose & Impact

• Model Training Simplification: By specifying more precise model types in training commands, users can more easily ensure they are using the correct model for their tasks, minimizing errors and confusion.
• Kaggle Integration: This new guide allows users to leverage Kaggle's free cloud computing resources more effectively, helping them train and evaluate models without needing expensive local hardware.
• Enhanced Documentation: Providing clearer documentation with examples for managing annotations and results makes the framework more accessible, particularly for newer users.
• Improved Result Handling: Updates to result saving procedures ensure that key prediction data are accurately recorded, thereby enabling better analysis and reproducibility of experiments.
• User Notifications: Adding a warning for save_hybrid=True prevents users from misinterpreting their models' mAP scores, promoting a more accurate evaluation process.

What's Changed

• Fix model parameter in Pose, Segment dataset train examples by @lunarifish in #14505
• Warn on save_hybrid=True by @glenn-jocher in #14484
• Remove redundant assignment by @aryanjassal in #14498
• Update FastSAM and SAM docs by @Laughing-q in #14499
• Add Kaggle Integrations Docs Page by @abirami-vina in #14487
• Fix Enable Tracking Button and Optimize FPS in Streamlit Application by @lakshanthad in #14508
• ultralytics 8.2.59 use Results.save_txt for validation by @Laughing-q in #14496

New Contributors

• @lunarifish made their first contribution in #14505
• @aryanjassal made their first contribution in #14498

Full Changelog: v8.2.58...v8.2.59

v8.2.58 - `ultralytics 8.2.58` FastSAM code refactor (#14450)

🌟 Summary (v8.2.58)

Enhanced YOLO model and GUI integrations with updated dependencies and optimized image processing methods.

📊 Key Changes

• Dependency updates: Increased minimum Streamlit version to 1.29.0.
• Image processing: Changes in how images and masks are handled within the fastsam and validation modules.
• Code optimization: Refined method docstrings and added in-document scoping of tkinter and Pillow (PIL) imports.

🎯 Purpose & Impact

• Dependency updates: Ensures compatibility with the latest features and bug fixes on Streamlit, improving overall application stability and performance.
• Image processing: More efficient handling of image crops and mask segmentations improves the accuracy of model predictions.
• Code optimization: Enhances readability and maintainability. Scoped imports prevent unnecessary loading of large libraries, reducing startup times and potential conflicts.

What's Changed

• Scope tkinter and streamlit in Solutions tests by @glenn-jocher in #14426
• Fix _process_batch() docstrings by @glenn-jocher in #14454
• ultralytics 8.2.58 FastSAM code refactor by @maycuatroi in #14450

Full Changelog: v8.2.57...v8.2.58

v8.2.57 - `ultralytics 8.2.57` new Solutions Tests and Docs (#14408)

🌟 Summary

The v8.2.57 update for Ultralytics introduces a minor but crucial terminology update across several documentation and code files and enhances instance segmentation and tracking visualizations.

📊 Key Changes

• Terminology Update: Changed the term classes_names to just names in documentation and code.
• Instance Segmentation Enhancements: Introduced txt_color for customizing label text colors.
• New Tests Added: Created test_solutions.py to test various features like object counting, heatmaps, etc.
• Core Model Update: Ensured the model is appropriately copied before saving.

🎯 Purpose & Impact

• Simplified Terminology: The change from classes_names to names simplifies variable names, making the codebase easier to read and maintain.
• Improved Visuals: Allowing customization of text color for labels in segmented objects will enhance visualization clarity.
• Robust Testing: Adding comprehensive tests ensures the reliability and stability of various solutions provided by Ultralytics.
• Enhanced Model Handling: Copying the model before saving prevents potential issues with model state consistency, improving the robustness of model persistence.

These updates improve usability and ensure the software is more reliable and user-friendly, benefiting both new and experienced users.

What's Changed

• Fix TORCHVISION_0_18 for allow_empty=True by @glenn-jocher in #14415
• Add https://youtu.be/gX5kSRD56Gs to docs by @RizwanMunawar in #14417
• Fix model.save() method to FP16 by @glenn-jocher in #14418
• ultralytics 8.2.57 new Solutions Tests and Docs by @RizwanMunawar in #14408

Full Changelog: v8.2.56...v8.2.57

v8.2.56 - `ultralytics 8.2.56` Streamlit tracking app (#14269)

🌟 Summary

Release v8.2.56 introduces updates aimed at optimizing model selection and inference processes in the Ultralytics YOLO framework.

📊 Key Changes

• 🔄 Updated YOLOv8 example code link in the Rust demo README.
• 🔧 Streamlined model selection in the Streamlit inference solution by fetching model names dynamically.
• 🚀 Added an option to enable or disable tracking in the Streamlit app.
• 📉 Updated confidence and IoU threshold slider labels.

🎯 Purpose & Impact

• 📖 Easier Navigation: The clearer link in the Rust demo README makes it simpler for users to find updated example code.
• 🍃 Optimized Model Selection: Dynamically fetching available models reduces the likelihood of user error and simplifies maintenance.
• 👣 Enhanced Inference Control: Allowing users to toggle tracking on and off provides more control over the inference process, adding flexibility for different use cases.
• 🎛 Improved User Interface: Renaming sliders for confidence and IoU thresholds enhances readability and user experience.

These changes focus on improving the ease of use, flexibility, and efficiency of the YOLO framework for both developers and non-expert users, making the deep learning process smoother and more intuitive.

What's Changed

• ultralytics 8.2.56 Streamlit tracking app by @RizwanMunawar in #14269

Full Changelog: v8.2.55...v8.2.56

v8.2.55 - `ultralytics 8.2.55` adaptive `tflite_support` logic (#14385)

🌟 Summary

The Ultralytics v8.2.55 update comes with improved documentation, better compatibility for ARM64 systems, an adjusted rate limit, and enhanced logging readability.

📊 Key Changes

• 📝 Documentation Enhancement: Added a YouTube video tutorial link to the CIFAR-10 dataset documentation.
• 🔄 Exporter Update: Improved compatibility for TensorFlow Lite models, especially on ARM64 systems.
• ⏱️ Rate Limit Adjustment: The event rate limit has been reduced from 60 seconds to 30 seconds.
• 📈 Logging Improvement: Enhanced log readability by formatting numerical values with commas.

🎯 Purpose & Impact

• Documentation Enhancement: Provides users with a video tutorial, making it easier to understand how to train an image classification model with the CIFAR-10 dataset, especially beneficial for beginners.
• Exporter Update: Ensures smoother operation on ARM64 systems by handling potential import issues more gracefully, improving reliability for users working on such architectures.
• Rate Limit Adjustment: More frequent event updates, potentially leading to more responsive user feedback during operations.
• Logging Improvement: Makes the logs clearer and more interpretable by formatting large numbers with commas, aiding in quick comprehension of complex model summaries.

What's Changed

• Add https://youtu.be/fLBbyhPbWzY to docs by @RizwanMunawar in #14356
• ultralytics 8.2.55 adaptive tflite_support logic by @glenn-jocher in #14385

Full Changelog: v8.2.54...v8.2.55

v8.2.54 - `ultralytics 8.2.54` resolve YouTube bug with switch to `pytubefix` (#14354)

🌟 Summary

The v8.2.54 release of Ultralytics brings a minor update primarily focusing on improving video stream handling and expanding model assets.

📊 Key Changes

• 🚀 Video Stream Handling
  • Switched from using pytube to pytubefix for fetching YouTube video streams to resolve a known issue (pytube issue #1954).
• 🔄 Model Assets Expansion
  • Added support for additional model assets, including the yolov8 series with the -oiv7 suffix.

🎯 Purpose & Impact

• Enhanced Reliability: The switch to pytubefix ensures more reliable and error-free downloading of YouTube videos, addressing a specific bug that users encountered with pytube. This change will improve the experience for users who rely on video stream data.
• Broader Model Options: By expanding the range of supported model assets, users now have more flexibility and options to choose from for various applications. This can help in tailoring model performance to specific needs more effectively.

What's Changed

• ultralytics 8.2.54 resolve YouTube bug with switch to pytubefix by @glenn-jocher in #14354

Full Changelog: v8.2.53...v8.2.54

v8.2.53 - `ultralytics 8.2.53` Heatmaps fix for empty images (#14329)

🌟 Summary

Ultralytics v8.2.53 release primarily focuses on updating download links in various scripts, configuration files, and documentation.

📊 Key Changes

• Dockerfiles Update: Changed the source URLs for downloading tensorstore, torch, and torchvision for ARM packages.
• Dataset Configuration: Updated the download URLs for multiple datasets (e.g., ImageNet10, GlobalWheat2020, VOC, VisDrone) to new asset links.
• Documentation Update: Changed image and dataset links in several documentation files.
• Test Scripts Update: Modified test scripts to use new URLs for downloading test resources.
• Minor Code Tweaks: Adjusted the __init__.py version and made small enhancements to heatmap and training scripts.

🎯 Purpose & Impact

• URL Standardization: Streamlining access to resources by consolidating download links to a unified assets repository ensures consistency and reliability.
• Improved User Experience: Simplifies the process for users to find and download necessary components without broken links.
• Enhanced Documentation: Keeping documentation up-to-date with the latest links ensures users can follow instructions without issue.
• Stable Testing Environment: Ensures that test environments remain consistent and dependable with updated resource links.
• Minor, Yet Crucial: These changes, although minor, are crucial for maintaining the usability and stability of the software, especially for automated scripts and pipelines.

What's Changed

• BaseTrainer with find_unused_parameters=True when using DistributedDataParallel() by @Tsai-chia-hsiang in #14323
• Ultralytics Asset URL Update by @glenn-jocher in #14345
• ultralytics 8.2.53 Heatmaps fix for empty images by @ambitious-octopus in #14329

New Contributors

• @Tsai-chia-hsiang made their first contribution in #14323

Full Changelog: v8.2.52...v8.2.53

v8.2.52 - `ultralytics 8.2.52` fix CenterCrop transforms for PIL Image inputs (#14308)

🌟 Summary

Release v8.2.52 of Ultralytics introduces detailed guidance on maintaining computer vision models, along with modifications to various dataset download links and a few minor updates for better functionality.

📊 Key Changes

• New Guide Added: 📘 "Maintaining Your Computer Vision Model."
• Dataset Links Updated: 🗂 Many dataset download links now point to GitHub instead of the Ultralytics website.
• Testing URLs Adjusted: 🔄 Updated certain URLs in test scripts to point to new locations.
• Conversion Update: 🔄 Adjustments to ensure compatibility with torch 1.13+ for some model and dataset handling functions.
• Various Fixes: 🛠️ Including typo corrections, descriptor enhancements, and code improvements.

🎯 Purpose & Impact

• Comprehensive Model Maintenance Guide:

  • Purpose: 📚 To help users understand how to keep their models accurate and functional post-deployment through monitoring, anomaly detection, and retraining strategies.
  • Impact: 🌟 Enhanced reliability and performance of deployed computer vision models; helps users identify and address model drift and anomalies proactively.

• Updated Dataset Links:

  • Purpose: 🌐 Ensure the datasets are accessible with feature improvements using GitHub releases.
  • Impact: 📦 Users will experience streamlined, reliable access to datasets, avoiding potential downtime or slowdowns previously encountered with older links.

• Compatibility Fixes:

  • Purpose: 🔄 To maintain seamless conversion and implementation especially in environments running torch 1.13+.
  • Impact: 🚀 Users working with recent versions of torch will benefit from increased stability and compatibility, facilitating smoother model training and deployment.

In Summary

The v8.2.52 update is geared toward ensuring continuous model effectiveness and improving data access and functional resilience, leading to a more robust and user-friendly experience.

What's Changed

• Allow OpenVINO export from CUDA by @wh1t3h47 in #14256
• Fix end2end attribute in init_criterion by @Alejandro-Casanova in #14267
• Add Maintaining Your Computer Vision Models Docs Page by @abirami-vina in #14304
• allow_empty=True for Classify dataset class directories by @Alexis-IMBERT in #14301
• Fix Annotator PIL Image size (width, height) order by @SheffeyG in #14227
• ultralytics 8.2.52 fix CenterCrop transforms for PIL Image inputs by @glenn-jocher in #14308

New Contributors

• @wh1t3h47 made their first contribution in #14256
• @Alejandro-Casanova made their first contribution in #14267
• @Alexis-IMBERT made their first contribution in #14301
• @SheffeyG made their first contribution in #14227

Full Changelog: v8.2.51...v8.2.52