Releases: ultralytics/ultralytics
v8.2.74 - `ultralytics 8.2.74` add `fuse_score=True` BoT-SORT and ByteTrack arg (#14965)
🌟 Summary
Release v8.2.74 introduces enhancements for NVIDIA Jetson support, improved export options, and tracking updates.
📊 Key Changes
- GitHub Actions: Updated to include additional status code (
401
). - Documentation Updates:
- Expanded NVIDIA Jetson support, now including JetPack 6.0.
- Improved OpenVINO export options, adding support for dynamic input sizes.
- Added a contributor
mfloto
.
- Trackers Configuration:
- Added
fuse_score
option to BoT-SORT and ByteTrack trackers.
- Added
- Code Adjustments: Conditionally fuse confidence scores with IoU distances in trackers.
🎯 Purpose & Impact
- Security and Reliability:
- The GitHub Actions update makes the system more robust by handling
401 Unauthorized
statuses 💪.
- The GitHub Actions update makes the system more robust by handling
- Enhanced Hardware Support:
- Documentation on JetPack 6.0 expands the usability of YOLOv8 across more NVIDIA Jetson devices, providing detailed setup instructions 🚀.
- Export Flexibility:
- The dynamic input size support in OpenVINO exports enhances the model's flexibility and compatibility 🎨.
- Tracking Improvements:
- The
fuse_score
option in trackers improves the accuracy and reliability of object tracking by leveraging both confidence and IoU metrics 🎯.
- The
What's Changed
- Ignore Vimeo 401 'unauthorized' errors by @glenn-jocher in #14980
- fix example for plotting Ray Tune history by @mfloto in #14970
- Update NVIDIA Jetson Docs with JetPack 6 by @lakshanthad in #14939
- Fix OpenVINO Export Docs by @ambitious-octopus in #14918
ultralytics 8.2.74
addfuse_score=True
BoT-SORT and ByteTrack arg by @Laughing-q in #14965
New Contributors
Full Changelog: v8.2.73...v8.2.74
v8.2.73 - `ultralytics 8.2.73` Meta SAM2 Refactor (#14867)
🌟 Summary
The v8.2.73 release of Ultralytics introduces significant updates to the Segment Anything Model (SAM) architecture, including new support for SAM 2, which enhances real-time image segmentation capabilities and documentation improvements.
📊 Key Changes
- Addition of SAM 2 Models: Introduced new methods for building various SAM 2 models, including
build_sam2_t
,build_sam2_s
,build_sam2_b
, andbuild_sam2_l
. - Enhanced Documentation: Added comprehensive documentation for SAM and SAM 2 modules, including blocks, decoders, encoders, and memory attention modules.
- Updated Predictors: Integrated SAM2Predictor for advanced segmentation prediction, enhancing the overall prediction framework.
- Expanded API References: Updated the API references to include new SAM2 modules and their functionalities, ensuring better clarity and usability for developers.
🎯 Purpose & Impact
- Improved Segmentation Capabilities: The addition of SAM 2 models significantly enhances the real-time image segmentation capabilities, allowing users to leverage advanced features for more accurate and efficient segmentation tasks.
- Comprehensive Documentation: The improved and expanded documentation provides detailed references and examples for both SAM and SAM 2 models, making it easier for users to understand and implement the models in their projects.
- Upgraded Prediction Framework: The integration of SAM2Predictor ensures that the prediction framework is robust and capable of handling advanced segmentation tasks, benefiting users with more reliable and efficient predictions.
🔧 Technical Details
- Model Initialization: Enhanced the initialization and building process for SAM and SAM 2 models, ensuring that the models are correctly configured with the specified architecture parameters.
- Attention Mechanisms: Improved attention mechanisms with the introduction of SAM2TwoWayAttentionBlock and SAM2TwoWayTransformer, providing more efficient attention computations in the models.
What's Changed
ultralytics 8.2.73
Meta SAM2 Refactor by @Laughing-q in #14867
Full Changelog: v8.2.72...v8.2.73
v8.2.72 - `ultralytics 8.2.72` SAM 2 multiple-`bboxes` support (#14928)
🌟 Summary
Ultralytics 'v8.2.72' release focuses on enhancing disk space management, updating documentation, refining model code, and improving testing.
📊 Key Changes
- Disk Space Management: Added a step in CI workflows to clean up tool cache, potentially freeing up to 10GB of disk space.
- Documentation:
- Improved
CONTRIBUTING.md
with better references and added an FAQ section. - Updated ROS Quickstart guide for better clarity and specific use cases.
- Improved
- Model Enhancements:
- Adjustments in SAM prediction models for better handling of bounding boxes and points.
- Refined Transformer model's
forward
method.
- Testing Improvements: Added extra checks for Torch compatibility in tests.
🎯 Purpose & Impact
- Disk Space Management:
- Purpose: Ensure workflows run smoothly without disk space issues.
- Impact: Enhanced reliability in CI pipelines, preventing build failures due to insufficient disk space. 🧹💾
- Documentation:
- Purpose: Provide clearer guidance and support to contributors.
- Impact: Makes contributing easier and more accessible, fostering community contributions. 📖👥
- Model Enhancements:
- Purpose: Optimize model predictions and enhance functionality.
- Impact: Improves performance and accuracy of object detection and other tasks. 📈🤖
- Testing Improvements:
- Purpose: Ensure better compatibility and comprehensive testing.
- Impact: Increases robustness and reliability of the software across different environments. ✅🧪
These updates collectively aim to improve the project's maintainability, user experience, and performance.
What's Changed
- Delete Docker Cache before building Image by @lakshanthad in #14894
- ROS quickstart meta description by @ambitious-octopus in #14932
- Fix the docstring of xywhr2xyxyxyxy by @patel-zeel in #14934
- Fixed
circle_label
andtext_label
docstrings by @jk4e in #14909 - Fix
is_url()
andcheck_disk_space()
docstrings in downloads.py by @jk4e in #14923 - Update MLP module for RTDETR backward compatibility by @Laughing-q in #14901
- Corrects CONTRIBUTING.md relative link to CLA.md by @Burhan-Q in #14857
ultralytics 8.2.72
SAM 2 multiple-bboxes
support by @Laughing-q in #14928
New Contributors
- @patel-zeel made their first contribution in #14934
Full Changelog: v8.2.71...v8.2.72
v8.2.71 - `ultralytics 8.2.71` Multinode DDP training (#14879)
🌟 Summary
The Ultralytics v8.2.71
release focuses on refining model training commands, enhancing the CLI, and fixing some documentation inconsistencies.
📊 Key Changes
- Matrix Parallelism: Increased max-parallel jobs from 6 to 10.
- Retry Logic Modification: Adjusted retry wait time from 30 to 60 seconds.
- Documentation Fixes: Corrected CLI commands for training (
yolo classify train
,yolo segment train
,yolo obb train
,yolo pose train
) and removed some outdated notes on SAM 2. - Gitignore Update: Added
requirements.txt
,setup.py
, and other files to the.gitignore
list. - CLI Enhancements: Updated CLI task options to include
obb
and added a new modebenchmark
.
🎯 Purpose & Impact
- Enhanced Parallelism: Increasing max parallel jobs accelerates CI workflows, thereby improving overall efficiency and quicker feedback loops. 🚀
- Improved Job Reliability: With longer wait times before retrying jobs, the update aims to reduce the number of job failures, making the process more robust. 🛠️
- Command Accuracy: The corrected training commands ensure that users are employing the right commands, reducing the chances of errors and improving user experience. ✅
- Refined CLI Use: The added
obb
task option andbenchmark
mode provide users with more versatility and capabilities within the CLI, enabling better performance analysis and use-case flexibility. 💡
v8.2.70 - `ultralytics 8.2.70` Segment Anything Model 2 (SAM 2) (#14813)
🌟 Summary
The v8.2.70
release of Ultralytics brings valuable enhancements, such as integrating Black
code formatter and adding substantial documentation, especially around SAM2 (Segment Anything Model 2).
📊 Key Changes
- Workflow Adjustment: The GitHub Actions workflow now includes the installation of
Black
for consistent code formatting. - Documentation Expansion: Significant additions to the documentation, including detailed guides and FAQ sections for various integrations and models.
- New Model Support: Introduction of support for SAM2, an enhanced model for comprehensive object segmentation in images and videos.
🎯 Purpose & Impact
- Consistency in Code Formatting: Adding
Black
to the GitHub Actions workflow ensures that the code base maintains consistent formatting standards, reducing errors and improving readability. - Enhanced Documentation: New and updated documentation provides users with better guidance and understanding of using various models and integrations. Non-expert users can benefit from the clear explanations and examples, while expert developers can delve into the detailed technical aspects.
- Advanced Model Capabilities: Supporting SAM2 broadens the capability of the Ultralyics framework, especially for complex image and video segmentation tasks. The improvements and new features in SAM2, such as real-time performance and zero-shot generalization, can significantly impact applications in fields requiring precise and efficient object segmentation.
This release improves the user experience with better documentation and more powerful tools for both developers and end-users. 🚀
What's Changed
- Update comet from init() to login() by @glenn-jocher in #14793
- New Meta Segment Anything Model 2 (SAM2) Docs page by @glenn-jocher in #14794
- SAM2 mkdocs.yml fix by @glenn-jocher in #14796
py-cpuinfo
Exception context manager fix by @glenn-jocher in #14814- Add https://youtu.be/_gRqR-miFPE & https://youtu.be/CfbHwPG01cE to docs by @RizwanMunawar in #14817
ultralytics 8.2.70
Segment Anything Model 2 (SAM 2) by @Laughing-q in #14813
Full Changelog: v8.2.69...v8.2.70
v8.2.69 - `ultralytics 8.2.69` FastSAM prompt inference refactor (#14724)
🌟 Summary
Release v8.2.69
brings improvements to issue templates, documentation, FastSAM’s prompt handling, and the introduction of IBM Watsonx and JupyterLab integrations.
📊 Key Changes
- Issue Templates Updates:
- Standardized language in bug report and question templates to consistently reference "Ultralytics YOLO".
- Documentation Enhancements:
- Detailed guides on integrating IBM Watsonx and JupyterLab for model training.
- Minor formatting and structural enhancements in various dataset and integration documentation.
- FastSAM Enhancements:
- Merged all prompting functionalities directly into
FastSAM
class, removing the need forFastSAMPrompt
. - Simplified and optimized the handling of multiple prompt types (bounding boxes, points, texts).
- Merged all prompting functionalities directly into
- Bug Fixes:
- Fixed class imbalance augmentation issue in data augmentation script.
- Formatting corrections in Google Cloud region latency output.
🎯 Purpose & Impact
- Improved Usability:
- Standardized templates help users provide more consistent and comprehensive bug reports and questions, facilitating faster and more accurate triaging.
- Enhanced Documentation:
- New IBM Watsonx and JupyterLab integration guides enable users to expand their training environments, offering more flexibility and productivity in model training and experimentation.
- Streamlined FastSAM Prompting:
- Integrating prompts into the FastSAM class simplifies usage, making it easier for users to apply multiple prompt types in a single call and improving model inference performance.
- Enhanced Augmentation and Debugging:
- Fixing the class imbalance augmentation issue ensures more reliable and balanced training data generation.
- Clearer formatting in latency outputs aids in the easier identification of the best Google Cloud regions for deployment.
These updates focus on enhancing user experience, expanding documentation, and improving model functionalities for more efficient and effective workflows. 🚀
What's Changed
- Add JupyterLab Integrations Docs Page by @abirami-vina in #14755
- Fixed typo and formatting OpenVINO integration by @ambitious-octopus in #14773
- Add IBM Watsonx Integrations Docs Page by @abirami-vina in #14785
- Update Ultralytics issue templates by @ambitious-octopus in #14718
- Eliminate
set()
to avoidset()
+list()
op by @QinghongLin in #14745 ultralytics 8.2.69
FastSAM prompt inference refactor by @Laughing-q in #14724
New Contributors
- @QinghongLin made their first contribution in #14745
Full Changelog: v8.2.68...v8.2.69
v8.2.68 - `ultralytics 8.2.68` new HUB GCP region latency tests (#14753)
🌟 Summary
Release v8.2.68 includes significant updates to Ultralytics' Docker setup and adds a new class for Google Cloud Platform (GCP) region analysis.
📊 Key Changes
- Dockerfile Updates:
- Standardized working directory to
/ultralytics
across all Dockerfiles. - Removed redundant environment variable
APP_HOME
. - Updated volume mounts to
/datasets
instead of/usr/src/datasets
.
- Standardized working directory to
- GCP Support:
- Added a new
GCPRegions
class inultralytics/hub/google/__init__.py
to check and analyze latency across GCP regions.
- Added a new
- Documentation:
- Added new reference docs for the
GCPRegions
class.
- Added new reference docs for the
🎯 Purpose & Impact
- Dockerfile Updates:
- Purpose: Simplify the Docker setup and ensure consistency across different Docker environments.
- Impact: Easier management and setup of Docker containers. This can lead to fewer errors and a smoother development experience.
- GCP Support:
- Purpose: Provide tools for analyzing and selecting the optimal GCP regions based on latency.
- Impact: Users deploying Ultralytics services on GCP can now optimize latency, which can enhance performance and user experience.
- Documentation:
- Purpose: Offer clear guidance and reference for the new
GCPRegions
functionality. - Impact: Improves user understanding and usability of the new features, making it easier for users to implement and benefit from them.
- Purpose: Offer clear guidance and reference for the new
What's Changed
- Simplify Dockerfile
WORKDIR
by @glenn-jocher in #14750 ultralytics 8.2.68
new HUB GCP region latency tests by @glenn-jocher in #14753
Full Changelog: v8.2.67...v8.2.68
v8.2.67 - `ultralytics 8.2.67` new NVIDIA Jetson Orin Jetpack 6 Docker image (#14740)
🌟 Summary
Ultralytics has released version 8.2.67 focused on workflow enhancements and better support for the latest Jetson platforms.
📊 Key Changes
- Workflow Drop: Removal of Windows CI testing.
- New Docker Support: Added Docker support for Jetson JetPack 6.
- Dockerfile Improvements: Enhanced installation process and dependency management for JetPack 4 and 5.
- Enhanced Settings: Updated configuration warnings and help messages.
🎯 Purpose & Impact
- Streamlined CI: By dropping Windows from the testing matrix, CI runs might become more efficient and focused on the most used OS environments.
- Expanded Platform Support: Added support for JetPack 6 dramatically improves Ultralytics' compatibility with the latest NVIDIA Jetson hardware, benefiting developers leveraging these platforms.
- Better Dockerfile Maintenance: Enhanced Dockerfile management and dependency handling ensure smoother and more reliable builds, particularly for JetPack 4 and 5 users.
- User-Friendly Updates: Clearer error messages and guidance within the settings ensure users can handle configurations more effectively, reducing potential for configuration errors.
Overall, this update enhances usability, support for cutting-edge hardware, and streamlines essential CI workflows. 🚀
What's Changed
SETTINGS
type and key checks by @glenn-jocher in #14703- Add JetPack6 Docker for NVIDIA Jetson Orin Series by @lakshanthad in #14707
ultralytics 8.2.67
new NVIDIA Jetson Orin Jetpack 6 Docker image by @glenn-jocher in #14740
Full Changelog: v8.2.66...v8.2.67
v8.2.66 - `ultralytics 8.2.66` HUB model autodownload (#14702)
🌟 Summary
Ultralytics' v8.2.66 update includes code optimizations, documentation improvements, and adjustments for better compatibility and functionality, particularly for ARM-based architectures 🛠️.
📊 Key Changes
- 📜 Adjusted Dockerfile for ARM-based architectures to simplify package installations.
- 📹 Embedded a YouTube tutorial in the SKU-110k dataset documentation.
- 🔍 Refined
check_file
function for better handling and downloading of files. - 🛠️ Minor code refactoring and removal of redundant functions.
🎯 Purpose & Impact
- Simplified Workflows: Streamlining Dockerfile operations reduces complexity, making it easier to set up on ARM-based systems, including Raspberry Pi users.
- Enhanced Learning Resources: Integrating video tutorials into documentation aids users in understanding how to use datasets effectively.
- Improved File Handling: Enhanced
check_file
function ensures more robust downloading and local file retrieval, which helps in maintaining code efficiency and reliability. - Cleaner Codebase: Removing unused functions and refactoring enhances code readability and maintainability, benefiting developers working with or contributing to the codebase.
What's Changed
- Remove duplicate
make_divisible
function by @Burhan-Q in #14690 - Add https://youtu.be/_gRqR-miFPE to docs by @RizwanMunawar in #14698
- Updates
save_period
to include first epoch by @Burhan-Q in #14700 - Add compatible
tensorstore
versions foraarch64
by @lakshanthad in #14697 ultralytics 8.2.66
HUB model autodownload by @glenn-jocher in #14702
Full Changelog: v8.2.65...v8.2.66
v8.2.65 - `ultralytics 8.2.65` fix YouTube throttling bug (#14684)
🌟 Summary
The 'v8.2.65' update for Ultralytics includes important changes such as the addition of a new activation function, bug fixes, and updates to caching logic.
📊 Key Changes
- New Activation Module: Introduction of the AGLU activation function within the neural network modules.
- Update to
cache_ram
Feature: Disabled due to a known memory leak issue. - Dependency Update for
pytubefix
: Now requires version 6.5.2 or later. - NMS (Non-Maximum Suppression) Logic Update: Improvements in handling single class and agnostic NMS.
- Documentation Enhancement: Addition of detailed documentation for activation functions.
🎯 Purpose & Impact
- Enhanced Flexibility in NN Customization: The new AGLU activation function allows for more sophisticated activation handling, potentially improving model performance on specific tasks.
- Operational Stability: Disabling
cache_ram
prevents potential training disruptions due to memory leaks, ensuring smoother classifier training sessions. - Improved Video Handling: Updating
pytubefix
enhances compatibility and fixes bugs related to YouTube video processing. - Better NMS Accuracy: The update to NMS logic will result in more accurate detections by properly handling edge cases with single class and agnostic settings.
- User-Friendly Documentation: Introductory guide and references for neural network modules help both new and existing users understand the tools available for model customization.
Overall, these updates aim to enhance both the functionality and reliability of the Ultralytics framework.
What's Changed
- New AGLU activation module by @glenn-jocher in #14644
- Fix
torch.amp
has no attributeGradScaler
by @Laughing-q in #14647 - Add warning when
cache_ram
works with classify by @Laughing-q in #14650 - Allow
agnostic_nms
option for validation by @Laughing-q in #14675 ultralytics 8.2.65
fix YouTube throttling bug by @glenn-jocher in #14684
Full Changelog: v8.2.64...v8.2.65