Announcing Ultralytics Release v8.3.85!
We’re excited to unveil Ultralytics v8.3.85
, a release geared towards enhancing export and inference workflows. This update includes critical fixes for TensorRT exports and substantial improvements to ONNX segmentation examples, making them more efficient, usable, and robust. Dive into the details below!
Summary
This release focuses on:
- TensorRT export fixes for smoother dynamic model handling.
- Optimized ONNX segmentation examples with streamlined setup.
If you’re working on model exports or leveraging ONNX/TensorRT for inference, this update is for you!
Key Changes
TensorRT Updates
- Resolved
max_shape
Calculation Bug: Fixed inconsistent calculations during TensorRT model export with non-zero workspace values, ensuring precise shape handling. - Improved Default Workspace Behavior: When workspace settings are not specified, they now correctly default to
0
, simplifying the export process.
ONNX Segmentation Example Enhancements
- Simplified Preprocessing/Postprocessing:
- Consolidated key parameters (
iou
,imgsz
, andconf
) for improved readability and ease of modification. - Streamlined function calls for faster and cleaner code execution.
- Consolidated key parameters (
- Unified Threshold Naming:
- Adjusted
iou
andconf
to follow YOLO conventions, promoting consistency across examples.
- Adjusted
- Optimized Segmentation Masks:
- Boosted mask accuracy and computational efficiency to provide reliable results for object segmentation tasks.
- Automated GPU Integration:
- Integrated GPU support automatically where available, reducing manual setup effort for backend processing.
Purpose & Impact
-
For TensorRT Users:
- Purpose: Fix TensorRT export bugs to ensure dynamic model compatibility and reduce export crashes.
- Impact: Enhanced stability and accuracy for
.engine
format exports, providing an optimal inference experience.
-
For ONNX Developers:
- Purpose: Simplify segmentation workflows and boost efficiency for ONNX Runtime users.
- Impact: Easier setup with reliable and intuitive tools for mask-based object segmentation.
-
Overall Benefits:
Improved model export and refined inference logic empower developers to seamlessly transition between development and deployment phases with confidence.
What’s Changed
-
Cleanup and Fix ONNX Segmentation Examples
Contribution by @Y-T-G.
PR: ONNX Segmentation Fix. -
TensorRT Export
max_shape
Bug Fix
Contribution by @Y-T-G.
PR: TensorRT Bug Fix.
Full Changelog: See all changes for v8.3.85.
Try It Out and Share Feedback
We encourage you to update to v8.3.85
today and explore these new improvements! Check out the release page for v8.3.85 for further details and installation instructions.
Your feedback is invaluable to us—let us know your thoughts, suggestions, or any challenges you face in the Ultralytics GitHub Discussions or by opening an Issue.
Thank you for being part of the YOLO and Ultralytics community! These updates would not be possible without your support and contributions.