v0.4.14 1.9.2
roboflow/rf-detrAug 10, 2026 → Aug 11, 20261 day between releases
What's New
Major version jump introduces breaking COCO label mapping changes and significant HungarianMatcher performance gains
Breaking Changes
- Hierarchical COCO datasets now use a correct shared label mapping, removing unannotated grouping categories from label slots.
- Unannotated grouping categories no longer consume a label slot, potentially affecting metrics on hierarchical datasets.
New Features
- Training resume from lightweight checkpoints now restores per-callback state instead of restarting cold.
- HungarianMatcher optimized for faster speed and lower memory usage.
Improvements
- HungarianMatcher is approximately 51% faster and uses 73-76% less peak CUDA memory.
- Training and evaluation workflows include several bug fixes and refinements.
Bug Fixes
- Fixed training resume behavior to correctly restore per-callback state.
- Resolved various training and evaluation issues.
Upgrade Notes
Upgrading requires attention to hierarchical COCO datasets as label mapping logic changed to exclude unannotated grouping categories. Ensure training scripts handle this shift to avoid metric discrepancies. The version jump also brings significant performance improvements to the HungarianMatcher.
The Two Releases
Release v0.4.14
Aug 10, 2026· 2 months ago
A maintenance release focused on resolving installation compatibility issues with the `uv` package manager.
Full release notesv1.9.2: Fixed Roboflow COCO & perf. optim
Aug 11, 2026· 2 months ago
A fix-and-performance release introducing a breaking change for hierarchical COCO datasets. It significantly speeds up the HungarianMatcher and restores callback state during training resume.
Full release notes