Skip to content

Release v1.13.0

Choose a tag to compare

View all tags
@AlbertvanHouten AlbertvanHouten released this 22 May 11:11
· 10 commits to develop since this release
v1.13.0
66832fa
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Release 1.13.0

This release is centered around the maturation of the experimental module introduced in 1.12.0, while also delivering improvements to core Datumaro, dependency management, CI, and security. The changes below are split between the experimental module and the rest of Datumaro.

Experimental module

New features

  • Dataset format support
    • YOLO dataset import/export
    • Pascal VOC import/export, including classification-only datasets
    • Refactored COCO dataset import/export interface
    • Importing legacy Datumaro datasets, including datasets with videos
    • Conversion of legacy anomaly datasets
    • Automatic dataset format detection and import
    • Automatic sample-type detection on data import
    • Detect Roboflow COCO layout
  • Media and image
    • Video media support
    • 16-bit and 32-bit image support
    • EXIF orientation honored in image loading and processing
    • Support for mask callables
    • MediaBridgeConverter for media field type conversions
    • LazyImage loading from ImagePathField
  • Dataset and transforms
    • Tiling support, with picklable transform factories
    • Keypoint support, including KeypointCategories
    • Multi-label and hierarchical-label classification support
    • Saving datasets as zip archives
    • append_dataset method
    • Dataset filtering on subset
    • Label filtering on the dataset class, with option to keep empty samples
    • Multi-worker dataloader support
    • Lazy-loading during attribute access
    • Dataset export/import implementation
  • Schema and type system
    • New fields: ScoreField, EllipseField, StringField
    • Type converters for NumericField, BoolField, StringField, and a bool Polars converter
    • Optional field handling and optional attributes in schema conversions
    • Typed numpy array support
    • Categories validation and propagation across conversions
    • Conversion between categories
    • Sample and dtype validation
    • Legacy mapper for classification
    • Uppercase subset enums

Enhancements

  • Batch dataset building optimized with PyArrow integration
  • Optimized image loading and channel swapping
  • Improved dtype conversions and RGB/BGR color converter
  • Reimplemented label shape converter to support multi-label / is-list swaps
  • Extended label converter to handle both multi-label and is-list
  • Allow label list conversions
  • Added several converters and fixed an out-of-sync dataframe
  • Extended field converters and enhanced cost calculation for direct field-type converters
  • Added a direct_only option to dataset import/export functions
  • Extended automatic schema detection checks
  • Cross-platform filename sanitization for dataset imports
  • Prevent file overwrite failure when exporting to an existing directory
  • Support for missing images in dataset export
  • Ensure new Datumaro datasets are picklable
  • Enhanced conversion error handling
  • Restructured experimental tests directories and reorganized files in the experimental folder
  • Documentation improvements and a README for the experimental module

Core Datumaro

New features

  • format parameter added to the datum patch command
  • Include empty annotation files for YOLO export
  • Export segment info attributes in COCO panoptic

Enhancements

  • Added support for Python 3.14 and restored Python 3.10 support
  • Migrated dependency management to uv
  • Switched linting to prek and ruff with an updated rule set
  • Simplify original_size check in CVAT base
  • Documentation workflow fixes and renamed AttributeInfo attribute

Bug fixes

  • Fix RotatedBbox.wrap() by assigning correct attributes

CI / tooling

  • Enabled Renovate and removed Dependabot configuration
  • Added auto-approval flow for Renovate uv.lock refresh PRs
  • Refactored GitHub Action workflows
  • Moved linter and docs workflows to a dedicated runner group
  • Added harden-runner step for audit traffic and enabled traffic blocking in workflows
  • Security workflow consolidation and GitHub Actions upgrades
  • Added Collect Library Licenses workflow

Security

  • Pin keras to >=3.13.2
  • Pin jupyterlab>=4.5.7
  • Update Pillow to 12.1.1
  • Update nbconvert to 7.17.0

Removed

  • pytest-csv dependency
Assets 2
Loading