Skip to content

v1.3.0

Latest

Choose a tag to compare

@github-actions github-actions released this 12 Nov 15:55
· 88 commits to main since this release
6b3d5cd

Changelog

Added

  • Annotations

    • A new "Annotation classes" tab in the settings to add, find, rename, recolor and delete the annotation classes of a dataset.
    • Annotation class colors are saved in the dataset and shared by all users of it.
  • Filtering

    • Filter samples by evaluation run and per-sample metrics in the GUI query filter using sample_evaluation(...).
    • Type exact minimum and maximum filter values alongside the sliders for numeric metadata, image and video dimensions, video frame numbers, FPS, and duration.
    • Typing (, [, " or ' in the query editor wraps the selected text in that pair.
  • Evaluation

    • Python SDK: Compute per-class and mean average precision for object detection with compute_average_precision=True in ObjectDetectionEvaluationConfig. Read the results with dataset.evaluate().metrics(run_id), averaged over IoU thresholds 0.50 to 0.95 and at each threshold.
  • Embedding plot

    • The embedding plot shows the active plot filter and lets you clear it right on the plot.
  • Other

    • Python SDK: Added Python 3.15 support.

Changed

  • Annotations

    • The class for new annotations is remembered per collection across browser sessions and follows class renames and deletions. Selecting an existing annotation no longer changes it.
    • Users with the labeler or viewer role can only choose existing annotation classes in the labeling tools.
    • BREAKING: Annotation classes have new database fields for their custom color. DuckDB databases created with an earlier version are not upgraded and stop working. Re-index your data. Enterprise users are not affected.
  • Embedding plot

    • Embedding plots color faster when no filters are active and when coloring image and crop embeddings by annotation class.
  • Other

    • Python SDK: BREAKING: Upgraded scikit-learn to 1.9.1 on Python 3.11 and newer. Retrain saved random-forest classifiers created with scikit-learn 1.7.2 before loading them in the upgraded environment.
    • Python SDK: BREAKING: multi_strategies accepts only the built-in sampling strategies, and SamplingStrategy is removed. Type strategy lists as Sequence[Strategy] from lightly_studio.sampling.sampling_config.

Fixed

  • Annotations

    • Renaming an annotation class that has annotations no longer fails on DuckDB. If your DuckDB database was created before this version, re-index your data to get the fix.
    • Smart select stays active when you select a segmentation mask or change its class in the sample details sidebar.
    • Zooming with the mouse wheel or trackpad works while the Smart select tool is active.
    • Parts of a segmentation mask stroke that land on locked annotations no longer trim other annotations' masks.
  • Filtering

    • The browser no longer freezes when a collection has an integer metadata field with a very wide range, such as an ID.
    • The value counts of categorical metadata now match the filters of the images grid.

What's Changed

Full Changelog: v1.2.1...v1.3.0

Released by @lukas-lightly