3LC Python Package Version 2.18¶

2.18.0¶

Features¶

  • [15280] Added support for pose/keypoints, oriented bounding-boxes, as well as other 2D/3D geometry (points, lines, triangles)

    • Pose is supported for SuperGradients, Ultralytics YOLO, and PyTorch

    • Oriented bounding boxes are supported for Ultralytics YOLO and PyTorch

    • See Computer Vision for more details.

  • [15982] Made it possible to collect and reduce embeddings with the SuperGradients integration for any model with a backbone attribute, which applies to the YOLO-NAS family of models

  • [15980] Added support for recipe-based training in SuperGradients integration

Enhancements and Fixes¶

  • [15944] Avoid unnecessary sample type roundtrip

  • [15944] Add random_state to base reduction args, as it is available in both pacmap and umap

  • [15640] Fixed issues with relative paths in configurations by resolving to absolute paths

  • [16010] Fixed regression in TableFromCoco leading to slow processing by bypassing default values when serializing and iterating

  • [15989] Improve performance by not considering default value population when data is not None and not composite schema

  • [15991] Perform SuperGradients metrics collection in chunks to avoid out-of-memory issues

  • [15986] Made it possible to export segmentations and keypoints with COCOExporter