This release streamlines Datumaro by removing a number of lesser-used features, helping to simplify the tool and reduce its dependencies. These changes are part of an effort to keep Datumaro focused on its core strengths: dataset management and integration with machine learning frameworks. As part of this update, inference-related features have been removed. For inference tasks, we recommend using the OpenVINO model API. If you rely on a specific feature that is no longer available, you can still access it from the previous version of Datumaro.
Removed features
- CLI commmands:
- API features:
- Model inference
(#1831, #1825) - Model-based transformations
(#1826, #1839) - Crypter
(#1829) - Synthetic dataset generation
(#1815) - Data exploration
(#1814) - BBox to mask using SAM
(#1826) - Telemetry
(#1828) - Anchor generation
(#1832) - Missing annotation detection
(#1826) - Model inference explanation
(#1812) - Near-duplicate removal
(#1835) - Pruning
(#1813) - Pseudo-labels
(#1814) - Project
(#1824) - Noisy label detection
(#1833) - Data shift analysis
(#1827, #1892)
- Model inference
- SAM Docker image
(#1830)
New features
- Experimental dataset class
(#1807, #1810, #1811, #1834, #1858, #1845, #1863, #1868, #1876, #1877, #1879, #1881, #1891) - New OpenVino Accuracy Checker semantic segmentation format
(#1893)
Enhancements
- Mark several dependencies as optional
(#1849, #1862) - Removal of unneeded dependencies
(#1837) - Documentation tidy-up
(#1840) - DCO introduction; readme, PR template, and contribution guide tidy-up
(#1844, #1860, #1856, #1847) - Fix code coverage upload to Codecov in the CI
(#1861) - Fix crashes with certain datasets in the compare command
(#1892) - Added Semgrep security scan in the CI
(#1883)