Repository navigation
What's new
Fixes
- Packaging: The 2.0.0 wheel and sdist contained only
aitlas/__init__.pyandaitlas/run.py, so importing any subpackage failed afterpip install aitlas. All subpackages are now included (#26). - Dependencies: Capped setuptools below 82. setuptools 82 removed
pkg_resources, whichfs(a dependency ofeo-learn) still imports, soimport aitlas.datasetsfailed in fresh installs. - Change detection evaluation: Resized ground truth masks are now binarised at 0.5 before scoring. Previously they were cast to integers, which removed every partially covered border pixel from both classes.
- Change detection precision: Training with
automatic_mixed_precisionnow uses bfloat16 where the GPU supports it and float32 otherwise, and evaluation always runs in float32. float16 could overflow in some models (e.g., the attention in CGNet) and produce NaN losses. - Checkpoint selection: An epoch with a NaN validation loss is never kept as the best model, and a validation loss of exactly 0 is no longer treated as missing.
Other changes
- Package check: Added a CI workflow that builds the sdist and wheel and verifies that they contain the complete package.
- README: The installation check now imports
aitlas.base. - Tests: Fixed the change detection unit tests, which failed since AMP was added, and added tests for the fixes above.
Behavior changes
- Change detection scores change for datasets whose masks are resized to the model input size.
- Change detection models trained with
automatic_mixed_precisionnow use bfloat16 (or float32) instead of float16.
Notes
- 2.0.1 was not published separately. Its changes are included in this release.
- 2.0.0 has been yanked on PyPI because of the missing subpackages. Use 2.0.2 or newer.
Known issues
- Importing
aitlas.datasetsshows apkg_resourcesdeprecation warning fromfs. Makingeo-learnoptional is planned for 2.1.0.
Full Changelog: v2.0.0...v2.0.2