Version 3.4.0
Major:
- Allow to train with large Zarr/H5 files without loading data in memory
- Adapt by_chunks to multichannel data
- Add option to reuse predictions in test
Minor:
- Calculate detection metrics in by_chunks setting
- Add DATA.NORMALIZATION.CUSTOM_MODE option
- Add percentile normalization
Fix:
- Correct bug when discading samples from training when using TRAIN.MINIMUM_FOREGROUND_PER
- Fix PSNR calculation in I2I and SR workflows
- Update notebooks and correct errors related to gdown package
- Correct validation paths in generators when in memory is False
Full Changelog: v3.3.14...v3.4.0