- Fixed MPS patch failing for custom models with non-default input channel counts
- MPS compatibility check now gracefully handles unexpected errors instead of crashing
- Tests for MPS patch across smp and fastai models with various backbones and channel counts
- New v4 model weights with improved performance
- Optional legacy dependency group for backward compatibility with older models
- Migrated default model backend from fastai to segmentation-models-pytorch (smp)
- Moved fastai to optional
legacydependency group for continued support of older models
- Added cache to load_model_from_weights for faster model loading
- Modified patch compilation to use a queue to avoid overloading to cpu
- Fixed nodata bug in get_patch
- New model weights (v3) trained on larger dataset
- Fixed broken links in README for PyPI compatibility
- GDAL style (255-0) nodata mask for geotiff exports when using
apply_no_data_mask=True
- Default model download destination now uses
platformdirs.user_data_dir - Exported geotiff metadata nodata value changed from 0 to None
- torch.compile model compilation support
- New model release with improved speed and robustness across different resolutions
- Initial release