Code map¶
| Location | Responsibility |
|---|---|
src/geo2wf/data/ | Raster/cache loading, contracts, normalization, features, sampling, and augmentation |
src/geo2wf/models/ | Field, joint/encoder-only, correction, and forecast models |
src/geo2wf/config/, training.py | Hydra composition, compatibility checks, and Lightning training |
src/geo2wf/metrics/, evaluation/ | Physical errors, storm structure, and evaluation adapters |
src/geo2wf/visualization/, tracking/ | Figures, CSV/W&B logging, and run provenance |
src/geo2wf/cli/, scripts/ | Retained export, inference, evaluation, and website workflows |
configs/ | Grouped training presets and the export-only config.yaml |
tests/ | Data/model contracts, numerical behavior, integration, and dashboard checks |
docs/ | These guides, paper results, and the StormSense frontend |
The root train.py and compatibility modules under src/ preserve existing imports/checkpoints. New model and loader work belongs under src/geo2wf/. Data/model configs name their _target_; training composes and validates them without a central architecture registry.
Large scientific inputs and model releases belong in the linked Hugging Face repositories. StormSense's external ViT and ConvLSTM comparison outputs do not imply maintained training implementations.
Use commands for entry points, training for configuration, and troubleshooting for common failures.