Commands¶
Run commands from the repository root after installation. CLI workflows use local data and checkpoint files; Hugging Face access is a separate download step.
Train and export¶
uv run geo2wf-train --help
uv run python scripts/export_geo_sar_geotiffs.py --help
uv run python scripts/export_unet_intensity_cache.py --help
uv run python scripts/export_joint_intensity_cache.py --help
uv run python scripts/export_intensity_forecast_cache.py --help
See training for Hydra overrides and resume flags, model presets for experiment choices, and GEO–SAR export for source requirements.
Evaluate local checkpoints¶
uv run python scripts/evaluate_intensity_models.py --help
uv run python scripts/evaluate_intensity_correction.py --help
uv run python scripts/evaluate_intensity_forecast.py --help
The script help lists workflow-specific arguments; the installed CLI wrappers provide the workflow selector. The comparison evaluator works with field, correction, and joint-model artifacts; it does not rerun the removed experiment orchestration scripts. Metric meanings are documented under evaluation.
Infer over source observations¶
The field workflow reads an observation manifest and source archive:
uv run geo2wf-infer deterministic-residual \
--config /path/to/resolved-config.yaml \
--checkpoint /path/to/model.ckpt \
--stats /path/to/paired/stats.json \
--data-root /path/to/source-archive \
--manifest /path/to/observation_manifest.csv \
--reference-root /path/to/reference-series \
--ibtracs-file /path/to/ibtracs.ALL.list.v04r01.csv \
--output-root inference/my-run
The reference root must contain <storm_id>/inference-summary.csv files; their observation IDs define which source observations are processed. --storms and --limit restrict that selection. Scalar workflows consume their corresponding local caches. For all inference options:
uv run python scripts/run_storm_unet_inference.py --help
uv run python scripts/run_intensity_correction_inference.py --help
uv run python scripts/run_intensity_forecast_inference.py --help