Skip to content

Get started

Use Python 3.10 or 3.11 and uv. From the repository root:

uv sync --frozen

This installs the locked dependencies and the geo2wf-train, geo2wf-export, geo2wf-evaluate, and geo2wf-infer commands.

Get data and model files

Download without logging in, starting with the small metadata files:

uv tool install huggingface_hub
python3 scripts/download_artifacts.py
python3 scripts/download_artifacts.py all --dry-run
python3 scripts/download_artifacts.py all

The script saves pinned releases under downloads/data/ and downloads/models/. Use data or models instead of all to fetch one release, and --output-dir /path/to/storage to choose another destination. The full download includes about 32.9 GB of scientific assets and 1.27 GB of checkpoint weights, plus metadata and source files.

Read the dataset guide for selective downloads and using the matching source archive for reproduction. Current training in this checkout consumes a local raster export or task cache; the Hub catalog uses the loaders in the released source archive.

Once data are available locally, run the first experiment or select a model preset. Inference requires a compatible checkpoint, its resolved configuration, and the original normalization statistics.

Local paths and logging

Pass data paths explicitly in commands. Optional machine defaults can be kept in .local.env, copied from .local.example.env. TCD_DATA_ROOT selects the source archive for export/inference; WANDB_DISABLED=true disables W&B while keeping local CSV metrics and run records. WANDB_MODE=offline records W&B artifacts locally.

Development and docs

uv sync --frozen --group dev --group docs
uv run python -m pytest
uv run mkdocs build --strict
uv run python scripts/check_site_links.py
uv run mkdocs serve

The preview opens at http://127.0.0.1:8000. GPU training requires a compatible PyTorch/CUDA installation and allocated devices; use the trainer overrides for your machine.