Start here¶

The supported workflow is an installable geo2wf package with a composed training configuration and thin export, evaluation, and inference commands.

flowchart LR
  A[Source manifest] --> B[geo2wf-export]
  B --> C[WindFieldBatch + DataSpec]
  C --> D[geo2wf-train]
  D --> E[CheckpointLoader]
  E --> F[geo2wf-evaluate / geo2wf-infer]

Recommended reading order¶

1. Install and verify the package.
2. Run the first smoke experiment.
3. Learn how config groups and overrides compose.
4. Read the two-stage scientific workflow.
5. Use the command reference for evaluation and inference.

Main runtime pieces¶

geo2wf-train

Composes data, model, trainer, logging, and optional experiment groups; validates the model against the dataset DataSpec; creates the run directory; and starts Lightning.

PairedDataModule

Builds datasets and loaders from split manifests. Its canonical collator stacks tensors while keeping metadata sample-oriented.

WindFieldLightningModule

Defines the common training-objective and physical-prediction extension points used by modular models.

CheckpointLoader and PredictionService

Strict-load old or new checkpoints and expose deterministic and ensemble predictions through one physical-unit PredictionBatch.

Choose a route¶

• To prove the installation and configuration, follow First experiment.
• To train the main stack, use the Stage 1 → Stage 2 sequence.
• To add a component, follow Adding models, datasets, and metrics.
• To understand package ownership, read Modular package architecture.