Skip to content

Model and training interfaces

Status: The ML layer is under redesign. Existing classes and configuration shapes may change.

Scope

The ML package connects geospatial samples to PyTorch and PyTorch Lightning. It owns model contracts, encoders, decoders, heads, tasks, transforms, and training-time callbacks.

Design areas still open

  • GeoStack and GeoSample tensor conversion.
  • Dataset and datastore responsibilities.
  • Model context and stage composition.
  • Task and DataModule interfaces.
  • Prediction stitching and output writing.
  • Stable configuration and registry shapes.

Training examples and configuration reference will be written after the Spatial and ML interfaces settle.