Datasets#

A dataset expands one or more geographic scenes into fixed or seeded-random radio maps. Scene geometry is compiled once per scene and shared; each radio map stores only its own radio products and metadata. The same generator runs from Python, from the batch web page, and from a cluster job.

Planning and generation

Request schema, plan inspection, seeded expansion, buildings on flat ground, stored arrays and previews.

Planning and generation
Dataset layout

Manifest, scene directories, radio map directories, the stored arrays, loading them and the Parquet index for finding them.

Dataset layout
Resume and extend

Interruption recovery, deterministic extension, the compatibility checks.

Resume and extend
Packaging for release

One zstd-compressed archive per scene, splits by scene group, licence checks, checksums.

Packaging for release
PyTorch

A PyTorch dataset of whole radio maps or single transmitters, selection by index, random crops for batching.

PyTorch
Cluster jobs

The Slurm example for a GPU node.

Cluster jobs
Tutorial: viewing a package

A notebook that finds, plots, views in 3D and solves again the radio maps of a package, and compares one with the same radio map on flat ground.

Viewing the radio maps of an OWRT-RF package
Tutorial: PyTorch

A notebook that reads a package with RadioMapDataset and trains a small network on crops.

Radio maps as a PyTorch dataset