Inspect and consume a dataset

rfgen inspect is the supported read-only command for a stored dataset. Run it after generation before relying on the output in a downstream experiment.

rfgen inspect ./rfgen-output

Use --sample-size N when you want the command to include a bounded sample of records in its report:

rfgen inspect ./rfgen-output --sample-size 3

For Signal Dataset, the report contains record_count, a field inventory, and the first at most N record IDs in ordinal order. The inventory reports each field’s observed shapes, dtypes, and axes using metadata-only reads; inspection does not fetch tensor payloads. This bounded prefix is not a random or representative sample.

The native report carries a sample_ids field containing the same bounded ordinal prefix. --sample-ids prints only those IDs, one per line:

rfgen inspect ./rfgen-output --sample-size 8 --sample-ids

The command does not repair a store, calculate a distribution dashboard, or certify a dataset for deployment. For a known baseline, compare what you see with that path’s qualification boundary.

To consume native fields in PyTorch:

import torch

import signal_dataset as sd

dataset = sd.open("./rfgen-output")
record = dataset[0]
iq = torch.from_numpy(record["projections/receiver/receivers/rx0/iq"].data.copy())

Field names are namespaced by the projection that produced them; rfgen inspect prints the full list. Tensor conversion, batching, and split policy belong to the training application. For metadata-only scans, use dataset.iter_record_metadata() instead of indexing full records: it yields RecordMetadata without loading any field payload.