Narrowband classifier baseline¶
Purpose¶
Build a small synthetic dataset for a modulation-classification experiment. Each record contains complex baseband IQ, a structured class label, and the provenance needed to understand how it was generated.
Audience and prerequisites¶
Use this path for a first classifier-shaped dataset or as a stable baseline for
an editable experiment. Install rfgen[torchsig,sionna] (or, from a source checkout,
uv pip install -e '.[torchsig,sionna]') because this template selects a TorchSig
emitter and the 3GPP TR 38.901 Urban Microcell (UMi) propagation
model, an urban-microcell propagation scenario. The Quickstart
walks through the same commands and explains the resulting artifacts.
Inputs¶
The shipped template uses a deterministic seed and four labels: BPSK, QPSK, 16-QAM, and 64-QAM. It is an ordinary YAML file after materialization; it is not a scenario catalog and editing it creates your own experiment. The communication waveforms are built with TorchSig’s constellation-signal primitive; the exact selected class remains stored as rfgen metadata.
What the records carry¶
Each sample publishes one record: the receiver capture, one declared box per
emitter, and a segmentation raster. Label
schema states the frames those labels
are in — box edges are offsets from the receiver’s tuned centre, and the raster
is not a torch.stft grid.
Run¶
rfgen init narrowband-baseline ./narrowband-config
cd ./narrowband-config
rfgen validate --config-dir .
rfgen generate --config-dir .
rfgen inspect ./rfgen-output
Observable output and interpretation¶
The template writes 100 records to the local Signal Dataset snapshot
rfgen-output under the configuration directory. Every record is a synthetic single-emitter narrowband example whose
label is one of the four named modulation classes. The dataset also retains
metadata, its RX profile (the
receiver-side capture and hardware configuration/effects recorded with the
dataset), and Sionna UMi provenance; it is not a
directory of unlabeled waveform files.
Verification boundary¶
The installed-wheel qualification checks the stated record count and label set, store readability, and repeatability for the fixed seed and environment. It does not assert class balance, broad modulation coverage, channel realism, classifier accuracy, or equivalence to over-the-air data.
Customize safely¶
Copy the materialized configuration to a new directory before changing it.
After every edit, run rfgen validate --config-dir DIRECTORY, generate into a
new output directory, and record the resulting configuration alongside your
training results. For terminology and the data model, read the
RF & ML Quick Overview.