Repository navigation
Release v2.2.0 introduces a full geolocation stack, a config-based data module, and multi-label classification support, alongside a very large set of new signal families and builders, extensive test-coverage expansion, numerous bug fixes, and performance work across the transforms and dataset pipelines.
Features
- Geolocation support – Added a complete geolocation stack covering types,
propagation models, file handling, transforms, datasets, an example helper
script, and example notebooks. - Config-based TorchSig Lightning DataModule – A new declarative
TorchSigDataModulewithfrom_configconstruction and target-label support. - Multi-label classification – New Multi-HotLabel metadata transform and
generalized label-groupings beyond hard-coded families. - New signal families & builders – Structured builders and signal lists for
LoRa, GSM, 802.11a, Zigbee, Bluetooth LE, ADS-B, DVBS2, LMR, cellular, wifi,
and more, plus generalizations to existing builders. - New readers & utilities – Added
SIGMFReader,WAVReader, andOGGReader,
plus aMetadataReaderutility thatNPYReadernow inherits from. - Performance benchmarks – Optional benchmarks for transforms and for default
dataset generation, including dedicated Doppler benchmarks. - Build profiles – Introduced build profiles for TorchSig, including
documentation dependencies.
Fixes
- Fixed scalar Doppler correctness and input validation.
- Corrected fractional resampling rates near unity.
- Fixed a circular import surfaced by importing
TorchSigDefaults. - Fixed the narrowband-dataset passband-ripple design failure.
- Resolved
publish:pypiCI failures (libxcb.so.1) and a failing docs-build
job. - Fixed
DataLoadertest hangs (and orphaned workers) with multiple workers. - Fixed malformed notebook metadata.
- Fixed a typo that conflated
SigMFReaderandWAVReader. - Made the
sigmfdependency optional. - Added bounding-box corrections for multiple signal families (802.11a, Zigbee,
GFSK, ADS-B) and removed obsolete 802.11a control-frame signal classes. - Added safe fail-over behavior for component transform failures during signal
generation. - Fixed QAM signals being excluded from the AM family.
- Accelerated
digital_agc,clock_drift, andclock_jittertransforms with
Numba, and optimized the spur transform by replacing the per-spur FFT with an
analytic peak. - Cached polyphase filters and preserved resampler input precision.
Other
- Test coverage expansion – Added unit-test modules across many components
(builders, signal lists, dataset utilities, coordinate system, abstractions,
signal builders, DSP/numba, YAML, verification, base handling, printing, data
loading, defaults, and more). - Debug logging – Added debug-mode metadata generation logging.
- Module hygiene – Added
__all__and__init__.pyto every TorchSig module
and standardized signal-class family dictionaries. - Developer experience – Added a
make-based task runner for common tasks. - Documentation – Updated README, Makefile help, and example READMEs.