TorchSig 2.0: Dataset Customization, New Transforms and Future Plans

Authors

  • Erebus Oh
  • Justin Mullins
  • Matt Carrick
  • Matt Vondal
  • Jared Hoffman
  • Frank Leonardo
  • Paul Toliver
  • Rob Miller

Abstract

TorchSig’s v2.0 release improves the training and
testing of machine learning (ML) models with
synthetic datasets by providing transforms that
model important real-world effects like phase
noise, quantization, or amplifier nonlinearities.
Users also have more dataset generation controls
and customization options, such as parameterizing
the range of center frequencies, bandwidths,
and burst durations. Additionally, users can bring
their own data and use TorchSig’s framework for
data augmentation and model training.

Downloads

Published

2026-09-29