---

TorchSig is an open-source signal processing machine learning toolkit based on the PyTorch data handling pipeline. The user-friendly toolkit simplifies common digital signal processing operations, augmentations, and transformations when dealing with both real and complex-valued signals. TorchSig streamlines the integration process of these signals processing tools building on PyTorch, enabling faster and easier development and research for machine learning techniques applied to signals data, particularly within (but not limited to) the radio frequency domain.

Getting Started

Prerequisites

• Ubuntu ≥ 22.04
• Hard drive storage with 1 TB
• CPU with ≥ 4 cores
• GPU with ≥ 16 GB storage (recommended)
• Python ≥ 3.10

We highly reccomend Ubuntu or using a Docker container.

Installation

Clone the torchsig repository and install using the following commands:

git clone https://github.com/TorchDSP/torchsig.git
cd torchsig
pip install -e .

Examples and Tutorials

TorchSig has a series of Jupyter notebooks in the examples/ directory. View the README inside examples/ to learn more.

Usage

Generating Datasets with Python API

TorchSig uses a unified dataset architecture. Create datasets using the Python API:

# define dataset metadata, can override defaults
dataset_metadata = TorchSigDefaults().default_dataset_metadata

# optionally, apply impairments
impairments = Impairments(level=0)
burst_impairments = impairments.signal_transforms
whole_signal_impairments = impairments.dataset_transforms

# create the dataset
dataset = TorchSigIterableDataset(
  metadata=dataset_metadata,
  transforms=[whole_signal_impairments, Spectrogram(fft_size=dataset_metadata["fft_size"])],
  component_transforms=[burst_impairments],
)
# create a dataloader (reproducible)
dataloader = WorkerSeedingDataLoader(dataset, batch_size=2)

# save the dataset to disk
dataset_creator = DatasetCreator(
  dataset_length=20,
  dataloader=dataloader,
  root="./sample_dataset",
  overwrite=True,
  multithreading=False,
)
dataset_creator.create()

# load the dataset in from disk
static_dataset = StaticTorchSigDataset(
  root="./sample_dataset",
)

print(static_dataset[0])

Docker

One option for running TorchSig is within Docker. Start by building the Docker container:

docker build -t torchsig -f Dockerfile .

Generating Datasets with Docker

To create datasets with the Docker container, create a Python script and run it:

# create_dataset.py
from torchsig.datasets.datasets import TorchSigIterableDataset
from torchsig.utils.writer import DatasetCreator
from torchsig.utils.defaults import TorchSigDefaults
from torchsig.transforms.impairments import Impairments
from torchsig.transforms.transforms import Spectrogram

# Classification dataset (single signal)
dataset_metadata = TorchSigDefaults().default_dataset_metadata
dataset_metadata["num_iq_samples_dataset"] = 100
dataset_metadata["num_signals_min"] = 1
dataset_metadata["num_signals_max"] = 1

impairments = Impairments(level=0)
burst_impairments = impairments.signal_transforms
whole_signal_impairments = impairments.dataset_transforms

dataset = TorchSigIterableDataset(
    metadata=dataset_metadata,
    transforms=[
        whole_signal_impairments,
        Spectrogram(fft_size=dataset_metadata["fft_size"]),
    ],
    component_transforms=[burst_impairments],
)

dataloader = WorkerSeedingDataLoader(dataset, batch_size=4)

dataset_creator = DatasetCreator(
    dataset_length=10,
    dataloader=dataloader,
    root="/path/to/classification_dataset",
    overwrite=True,
    multithreading=False,
)
dataset_creator.create()

# Detection dataset (multiple signals)
dataset_metadata = DatasetMetadata(
    num_iq_samples_dataset=100,
    num_samples=10,
    impairment_level=2,  # wireless
    num_signals_max=3,
)
dataset = TorchSigIterableDataset(
    metadata=dataset_metadata,
    transforms=[Spectrogram(fft_size=dataset_metadata["fft_size"])],
)
creator = DatasetCreator(dataset, root="/path/to/detection_dataset")
creator.create()

dataset_metadata = TorchSigDefaults().default_dataset_metadata
dataset_metadata["num_iq_samples_dataset"] = 100
dataset_metadata["num_signals_min"] = 1
dataset_metadata["num_signals_max"] = 3

impairments = Impairments(level=2)
burst_impairments = impairments.signal_transforms
whole_signal_impairments = impairments.dataset_transforms

dataset = TorchSigIterableDataset(
    metadata=dataset_metadata,
    transforms=[
        whole_signal_impairments,
        Spectrogram(fft_size=dataset_metadata["fft_size"]),
    ],
    component_transforms=[burst_impairments],
)

dataloader = WorkerSeedingDataLoader(dataset, batch_size=4)

dataset_creator = DatasetCreator(
    dataset_length=10,
    dataloader=dataloader,
    root="/path/to/detection_dataset",
    overwrite=True,
    multithreading=False,
)
dataset_creator.create()

docker run -u $(id -u ${USER}):$(id -g ${USER}) -v `pwd`:/workspace/code/torchsig torchsig python3 create_dataset.py

Running Jupyter Notebooks with Docker

To run with GPU support use --gpus all:

docker run -d --rm --network=host --shm-size=32g --gpus all --name torchsig_workspace torchsig tail -f /dev/null

To run without GPU support:

docker run -d --rm --network=host --shm-size=32g --name torchsig_workspace torchsig tail -f /dev/null

Run Jupyter Lab:

docker exec torchsig_workspace jupyter lab --allow-root --ip=0.0.0.0 --no-browser

To start an interactive shell:

docker exec -it torchsig_workspace bash

Then use the URL in the output in your browser to run the examples and notebooks.

Key Features

TorchSig provides many useful tools to facilitate and accelerate research on signals processing machine learning technologies:

• Unified Dataset Architecture: TorchSig features a single, flexible dataset system that supports both signal classification (single signal) and signal detection (multiple signals) tasks through configuration.
• Comprehensive Signal Library: Support for 60+ signal types across all major modulation families (FSK, QAM, PSK, ASK, OFDM, Analog) with realistic impairments and channel effects.
• Advanced Transform System: Numerous signals processing transforms enable existing ML techniques to be employed on signals data, with unified impairment models supporting perfect, cabled, and wireless channel conditions.

Core Classes

• Signal and SignalMetadataObject: Enable signal objects and metadata to be seamlessly handled and operated on throughout the TorchSig infrastructure.
• TorchSigIterableDataset: Unified dataset class that synthetically creates, augments, and transforms signals datasets. Behavior (classification vs detection) is determined by configuration parameters.
  • Can generate samples infinitely when num_samples=None, or finite datasets when num_samples is specified.
  • Dataset type determined by num_signals_max: 1 for classification, >1 for detection tasks.
• DatasetCreator: Writes a PyTorch DataLoader containing a TorchSigIterableDataset objects to disk with progress tracking and memory optimization.
• StaticTorchSigDataset: Loads previously generated datasets from disk back into memory.
  • Can access previously generated samples efficiently.
  • Supports both classification and detection datasets through unified interface.

Documentation

Documentation can be found online or built locally by following the instructions below.

cd docs
pip install -r docs-requirements.txt
make html
firefox build/html/index.html

License

TorchSig is released under the MIT License. The MIT license is a popular open-source software license enabling free use, redistribution, and modifications, even for commercial purposes, provided the license is included in all copies or substantial portions of the software. TorchSig has no connection to MIT, other than through the use of this license.

Publications

Title	Year	Cite (APA)
TorchSig 2.0: Dataset Customization, New Transforms and Future Plans	2025	Oh, E., Mullins, J., Carrick, M., Vondal, M., Hoffman, J., Leonardo, F., Toliver, P., Miller, R. (2025, September). TorchSig 2.0: Dataset Customization, New Transforms and Future Plans. In Proceedings of the GNU Radio Conference (Vol. 10, No. 1).
TorchSig: A GNU Radio Block and New Spectrogram Tools for Augmenting ML Training	2024	Vallance, P., Oh, E., Mullins, J., Gulati, M., Hoffman, J., & Carrick, M. (2024, September). TorchSig: A GNU Radio Block and New Spectrogram Tools for Augmenting ML Training. In Proceedings of the GNU Radio Conference (Vol. 9, No. 1).
Large Scale Radio Frequency Wideband Signal Detection & Recognition	2022	Boegner, L., Vanhoy, G., Vallance, P., Gulati, M., Feitzinger, D., Comar, B., & Miller, R. D. (2022). Large Scale Radio Frequency Wideband Signal Detection & Recognition. arXiv preprint arXiv:2211.10335.
Large Scale Radio Frequency Signal Classification	2022	Boegner, L., Gulati, M., Vanhoy, G., Vallance, P., Comar, B., Kokalj-Filipovic, S., ... & Miller, R. D. (2022). Large Scale Radio Frequency Signal Classification. arXiv preprint arXiv:2207.09918.

Citing TorchSig

Please cite TorchSig if you use it for your research or business.

@misc{torchsig,
  title={Large Scale Radio Frequency Signal Classification},
  author={Luke Boegner and Manbir Gulati and Garrett Vanhoy and Phillip Vallance and Bradley Comar and Silvija Kokalj-Filipovic and Craig Lennon and Robert D. Miller},
  year={2022},
  archivePrefix={arXiv},
  eprint={2207.09918},
  primaryClass={cs-LG},
  note={arXiv:2207.09918}
  url={https://arxiv.org/abs/2207.09918}
}