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

Use PyPI to install:

pip install torchsig

Or clone the torchsig repository and install using the following commands:

git clone https://github.com/TorchDSP/torchsig.git
cd torchsig

# Runtime only (for normal usage)
pip install -e .

# Full development environment (for contributors)
pip install -e .[dev]

# Documentation build environment
pip install -e ".[docs]"

# To run the notebooks
pip install -e .[notebooks]

# To run the notebooks, including examples
pip install -e .[examples]

When to use which command?

• pip install -e . — pulls in only the runtime dependencies (PyTorch, OpenCV, etc.).
  Use this if you just want to run the library in your own projects.

• pip install -e .[dev] — adds the extra dev group, installing testing, linting, and coverage tools.
  It also includes the documentation dependencies. Choose this when you plan to develop, run the test suite, build the documentation, or contribute code back to the project.

• pip install -e .[docs] — installs the Sphinx documentation toolchain without the testing and linting tools. Choose this when you only need to build the documentation.

• pip install -e .[notebooks] — adds the extra notebook group, jupyter notebook tools. Choose this when you plan to develop or run notebooks for the project.

• pip install -e .[examples] — adds the extra examples group, including optional dependencies to run all examples in the notebooks and scripts. Choose this when you plan to run the examples from the project.

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

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 docker/Dockerfile .

And then you can launch a Docker instance:

docker run -it torchsig

See docker/README.md to learn more.

Development

To contribute to our library, please make sure to run the following:

# pytests all pass
pytest

# pylint score > 9/10
pylint --rcfile=.pylintrc torchsig

# not required
# but helpful for maintaining PEP 8 Style Guide
ruff check torchsig

Both need to pass in order to contribute to our Github.

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.

pip install -e ".[docs]"
make docs
firefox docs/build/html/index.html

🛠️ Development Workflow

To simplify environment setup and maintain code quality, this project uses a Makefile. This provides a standardized set of shortcuts for common development tasks, ensuring consistency across different environments.

Common Commands

For a full list of available targets and descriptions, run:

make help

Note for Windows Users: make is a Unix utility. To use these commands on Windows, please use WSL (Windows Subsystem for Linux), Git Bash, or install make via Chocolatey.

Pre-commit Hooks

TorchSig uses pre-commit to run automated checks before commits. After installing the development dependencies, install the Git hooks with:

pre-commit install

To run all pre-commit checks manually:

pre-commit run --all-files

The hooks will automatically run on staged files when committing. If a hook modifies a file, review and stage the changes before committing again.

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

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}
}