TorchSig v1.0.0-beta

Since this is almost an entirely new version of TorchSig, it is not backwards compatible with previous versions.

What's Changed (a lot!)

tldr; Look at our example notebooks, and start with getting_started.ipynb

• Datasets
  • Dataset configuration with new class DatasetMetadata, which holds configuration information about dataset
    • supports YAML configuration
    • Official Narrowband and Wideband datasets are define in datasets/default_configs with loaders
  • Two types of dataset classes: NewDataset and StaticDataset
    • NewDataset support infinite or finite signal generation, but does not store samples in memory
      • NewNarrowband, NewWideband
    • StaticDataset supports loading a dataset from disk
      • StaticNarrowband, StaticWideband
  • Datamodules for custom Narrowband/Wideband dataset and Official Narrowband/Wideband datasets
• Signals
  • New signal classes: DatasetSignal, DatasetDict.
  • SignalMetadata redefined as class instead of TypedDict.
  • SignalBuilder for signal generation.
  • Signal lists variable names updated.
• Transforms
  • Transforms redefined as changing signal data and signal metadata
    • Grouped into three categories: SignalTransform, DatasetTransform, and Impairments.
    • Many transforms refactored, optimized, or removed.
  • Target transforms redefined as label/output creators, do not change signal data at all
    • Can ask for any configuration of labels
• Utils
  • Random class Seedable to manage randomness across all TorchSig objects
  • Various printing and verification code
  • Dataset writing and reading using Zarr and YAML
    • New class: DatasetCreator
• Examples
  • Examples no longer need to be run in any specific order.
  • Moved Jupyter annotation tool example to tools/examples.
• Tests
  • More comprehensive test suit with Pytest.
• Google-style docstrings have been added for most of the code.
  • Updated documentation.

What Has Not Changed

• gr-spectrumdetect
• image_datasets

Notes

• v0.6.1 pretrained models are out of date.
• OFDM signal not fully featured
  • 0.6.1 OFDM has a simulated protocol on it: Resource blocks, pilots, etc. that does not exist in 1.0 yet.
  • 0.6.1 OFDM has some windowing in the time domain acting as pulse shaping, that does not exist in 1.0 yet
  • Otherwise the OFDM is accurate and representative, it is idealized and textbook-like, just as our other signals are
• We are depreciating our support of models. Specifically, we plan on only maintaining our pretrained models.
  • Our code should be compatible with any PyTorch compatible model, or should be easy to make compatible.
  • Still keep example notebooks to show how to train and use a model with our datasets.
• gr-spectrumdetect had not been updated with most recent code.

Known Bugs

• NewDatasets are not perfectly reproducible.
  • To guarantee reproducibility, we recommend writing the dataset to disk and loading back in the ensure correctness of samples.

Open-Source Collaboration

• We would really like to involve the open-source community, and are moving development to Github.
• Please feel free to make contributions towards bugs, new features, current issues, or anything on our Roadmap on the wiki.
• Github workflows for Pylint and Pytest to merge with main:
  • Pylint score must be above 9/10
  • Pytests should all pass without error