CASTOR

CASTOR is a transformer-based pipeline for gravitational-wave detection, training, inference, and sensitivity evaluation.

CASTOR is under active development. Version 0.1.0 packages the analysis workflow described in the accompanying paper.

Installation

Choose and install the appropriate CPU, CUDA, or ROCm build of PyTorch for your platform before installing CASTOR. This prevents a CPU-only Linux environment from downloading an unintended CUDA runtime. Then install the released package with:

python -m pip install castor-gw

For an editable development installation, install from a clone:

git clone https://github.com/chayanchatterjee/castor.git
cd castor
python -m pip install -e .

Install the scientific extras needed by a workflow with:

python -m pip install -e ".[training]"   # ML4GW online training
python -m pip install -e ".[evaluation]" # PyCBC whitening
python -m pip install -e ".[plotting]"   # plotting dependencies

Verify the installed commands with:

castor-train --help
castor-eval --help

See the documentation for the training, evaluation, cluster, data-format, and reproducibility guides.

Reproducibility

The paper workflow uses 10--50 solar-mass component masses, 500 epochs, global seed 2026, deterministic validation, raw checkpoint weights, and a factor of 32 that reconciles the PyCBC and ML4GW whitening conventions. Checkpoints and evaluation caches record the configuration and input identities required to reject incompatible reuse.

License

CASTOR is distributed under the Apache License 2.0.