SwinYNet

Le résumé fourni par la source

SwinYNet – Transformer-based Real-time FRB & Pulsar Search Tool SwinYNet is an end-to-end Transformer model that simultaneously detects Fast Radio Bursts (FRBs) / pulsars, segments their signals at pixel-level, and estimates Dispersion Measure (DM) & Time-of-Arrival (ToA) – without de-dispersion preprocessing.Trained only on simulated data, it generalises to real FAST observations and enables real-time, petabyte-scale blind searches on a single GPU. 🌟 Key Features Feature Description One-for-All Detection + Segmentation + DM/ToA estimation in a single forward pass. No de-dispersion Operates directly on raw time–frequency data; no DM trials, no dedispersion cost. SOTA performance on FAST-FREX F1 = 97.8%, Recall = 95.7%, and Precision = 100%. Real-time speed Achieved 1.3–2.3× real-time speed on FAST CRAFTS data. Petabyte ready Already processed 2.8 PB CRAFTS data, found 2 known pulsars. Plug-and-play Outputs plug straight into PRESTO/prepfold and fitburst for refined folding & fitting. Open & Easy Pure PyTorch; install with uv in one minute; supports FITS & SIGPROC files. 🚀 Quick Start 1. Install uv (if not yet) curl -LsSf https://astral.sh/uv/install.sh | sh # macOS / Linux # Windows: powershell -c "irm https://astral.sh/uv/install.ps1 | iex" 2. Clone & create environment git clone https://github.com/expnn/SwinYNet.git cd SwinYNet uv venv --python 3.11 --managed-python uv sync # installs all deps (PyTorch, numpy, astropy, etc.) Mainland China?uv sync --index-url https://pypi.tuna.tsinghua.edu.cn/simple 3. Run inference swinynet -c cfgs/te8hjj4j.yaml \ -p data/input-fits-files \ -o data/output/manifest.json Get help anytime:swinynet -h 📁 Repository Layout SwinYNet ├── frbd/ # Core package │ ├── model/ # SwinYNet architecture │ ├── cfgs/ # YAML configs │ ├── data/ # Data structures, FAST-FITS & SIGPROC readers │ ├── config.py # global configurations of this project │ ├── theory.py # Formulas & basic operations related to FRB or FRB data processing. │ └── main.py # SwinYNet inference entrypoint. └── cache/ # weights of the trained models. 🔌 Integration with Existing Pipelines Tool How Benefit PRESTO / prepfold swinynet → ToA & DM → DM & Period → prepfold faster & much fewer candidate files fitburst Model masks & DM/ToA as priors fitting success ↑ 65% → 96%

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.