FRTSearch 0.1.3

Fast Radio Transients Search Tool — A deep learning-based detection tool for fast radio transients in pulsar search data

FRTSearch: Fast Radio Transient Search

FRTSearch is an end-to-end framework for discovering Pulsars, Rotating Radio Transients (RRATs), and Fast Radio Bursts (FRBs) in radio astronomical observation data. Single-pulse emissions from these sources all exhibit consistent dispersive trajectories governed by the cold plasma dispersion relation ($t \propto \nu^{-2}$) in time-frequency dynamic spectra. This shared signature serves as a key beacon for identifying these astrophysical sources. FRTSearch leverages a Mask R-CNN instance segmentation model and the IMPIC algorithm to directly detect and characterize Fast Radio Transients (FRTs), infer their physical parameters (DM, ToA), and generate diagnostic plots and candidate catalogs for manual verification and scientific analysis.

Core Components:

1. Mask R-CNN — Segments dispersive trajectories in time-frequency dynamic spectra, trained on the pixel-level annotated CRAFTS-FRT dataset.
2. IMPIC — Iterative Mask-based Parameter Inference and Calibration: infers DM and ToA directly from segmentation masks.   Code | Docs | Example

Supported formats:

Format	1-bit	2-bit	4-bit	8-bit	32-bit
PSRFITS (.fits)	✅	✅	✅	✅	—
Sigproc Filterbank (.fil)	✅	✅	✅	✅	✅

Installation

Option 1: pip install (Recommended)

Requires: Python 3.10+, PyTorch 2.0+, CUDA 11.7+

pip install FRTSearch

Option 2: From Source

git clone https://github.com/BinZhang109/FRTSearch.git && cd FRTSearch

# MMDetection
pip install -U openmim && mim install mmcv-full && pip install mmdet

# Other dependencies
pip install -r requirements.txt

Option 3: Docker

docker pull binzhang/FRTSearch

Download Model Weights

Download from Hugging Face and place into models/:

FRTSearch/
├── models/
│   └── hrnet_epoch_36.pth
├── configs/
│   ├── detector_FAST.py
│   └── detector_SKA.py
└── ...

Usage

Full Pipeline

# If installed via pip:
FRTSearch <data.fits|data.fil> <config.py> [--slide-size 128]

# If running from source:
python FRTSearch.py <data.fits|data.fil> <config.py> [--slide-size 128]

Argument	Description
data	Observation file (.fits or .fil)
config	Detector configuration file
--slide-size	Subintegrations per sliding window (default: 128)

Example

Test data can be downloaded from Hugging Face.

# FAST FRB detection
FRTSearch ./test_sample/FRB20121102_0038.fits ./configs/detector_FAST.py --slide-size 128

# SKA FRB detection
FRTSearch ./test_sample/FRB20180119_SKA_1660_1710.fil ./configs/detector_SKA.py --slide-size 8

Training

python train.py

Test Samples

python test_sample/test_samples.py --example FRB20121102

Available examples: FRB20121102, FRB20201124, FRB20180301, FRB20180119, FRB20180212

Dataset: CRAFTS-FRT

The first pixel-level annotated FRT dataset, derived from the Commensal Radio Astronomy FAST Survey (CRAFTS).

Instances	Source	Download
2,392 (2,115 Pulsars, 15 RRATs, 262 FRBs)	FAST 19-beam L-band	ScienceDB

Citation

@article{zhang2026frtsearch,
  title={FRTSearch: Unified Detection and Parameter Inference of Fast Radio Transients using Instance Segmentation },
  author={Zhang, Bin and Wang, Yabiao and Xie, Xiaoyao et al.}
  year={2026},
}

Test sample references: FAST — Guo et al. (2025)  |  SKA — Shannon et al. (2018)

Contributing

Open an Issue for bugs or questions. PRs welcome — see Contributing Guidelines.

License & Acknowledgments

This project is licensed under GPL-2.0.

Built upon: MMDetection | PRESTO

_{Exploring the dynamic universe with AI 🌌📡}