FETCH (Fast Extragalactic Transient Candidate Hunter) - ONNX Edition provides a lightweight, TensorFlow-free ONNX inference solution for fast radio burst (FRB) detection and classification.
This ONNX edition offers significant advantages over the original TensorFlow implementation:
- Zero TensorFlow dependencies - Pure ONNX Runtime for ultra-fast inference
- Cross-platform compatibility - Works on any ONNX-supported platform
- GPU/CPU flexibility - Automatic hardware detection with seamless fallback
FETCH uses the ONNX_HOME environment variable to determine where to store downloaded ONNX models:
# Set custom model directory (optional)
export ONNX_HOME=/path/to/your/models
# If not set, models are stored in $HOME/onnx_models by default
# Example: /home/user/onnx_models/The first time you run inference, models will be automatically downloaded from Zenodo and cached locally for future use.
# Install with CPU-only inference
uv pip install "fetch[cpu] @ git+https://github.com/devanshkv/fetch.git@onnx"
# Install with GPU acceleration
uv pip install "fetch[gpu] @ git+https://github.com/devanshkv/fetch.git@onnx"# Clone repository
git clone https://github.com/devanshkv/fetch.git
cd fetch
git checkout onnx
# CPU-only inference
pip install -r requirements_onnx.txt
pip install .
# GPU-accelerated inference
pip install -r requirements_onnx_gpu.txt
pip install .git clone https://github.com/devanshkv/fetch.git
cd fetch
git checkout onnx
# Install in development mode
pip install -e .The primary interface is through the fetch-predict command installed globally:
fetch-predict -m <model_id> -c <data_directory>Alternatively, use the script directly:
python src/fetch/predict_onnx.py -m <model_id> -c <data_directory># CPU inference with model 'a' (models auto-downloaded to $ONNX_HOME)
fetch-predict -m a -c /path/to/candidates -g -1
# GPU inference (automatically detects available GPU)
fetch-predict -m a -c /path/to/candidates
# Specific GPU with custom parameters
fetch-predict -m a -c /path/to/candidates -g 1 -b 16 -p 0.7
# High-throughput batch processing with custom model directory
fetch-predict -m j -c /path/to/large_dataset -b 32 --onnx_dir ./models
# Set custom model storage location
export ONNX_HOME=/data/models
fetch-predict -m a -c /path/to/candidates# Multiple model ensemble prediction
for model in {a..k}; do
fetch-predict -m $model -c /data/candidates -o results_$model.csv
done
# Custom threshold tuning
fetch-predict -m a -c /data/test_set -p 0.3 # High sensitivity
fetch-predict -m a -c /data/prod_set -p 0.8 # High precision| Argument | Short | Description | Default | Examples |
|---|---|---|---|---|
--model |
-m |
Model identifier (a-k) | Required | a, j, k |
--data_dir |
-c |
Directory containing H5 candidate files | Required | /data/candidates/ |
--gpu_id |
-g |
GPU device ID (-1 for CPU) | 0 | -1, 0, 1 |
--batch_size |
-b |
Inference batch size | 8 | 16, 32, 64 |
--probability |
-p |
Classification threshold | 0.5 | 0.3, 0.7, 0.9 |
--onnx_dir |
Model storage directory | $ONNX_HOME or $HOME/onnx_models |
./models/ |
|
--output |
-o |
Output CSV filename | results_{model}.csv |
predictions.csv |
Candidate files must be HDF5 format with the following structure:
candidate.h5
├── data_freq_time # 2D array: frequency vs time
├── data_dm_time # 2D array: DM vs time
└── metadata # Optional: SNR, DM, width, etc.
Expected dimensions:
- Frequency-time:
(256, 256)or(512, 512) - DM-time:
(256, 256)or(512, 512)
Results are saved as CSV with the following columns:
candidate,probability,label
cand_001.h5,0.9234,1.0
cand_002.h5,0.1456,0.0
cand_003.h5,0.8901,1.0
- candidate: Input filename
- probability: Model confidence score (0.0-1.0)
- label: Binary classification (>= threshold = 1.0, < threshold = 0.0)
# Clone and setup
git clone https://github.com/devanshkv/fetch.git
cd fetch
git checkout onnx
# Install build dependencies
pip install uv
# Build package
uv build
# Install locally
pip install dist/fetch-*.whl# Install development dependencies
pip install -r requirements_onnx.txt
pip install -e ".[dev]"
# Run tests
pytest tests/
# Code formatting
black .
ruff check .
# Type checking
mypy fetch/# Test all models
python bin/test_onnx_models.py
# Test specific model
python bin/test_onnx_models.py --model a
# Benchmark performance
python bin/test_onnx_models.py --benchmarkFETCH implements deep learning models for Fast Radio Burst (FRB) detection based on the methodology described in:
"FETCH: A deep-learning based classifier for fast transient classification"
Agarwal et al., Monthly Notices of the Royal Astronomical Society, 2020
DOI: 10.1093/mnras/staa1856
This project is licensed under the GNU General Public License v3.0. See LICENSE for details.
If you use FETCH in your research, please cite:
@article{Agarwal2020,
doi = {10.1093/mnras/staa1856},
url = {https://doi.org/10.1093/mnras/staa1856},
year = {2020},
month = jun,
publisher = {Oxford University Press ({OUP})},
volume = {497},
number = {2},
pages = {1661--1674},
author = {Devansh Agarwal and Kshitij Aggarwal and Sarah Burke-Spolaor and Duncan R Lorimer and Nathaniel Garver-Daniels},
title = {{FETCH}: A deep-learning based classifier for fast transient classification},
journal = {Monthly Notices of the Royal Astronomical Society}
}We welcome contributions! Please see our Contributing Guidelines for details.
- Fork the repository
- Create a feature branch (
git checkout -b feature/amazing-feature) - Commit your changes (
git commit -m 'Add amazing feature') - Push to the branch (
git push origin feature/amazing-feature) - Open a Pull Request
- Performance optimization: ONNX model quantization, pruning
- Hardware support: Apple Silicon, Intel GPU, AMD ROCm
- Data formats: Support for additional input formats
- Visualization: Real-time monitoring dashboards
- Integration: Pulsar search pipeline connectors
- Issues: GitHub Issues
- Discussions: GitHub Discussions
- Email: devansh.kv@gmail.com
FETCH ONNX Edition - Bringing state-of-the-art FRB detection to the masses with unprecedented efficiency and ease of use.