---
license: mit
library_name: pytorch
tags:
- astronomy
- fast-radio-burst
- transient-search
- object-detection
- image-classification
---

# DRAFTS Model Repository

Pretrained model weights for
[DRAFTS](https://github.com/SukiYume/DRAFTS), a deep-learning radio fast
transient search pipeline.

## Current search runtime

These are the two checkpoints used by the current DRAFTS observation-search
and injection-search entry points:

| Task | Architecture | File | Size | SHA-256 |
|---|---|---|---:|---|
| Candidate localization | CenterNet + ConvNeXt-Tiny, v10 | `object_best_model_centernet_conv_tiny_ema_v10.pth` | 114,449,795 bytes | `bcad4e710f5f1ccd3c8609d35a8d3fbfc36abd1d85bfefd035e945a573fb0629` |
| Candidate verification | ConvNeXt-Small binary classifier | `binary_best_model_conv_small_ema.pth` | 197,949,537 bytes | `2055745aab76ddc16074516aa7b9aafdfaedf16df37ce8924389573eab27ffd8` |

Download both files with the Hugging Face CLI:

```bash
mkdir -p runcode/models
hf download TorchLight/DRAFTS \
  object_best_model_centernet_conv_tiny_ema_v10.pth \
  binary_best_model_conv_small_ema.pth \
  --local-dir runcode/models
```

Or use direct HTTP downloads:

```bash
curl -L \
  -o runcode/models/object_best_model_centernet_conv_tiny_ema_v10.pth \
  https://huggingface.co/TorchLight/DRAFTS/resolve/main/object_best_model_centernet_conv_tiny_ema_v10.pth
curl -L \
  -o runcode/models/binary_best_model_conv_small_ema.pth \
  https://huggingface.co/TorchLight/DRAFTS/resolve/main/binary_best_model_conv_small_ema.pth
```

The injection runtime uses the same filenames under
`injection_experiment/search_runtime/models/`. Compare the downloaded files
with the SHA-256 values above before running a long search.

Model architecture definitions, preprocessing, command-line examples, and
deployment notes are maintained in the DRAFTS repository:

- [Observation-search runtime](https://github.com/SukiYume/DRAFTS/tree/main/runcode#readme)
- [CenterNet detector training](https://github.com/SukiYume/DRAFTS/tree/main/object_detection#readme)
- [Binary-classifier training](https://github.com/SukiYume/DRAFTS/tree/main/binary_classification#readme)
- [Injection experiment](https://github.com/SukiYume/DRAFTS/tree/main/injection_experiment#readme)

## Legacy checkpoints

Earlier ResNet detector and classifier checkpoints are grouped under
[`legacy/`](legacy/). They remain available for reproducing older DRAFTS
experiments, but they are not used by the current runtime.

| Task | File |
|---|---|
| CenterNet ResNet-18 detector | [`legacy/cent_resnet18.pth`](legacy/cent_resnet18.pth) |
| CenterNet ResNet-50 detector | [`legacy/cent_resnet50.pth`](legacy/cent_resnet50.pth) |
| ResNet-18 binary classifier | [`legacy/class_resnet18.pth`](legacy/class_resnet18.pth) |
| ResNet-50 binary classifier | [`legacy/class_resnet50.pth`](legacy/class_resnet50.pth) |

## License and citation

The accompanying DRAFTS code is released under the MIT License. If DRAFTS is
useful in your work, cite
[DRAFTS: A Deep Learning-Based Radio Fast Transient Search Pipeline](https://arxiv.org/abs/2410.03200)
and record the exact checkpoint and search configuration used.