README.md

SuperWASP Pulsator Detection

Automated classification of pulsating variable stars in SuperWASP light curves using deep learning (ResNet18 + Transfer Learning).

Results

Phase	F1-Score	Accuracy
Phase 1 — classifier only	0.694	69.9%
Phase 2 — fine-tuning	0.802	80.2%

Dataset

Zenodo VeSPA — SuperWASP variable star classifications
McMaster et al. 2021, Res. Notes AAS 5 228
DOI: 10.5281/zenodo.14937227
License: Creative Commons Attribution 4.0

Class	Images
Pulsators	1,999
Non-pulsators	2,000
Total	3,999

Method

• Architecture: ResNet18 pre-trained on ImageNet
• Phase 1: Only the classifier head is trained (frozen backbone)
• Phase 2: Fine-tuning of layer3 + layer4 + classifier with lower learning rate (5e-5)
• Augmentation: Random crop, horizontal/vertical flip, rotation, color jitter
• Early stopping: patience = 8 epochs

How to Use

Requirements

pip install -r requirements.txt

Step 1 — Download data

Open superwasp_01_download.ipynb in Google Colab and run all cells.
This downloads ~4,000 light curve images from Zenodo to your Google Drive.

Step 2 — Train model

Open superwasp_02_training.ipynb in Google Colab.
Enable GPU: Runtime → Change runtime type → GPU (T4)
Run all cells. The trained model is saved to your Google Drive.

Adjust paths

In both notebooks, update this line to match your Google Drive folder:

BASIS_PFAD = '/content/drive/MyDrive/SuperWASP'

Repository Structure

superwasp-pulsator-detection/
    superwasp_01_download.ipynb   ← Download images from Zenodo
    superwasp_02_training.ipynb   ← Train ResNet18 model
    requirements.txt              ← Python dependencies
    README.md

Citation

If you use this code or dataset, please cite:

McMaster et al. 2021, Res. Notes AAS 5 228
DOI: 10.5281/zenodo.14937227

Author

Desislava Lebessis
Citizen Scientist — SuperWASP / Zooniverse
GitHub · www.linkedin.com/in/desislava-lebessis-57aa77b4

License

MIT License — see LICENSE for details.