import torch
import timm

# Load the model directly from the Hub
model = timm.create_model(
    'hf-hub:parlange/cvt-gravit-b3',
    pretrained=True
)
model.eval()

# Example inference
dummy_input = torch.randn(1, 3, 224, 224)
with torch.no_grad():
    output = model(dummy_input)
    predictions = torch.softmax(output, dim=1)
print(f"Lens probability: {predictions[0][1]:.4f}")

⚡️ Training Configuration

Training Dataset: J24 (Jaelani et al. 2024)
Fine-tuning Strategy: all-blocks

🔧 Parameter	📝 Value
Batch Size	192
Learning Rate	AdamW with ReduceLROnPlateau
Epochs	100
Patience	10
Optimizer	AdamW
Scheduler	ReduceLROnPlateau
Image Size	224x224
Fine Tune Mode	all_blocks
Stochastic Depth Probability	0.1

📈 Training Curves

🏁 Final Epoch Training Metrics

Metric	Training	Validation
📉 Loss	0.1426	0.0469
🎯 Accuracy	0.9042	0.9898
📊 AUC-ROC	0.9816	0.9982
⚖️ F1 Score	0.9008	0.9897

☑️ Evaluation Results

ROC Curves and Confusion Matrices

Performance across all test datasets (a through l) in the Common Test Sample (More et al. 2024):

📋 Performance Summary

Average performance across 12 test datasets from the Common Test Sample (More et al. 2024):

Metric	Value
🎯 Average Accuracy	0.7963
📈 Average AUC-ROC	0.7784
⚖️ Average F1-Score	0.4644

📘 Citation

If you use this model in your research, please cite:

bibtex

@misc{parlange2025gravit,
      title={GraViT: Transfer Learning with Vision Transformers and MLP-Mixer for Strong Gravitational Lens Discovery}, 
      author={René Parlange and Juan C. Cuevas-Tello and Octavio Valenzuela and Omar de J. Cabrera-Rosas and Tomás Verdugo and Anupreeta More and Anton T. Jaelani},
      year={2025},
      eprint={2509.00226},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2509.00226}, 
}

Model Card Contact

For questions about this model, please contact the author through: https://github.com/parlange/