msiudek/astroPT_euclid_VIS_NISP_embeddings

AstroPT Euclid Embeddings

Pre-computed embeddings from the AstroPT VIS+NISP Model on Euclid dataset samples.

Overview

This repository contains pre-computed feature embeddings generated by AstroPT models applied to the Euclid Q1 galaxy dataset. These embeddings can be used for

• Efficient downstream task training (reduced computational cost)
• Feature analysis and visualization
• Similarity search and retrieval
• Clustering and unsupervised learning
• Fast fine-tuning on specialized tasks

Example notebooks are found in the AstroPT scripts

Available embeddings:

• VIS: Single-band embeddings from Euclid VIS imaging; VIS embeddings
• VIS+NISP: Multi-band embeddings from VIS + 3× NIR (Y, J, H); VIS+NISP embeddings
• VIS+NISP+SED: Multi-modal embeddings from imaging + 13-band photometry; VIS+NISP+SED embeddings

Quick Start

Load VIS+NSIP Embeddings

from datasets import load_dataset

# Load VIS embeddings
embeddings = load_dataset(
    "msiudek/astroPT_euclid_VIS_NISP_embeddings",
    split="train",
    streaming=True
)

# View a sample
sample = embeddings[0]
print(f"Object ID: {sample['object_id']}")
print(f"Embedding shape: {len(sample['embedding'])}")
print(f"Embedding: {sample['embedding']}")

Related Resources

Models:

• AstroPT VIS Model
• AstroPT VIS+NISP Model
• AstroPT VIS+NISP+SED Model

Datasets:

• AstroPT Euclid Dataset: Original imaging + photometry
• AstroPT Euclid Metadata: Galaxy properties

Code:

• AstroPT GitHub: Training and inference code

Citation

@article{Siudek2025,
  title={AstroPT: Astronomical Physics Transformers for Multi-modal Learning},
  author={Siudek, M and others},
  journal={Euclid Collaboration},
  eprint={2503.15312},
  archivePrefix={arXiv},
  year={2025},
  url={https://ui.adsabs.harvard.edu/abs/2025arXiv250315312E/abstract}
}

License

CC-BY-4.0

---

Last Updated: December 2025
Embeddings Version: 1.0