msiudek/astroPT_euclid_VIS_NISP_model

AstroPT Euclid VIS+NISP Model

Pre-trained AstroPT model for multi-band image analysis using Euclid VIS and NISP imaging.

Overview

This is a pre-trained checkpoint for the AstroPT framework, trained on multi-band imaging from the Euclid Q1 dataset. The model combines visible (VIS) and near-infrared (NISP: Y, J, H) bands for robust multi-wavelength astronomical image analysis.

Citation: Euclid Collaboration: Siudek, M et al. 2025 (arXiv:2503.15312)

Quick Start

Load Model

import torch
from pathlib import Path

# Load model checkpoint
model_path = "astropt/090M/ckpt.pt"
device = "cuda" if torch.cuda.is_available() else "cpu"

# Load state dict
checkpoint = torch.load(model_path, map_location=device)
# Initialize your model architecture here
# model.load_state_dict(checkpoint)

Inference

from datasets import load_dataset
import torch

# Load dataset
dataset = load_dataset(
    "msiudek/astroPT_euclid_dataset",
    split="train_batch_1",
    streaming=True
)

# Run inference
model.eval()
with torch.no_grad():
    for sample in dataset:
        # Stack 4 bands: VIS + Y + J + H
        images = torch.stack([
            torch.tensor(sample['VIS_image'], dtype=torch.float32),
            torch.tensor(sample['NISP_Y_image'], dtype=torch.float32),
            torch.tensor(sample['NISP_J_image'], dtype=torch.float32),
            torch.tensor(sample['NISP_H_image'], dtype=torch.float32),
        ])  # [4, 224, 224]
        images = images.unsqueeze(0)  # [1, 4, 224, 224]
        
        # Get embeddings
        embeddings = model(images)

Training Data

• Dataset: AstroPT Euclid Dataset
• Bands:
  • VIS: Euclid VIS (0.55–0.90 μm)
  • Y: Euclid NISP (1.0–1.2 μm)
  • J: Euclid NISP (1.2–1.4 μm)
  • H: Euclid NISP (1.4–2.0 μm)

Related Models

• AstroPT VIS Model: Single-band (VIS only)
• AstroPT VIS+NISP+SED Model: Multi-modal (imaging + photometry)

Datasets

• Imaging: astroPT_euclid_dataset
• Metadata: astroPT_euclid_metadata

Code & Documentation

For inference code, training scripts, and tutorials, visit the AstroPT GitHub Repository.

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
Model Version: 1.0