---
license: mit
pipeline_tag: feature-extraction
---

# 🌌 SpecCLIP: Aligning and Translating Spectroscopic Measurements for Stars

[![arXiv](https://img.shields.io/badge/arXiv-2507.01939-b31b1b.svg)](https://arxiv.org/abs/2507.01939)
[![GitHub](https://img.shields.io/badge/GitHub-Repo-black)](https://github.com/Xiaosheng-Zhao/SpecCLIP)
[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](https://github.com/Xiaosheng-Zhao/SpecCLIP/blob/main/LICENSE)

**SpecCLIP** is a contrastive + domain-preserving foundation model designed to align **LAMOST LRS** spectra with **Gaia XP** spectrophotometric data.
It learns a **general-purpose spectral embedding (768-dim)** that supports:

*   **Stellar parameter estimation**
*   **Cross-survey spectral translation** (LAMOST LRS ⟷ Gaia XP)
*   **Similarity retrieval** across LAMOST LRS and GAIA XP spectra

For full documentation, installation instructions, examples, and end-to-end usage, please visit the **GitHub repository**:
👉 [https://github.com/Xiaosheng-Zhao/SpecCLIP](https://github.com/Xiaosheng-Zhao/SpecCLIP)

---

## 🔧 Available Models

The following pretrained weights are included in this model repository:

| File                                         | Description                           | Embedding Dim | Param |
| -------------------------------------------- | ------------------------------------- | ------------- | ------|
| `encoders/lrs_encoder.ckpt`                  | LAMOST LRS masked transformer encoder | 768           |  43M  |
| `encoders/xp_encoder.ckpt`                   | Gaia XP masked transformer encoder    | 768           |  43M  |
| `encoders/xp_encoder_mlp.ckpt`               | Gaia XP autoencoder (MLP head)        | 768           |  43M  |
| `specclip/specclip_model_base.ckpt`          | Gaia XP  ⟷ LAMOST contrastive        | 768           |  100M |
| `specclip/specclip_model_predrecon_mlp.ckpt` | CLIP alignment + pred+recon           | 768           |  168M |
| `specclip/specclip_model_split_mlp.ckpt`     | CLIP alignment + split pred/recon     | 768           |  126M |

---

## 🧠 What the Model Does

SpecCLIP consists of:

*   **Two masked transformer encoders**
    – LAMOST LRS
    – Gaia XP
*   **Contrastive alignment loss (CLIP-style)**
*   **Domain-preserving prediction & reconstruction heads**
*   **Cross-modal decoder** for spectrum translation

It produces **shared embeddings** enabling multi-survey astrophysical analysis.

---

## Sample Usage
The following examples are adapted from the [official GitHub repository](https://github.com/Xiaosheng-Zhao/SpecCLIP).

### Installation

First, create a conda environment and install requirements:
```bash
conda create -n specclip-ai python=3.10
conda activate specclip-ai
conda install pytorch==2.5.1 torchvision==0.20.1 pytorch-cuda=11.8 -c pytorch -c nvidia
conda install numpy==2.0.1 scipy==1.15.3 pandas==2.3.3 mkl mkl-service -c defaults
pip install -r requirements.txt
pip install -e .
```

### Spectral Translation

Predict Gaia XP spectrum from LAMOST LRS:
```python
import json
from spectral_retrieval import SpectralRetriever
from predict_lrs_wclip_v0 import load_spectrum_data

# Configuration
with open('config_retrieval.json', 'r') as f:
    config = json.load(f)
retriever = SpectralRetriever(**config)

# Load the external spectra data
wavelength, flux = load_spectrum_data('./test_data/lrs/sample1_matrix.fits')

# Predict corresponding Gaia XP spectrum
prediction_external = retriever.predict_cross_modal(
    query_spectrum=(wavelength, flux),
    query_type='lamost_spectra'
)

# Plot
retriever.plot_cross_modal_prediction(
    prediction_external,
    save_path='./plots/external_lamost_to_gaia_prediction.png'
)
```

### Spectral Similarity Search

Find the top-4 most similar stars from Gaia XP catalog:
```python
# Download test data only
!python download_and_setup.py --test-data-only

# Build embedding database from test data
retriever.build_embedding_database(batch_size=1000, save_path='./test_embeddings.npz')

# Load external LAMOST spectrum
wavelength, flux = load_spectrum_data('./test_data/lrs/sample1_matrix.fits')

# Find similar Gaia XP spectra
results_external_cross = retriever.find_similar_spectra(
    query_spectrum=(wavelength, flux),
    query_type='lamost_spectra',
    search_type='cross_modal',
    top_k=4
)

# Plot
retriever.plot_retrieval_results(
    results_external_cross,
    save_path='./plots/external_lamost_to_gaia_cross.png'
)
```

### Parameter Prediction

**Coming soon.**
This section will include examples of using SpecCLIP embeddings with downstream models (e.g., MLP, SBI) for stellar-parameter prediction.

---

## 📄 Full Documentation

To keep the Hugging Face card concise, **all detailed instructions**, including:

*   Installation
*   Parameter prediction
*   Spectral translation
*   Retrieval
*   Full examples (Python + figures)
*   Acknowledgments

are available at the GitHub repo:

👉 **[https://github.com/Xiaosheng-Zhao/SpecCLIP](https://github.com/Xiaosheng-Zhao/SpecCLIP)**

---

## 📊 Citation

```bibtex
@ARTICLE{2025arXiv250701939Z,
       author = {{Zhao}, Xiaosheng and {Huang}, Yang and {Xue}, Guirong and {Kong}, Xiao and
                 {Liu}, Jifeng and {Tang}, Xiaoyu and {Beers}, Timothy C. and
                 {Ting}, Yuan-Sen and {Luo}, A-Li},
        title = "{SpecCLIP: Aligning and Translating Spectroscopic Measurements for Stars}",
      journal = {arXiv e-prints},
     keywords = {Instrumentation and Methods for Astrophysics, Solar and Stellar Astrophysics,
                 Artificial Intelligence, Machine Learning},
         year = 2025,
        month = jul,
          eid = {arXiv:2507.01939},
        pages = {arXiv:2507.01939},
          doi = {10.48550/arXiv.250701939},
archivePrefix = {arXiv},
       eprint = {2507.01939},
 primaryClass = {astro-ph.IM},
}
```

---

## 📬 Contact

*   GitHub Issues: [https://github.com/Xiaosheng-Zhao/SpecCLIP/issues](https://github.com/Xiaosheng-Zhao/SpecCLIP/issues)
*   Email: [xzhao113@jh.edu](mailto:xzhao113@jh.edu)