astra-clr

AstraCLR

Part of the light-curve family of open-source tools for astronomical time-series analysis.

Available from Python via the light-curve package: pip install light-curve. Documentation: light-curve.snad.space.

Paper

Majumder et al., 2026, in prep.

Original model

GitHub: https://github.com/TorshaMajumder/astra (private)

HuggingFace: https://huggingface.co/ashrot/astra-clr-base

Inference library: https://github.com/snad-space/astra-infer

License

MIT

Model overview

AstraCLR is a contrastive-learning encoder for multi-band photometric light curves. It maps a ZTF (g, r, i) light curve to a 512-dimensional embedding via a transformer architecture trained with a contrastive objective. Unlike ASTROMER, the model accepts multi-band input by concatenating per-band observation sequences with a log-wavelength channel. The ONNX file is distributed pre-built and requires no ML-framework conversion.

Inputs

The 700-element sequence is a concatenation of three per-band windows: g (0–299, 300 obs), r (300–649, 350 obs), i (650–699, 50 obs).

Output

Preprocessing steps

1. Sort observations chronologically within each band.
2. Select the first SEQ_PER_BAND[band] observations ("beginning" strategy).
3. Magnitude normalisation per band: subtract the inverse-variance-weighted mean. norm_mag = mag − Σ(mag/magerr²) / Σ(1/magerr²)
4. Time normalisation: norm_time = mjd − 58 000
5. Band channel: band_info = lg(eff_wavelength_Å) — g: lg(4746.48), r: lg(6366.38), i: lg(7829.03)
6. Padding: zero-pad shorter windows to the required length; set mask = 1 for padded positions, mask = 0 for real observations.

Model file

HuggingFace (ONNX): https://huggingface.co/light-curve/astra-clr

Source: https://huggingface.co/ashrot/astra-clr-base/blob/main/astra-clr.onnx Training data: ZTF Zubercal DR16 (https://huggingface.co/datasets/snad-space/astra-zubercaldr16_gaiadr3vclassre)