EBNet

Credit: Logan Sizemore, Marina Kounkel

EBNet is a pipeline for predicting stellar parameters for eclipsing binaries using light curves.

Installation

Clone the repository and run:

pip install .

from within the EBNet directory.

Usage

Run the pipeline with:

python -m ebnet <data_path>

Positional Arguments

• data_path: Path to an input FITS file or a directory containing FITS files.

Optional Arguments

• -h, --help: Show help information and exit.
• -m, --model_type: Model type for prediction. Default is mixed. Options:
  • tf_model
  • pt_model
  • mixed
• -mt, --meta_type: Metadata representation type. Default is magnitude. Options:
  • magnitude — raw magnitudes (catalog-style).
  • flux — log10(lambda * F_lambda) representation. When --download_flux is used, metadata are automatically converted from flux densities (Jy) to log-scaled fluxes.
• -o, --output: Path to save results as a FITS table. If not provided, results are printed as CSV to standard output.
• -d, --download_flux: Download SED flux metadata from VizieR and overwrite or fill metadata flux columns in the input table.
• --num_workers: Number of parallel download workers for SED flux retrieval (default: 1). Increasing this may speed up downloads but can trigger rate limits.
• --device: PyTorch device to run inference on. Default is "cpu".
• -v, --verbose: Enable verbose output, including SED download status and missing column warnings.

Input File Requirements

An input FITS file must contain a period column.

It may include any number of the following light curve columns:

• Kepler_mean
• TESS_T
• Johnson_U, Johnson_B, Johnson_V, Johnson_R, Johnson_I, Johnson_J, Johnson_H, Johnson_K
• Cousins_R, Cousins_I
• SDSS_u, SDSS_g, SDSS_r, SDSS_i, SDSS_z
• SDSS_uprime, SDSS_gprime, SDSS_rprime, SDSS_iprime, SDSS_zprime
• Gaia_BP, Gaia_G, Gaia_RP, Gaia_RVS
• PanStarrs_g, PanStarrs_r, PanStarrs_w, PanStarrs_open, PanStarrs_i, PanStarrs_z, PanStarrs_y
• Stromgren_u, Stromgren_v, Stromgren_b, Stromgren_y
• LSST_u, LSST_g, LSST_r, LSST_i, LSST_z, LSST_y3
• SWASP_default
• ZTF_g, ZTF_r, ZTF_i
• Tycho_B, Tycho_V
• Hipparcos_Hp

It may also include any number of the following metadata columns:

• GALEX.FUV, GALEX.NUV
• Johnson.U, Johnson.B, Johnson.V, Johnson.R, Johnson.I
• Cousins.U, Cousins.B, Cousins.V, Cousins.R, Cousins.I
• SDSS.u, SDSS.g, SDSS.r, SDSS.i, SDSS.z
• PAN-STARRS.PS1.g, PAN-STARRS.PS1.r, PAN-STARRS.PS1.i, PAN-STARRS.PS1.z, PAN-STARRS.PS1.y
• GAIA.GAIA3.Gbp, GAIA.GAIA3.G, GAIA.GAIA3.Grp
• 2MASS.J, 2MASS.H, 2MASS.Ks
• WISE.W1, WISE.W2, WISE.W3

It may include any number of the following radial velocity columns:

• rv1
• rv2

It may include any number of the following parallax columns:

• parallax
• parallax_error

Example Commands

1. Predict stellar parameters for all stars in sample.fits and write results to output.csv:

   python -m ebnet sample.fits > output.csv

2. Predict stellar parameters for all stars in all FITS files in a directory and write results to output.fits:

   python -m ebnet ./path/to/data/ -o output.fits

3. Run predictions using a GPU:

   python -m ebnet sample.fits --device cuda > output.csv

4. Download missing SED metadata from VizieR using four workers:

   python -m ebnet sample.fits --download_flux --num_workers 4 --verbose

---

Programmatic Usage

EBNet can also be used directly from Python without calling the command line interface.

import ebnet
from astropy.table import Table

# Load input data from a FITS file
path = "./sample.fits"
data = Table.read(path)

# Run predictions with the default "mixed" backend
result = ebnet.predict(
    data,
    model_type="mixed",
    meta_type="magnitude",
    device="cuda",
    download_flux=False,
    num_workers=1,
    verbose=True
)

# Save results to a FITS file
result.write("output.fits", format="fits", overwrite=True)

Inputs

The predict function accepts three types of inputs:

• An in-memory Astropy Table
• A path to a single FITS file
• A path to a directory containing one or more FITS files

Options

• model_type: Model backend for predictions. Options:
  • "tf_model" – TensorFlow-based model
  • "pt_model" – PyTorch-based model
  • "mixed" – Uses both models, selecting the backend per target
• meta_type: Metadata representation type. Options:
  • "magnitude" — raw magnitudes (catalog-style)
  • "flux" — log10(lambda * F_lambda) representation
  • Automatically set to "flux_jy" when download_flux=True
• device: PyTorch device for computation (e.g., "cpu", "cuda", "cuda:0", "mps").
• download_flux: If True, queries VizieR for SED fluxes and replaces existing metadata flux columns.
• num_workers: Number of threads for concurrent SED metadata downloads.
• verbose: Prints progress messages and missing column warnings to standard error.
• seed: Sets NumPy and PyTorch random seeds for reproducibility.

Outputs

The function returns an Astropy Table with:

• <target>_pred: Predicted value for each target parameter
• <target>_std: One-sigma uncertainty for each prediction
• per0_pred / per0_std: Derived argument of periastron in degrees and its uncertainty
• phase0_pred / phase0_std: Derived orbital phase zero in cycles and its uncertainty