Pysersic

Installation

pip install pysersic
# If arviz 1.0+ breaks it:
pip install 'arviz<1.0'
# If photutils is too old for numpy 2:
pip install --upgrade photutils

JAX backend:

• CPU (macOS/Linux): pip install 'jax[cpu]'
• GPU (Linux NVIDIA CUDA 12): pip install --upgrade 'jax[cuda12]'

Dependencies: JAX, NumPyro, astropy, photutils, scipy, matplotlib, arviz (<1.0), asdf, corner, equinox, interpax, tqdm.

Profile Types

Core Workflow

1. Prepare inputs (numpy arrays)

import numpy as np
from astropy.io import fits

# Image cutout, RMS (sigma/noise) map, PSF, mask (True = bad pixel)
image = fits.getdata("image.fits").astype(float)
rms = np.sqrt(fits.getdata("variance.fits").astype(float))
psf = fits.getdata("psf.fits").astype(float)
psf /= psf.sum()  # normalize
mask = fits.getdata("mask.fits").astype(bool)  # True = masked

2. Validate inputs

from pysersic import check_input_data
check_input_data(data=image, rms=rms, psf=psf, mask=mask)

3. Set up prior

Auto-prior (recommended for quick start):

from pysersic.priors import autoprior

prior = autoprior(
    image=image,
    profile_type="sersic",
    mask=mask,
    sky_type="none",  # "none", "flat", or "tilted-plane"
)

Manual prior via SourceProperties:

from pysersic.priors import SourceProperties

props = SourceProperties(image, mask=mask)
# Override guesses if needed:
# props.set_flux_guess(flux, flux_err)
# props.set_r_eff_guess(r_eff, r_eff_err)
prior = props.generate_prior("sersic", sky_type="none")

4. Create fitter and run

from pysersic import FitSingle
from pysersic.loss import gaussian_loss, student_t_loss
from jax.random import PRNGKey

fitter = FitSingle(
    data=image,
    rms=rms,
    psf=psf,
    prior=prior,
    mask=mask,
    loss_func=student_t_loss,  # robust to outliers
)

# MAP (fast, seconds)
map_params = fitter.find_MAP(rkey=PRNGKey(42))

# Posterior approximation (minutes)
svi_res = fitter.estimate_posterior(rkey=PRNGKey(43), method="svi-flow")

# Full NUTS sampling (slower, most accurate)
fitter.sample(rkey=PRNGKey(44))
sampling_res = fitter.sampling_results

5. Results

from pysersic.results import plot_residual

# Residual plot: data | model | residual
fig, ax = plot_residual(image, map_params["model"], mask=mask, vmin=-1, vmax=1)

# Parameter summary
sampling_res.summary()

# Corner plot
fig = sampling_res.corner()

# Quantiles as DataFrame
df = sampling_res.retrieve_param_quantiles(return_dataframe=True)

# LaTeX table
sampling_res.latex_table()

# Save to ASDF
sampling_res.save_result("fit_result.asdf")

Multi-Source Fitting

from pysersic import FitMulti
from pysersic.priors import PySersicMultiPrior, estimate_sky
from pysersic.results import parse_multi_results

# Source catalog (from sep, photutils, or manual)
catalog = {
    "flux": [1000, 500, 50],
    "x": [50.0, 30.0, 70.0],
    "y": [50.0, 60.0, 40.0],
    "r": [5.0, 3.0, 1.0],       # effective radius guess
    "type": ["sersic", "sersic", "pointsource"],
}

# Sky estimate
med_sky, std_sky, n_pix = estimate_sky(image, mask)

prior = PySersicMultiPrior(
    catalog=catalog,
    sky_type="flat",
    sky_guess=med_sky,
    sky_guess_err=2 * std_sky / np.sqrt(n_pix),
)

fm = FitMulti(data=image, rms=rms, psf=psf, prior=prior)
map_dict = fm.find_MAP(rkey=PRNGKey(99))

# Extract single source from multi-fit
fm.estimate_posterior(method="laplace", rkey=PRNGKey(100))
source_2 = parse_multi_results(fm.svi_results, 2)
source_2.summary()

Key Parameters (Sérsic fit)

Tips

• Loss function: Use student_t_loss for robustness to bad pixels/artifacts; gaussian_loss for clean data.
• Sky model: Use sky_type="none" if background is already subtracted; "flat" to fit a constant residual.
• PSF: Must be normalized (sum=1), same pixel scale as image. Odd dimensions preferred.
• Mask convention: True = bad/masked pixel (excluded from fit).
• Speed: MAP takes seconds; svi-flow takes minutes; NUTS sampling takes 5–30 min per source.
• macOS: JAX GPU not supported; CPU-only. Still fast for individual sources.
• Multi-band: Not natively supported as a single call. Fit bands independently, then combine posteriors. Or define custom NumPyro model for joint fitting.