Transform gallery¶

Curated before/after figures. Full API: Transforms reference. Compose smoke test: example_transforms.py.

pixi run python examples/gallery_images.py
pixi run python examples/gallery_tables_lc.py
# copy only the allowlisted PNGs below into docs/assets/gallery/

Gallery PNG allowlist (copy from examples/output/ when refreshing):

• image_compose_pipeline.png
• lightcurve_sigma_clip.png, lightcurve_asymmetric_sigma_clip.png
• table_fits_scale_columns.png
• lupton_rgb_sdss.png (from example_lupton_rgb_sdss.py)

Public samples: bash scripts/fetch_example_samples.sh. TORCHFITS_EXAMPLE_FAST=1 uses synthetic fallbacks for the gallery scripts when the cache is empty; example_lupton_rgb_sdss.py skips cleanly instead.

Image pipeline (Compose)¶

HorseHead alone has limited stretch contrast; the useful story is a Compose of background → arcsinh → zscale:

import torchfits
from torchfits.transforms import (
    ArcsinhStretch,
    BackgroundSubtract,
    Compose,
    ZScaleNormalize,
)

tensor = torchfits.read_tensor("horsehead.fits", hdu=0)
pipeline = Compose(
    [BackgroundSubtract(), ArcsinhStretch(a=0.1), ZScaleNormalize()]
)
out = pipeline(tensor)

Script: example_transforms.py. Cutouts live under Examples → Cutout, not here.

Tables / light curves¶

from torchfits.transforms import AsymmetricSigmaClip, SigmaClip

clean = AsymmetricSigmaClip(n_low=3.0, n_high=3.0)(flux)
sym = SigmaClip(n_sigma=3.0)(flux)

Lupton asinh RGB (real SDSS)¶

lupton_rgb matches Astropy's Lupton asinh mapping (per-pixel peak clip). On this reprojected SDSS g/r/i sample the object fluxes are faint, so the gallery uses Q=8, stretch=0.15 (Astropy's default stretch is 5; tutorials often use 0.5). Astropy convention maps the reddest band (i) to the R channel:

from torchfits.transforms import lupton_rgb

rgb = lupton_rgb(r=i, g=r, b=g, Q=8.0, stretch=0.15)

Script: example_lupton_rgb_sdss.py.

CLI synthetic RGB (no network) is under CLI recipes.