tools/make_spectra.py

import argparse


def parse_args():
    parser = argparse.ArgumentParser(
        description="Compute the angular power spectrum of a density catalog and write it as a "
        "single stacked PowerSpectrum parquet."
    )
    parser.add_argument(
        "--parquet-path",
        type=str,
        required=True,
        help="Path / glob to the density parquet(s). One example (shell) per row.",
    )
    parser.add_argument(
        "--output-path",
        type=str,
        required=True,
        help="Path to save the processed dataset.",
    )
    parser.add_argument("--codec", default="SNAPPY", help="target compression codec (default: SNAPPY)")
    parser.add_argument(
        "--name",
        type=str,
        required=True,
        help="Name (legend label) of the output dataset.",
    )
    parser.add_argument(
        "--lmax",
        type=int,
        default=1500,
        help="Maximum multipole moment. For masked (off-center observer) fields this is the "
        "underlying lmax that is then binned into nlb-width bandpowers.",
    )
    parser.add_argument(
        "--normalization",
        type=str,
        default="per_plane",
        choices=["per_plane", "global"],
        help="Normalization for the overdensity field.",
    )
    parser.add_argument(
        "--deconvolve",
        type=str,
        default="none",
        choices=["none", "ngp", "rbf"],
        help="Spherical mass-assignment-window deconvolution applied to the overdensity before "
        "the spectrum: 'none' (default), 'ngp' (HEALPix pixel window) or 'rbf' (-> 'rbf_neighbor').",
    )
    parser.add_argument(
        "--nlb",
        type=int,
        default=16,
        help="Bandpower bin width for the masked (off-center observer) estimator.",
    )
    parser.add_argument(
        "--apodization-deg",
        type=float,
        default=2.0,
        dest="apodization_deg",
        help="Apodization scale (degrees) for the observer-visibility mask (off-center only).",
    )
    return parser.parse_args()


def main():
    args = parse_args()
    import jax
    jax.config.update("jax_enable_x64", True)

    import numpy as np
    import pyarrow.parquet as pq
    from datasets import load_dataset

    from jax_fli.io import Catalog
    from jax_fli import PowerSpectrum
    from jax_fli import units, SphericalKappaField
    from jax_fli.data.masks import build_observer_visibility_mask
    from jax_fli.summary_statistics import compute_mcm
    import jax.numpy as jnp
    from tqdm.auto import tqdm

    # Stream the dataset so big multi-shell globs do not int32-overflow on combine_chunks.
    dataset = load_dataset("parquet", data_files=args.parquet_path, split="train", streaming=True)

    deconv_method = None if args.deconvolve == "none" else {"rbf": "rbf_neighbor"}.get(args.deconvolve, args.deconvolve)

    spectras = []
    cosmology_prev = None
    mask = None
    mcm = None
    mask_built = False


    def tree_all_close(tree1, tree2, rtol=1e-5, atol=1e-8):
        return jax.tree.all(jax.tree.map(lambda x, y: jnp.allclose(x, y, rtol=rtol, atol=atol), tree1, tree2))

    for ds in tqdm(dataset , desc="Processing shells", unit=" shell"):
        cat = Catalog.from_dataset(ds)
        field , cosmology = cat.field[0], cat.cosmology[0]

        if cosmology_prev is not None and not tree_all_close(cosmology, cosmology_prev):
            raise ValueError("All shells must share the same cosmology.")

        cosmology_prev = cosmology

        # Observer-visibility mask + MCM are constant across shells (same observer/nside) -> build
        # once. supersample=1 keeps the apodized footprint memory-bounded at nside 2048 (the
        # supersample=4 default builds an nside*4=8192 grid -> ~0.8B pixels -> OOM). A center
        # observer yields the scalar 1 -> mask=None (full sky, plain Cl).
        if not mask_built:
            vis = build_observer_visibility_mask(
                field.observer_position, field.nside, apodization_scale_deg=args.apodization_deg, supersample=1
            )
            mask = None if jnp.ndim(vis) == 0 else vis
            if mask is not None:
                mcm = compute_mcm(mask, lmax=args.lmax, nlb=args.nlb, pol=False, method="healpy")
            mask_built = True

        if isinstance(field, SphericalKappaField):
            # For kappa fields no conversion is needed
            overdensity = field
        else:
            overdensity = field.to(units.OVERDENSITY, normalization=args.normalization)

        if deconv_method is not None:
            # Deconvolve at the field's native bandlimit (default 3*nside-1, which
            # satisfies s2fft's lmax >= 2*nside-1); angular_cl truncates to args.lmax.
            overdensity = overdensity.deconvolve(deconv_method)

        spectra = overdensity.angular_cl(method="healpy", lmax=args.lmax, mask=mask, mcm=mcm, nlb=args.nlb)
        # Per-shell density names differ; normalize before stacking so stack's name check passes.
        spectras.append(spectra.replace(name=args.name))

    # A single source row already yields an (S, n_ell) batched spectrum -> do not re-stack it.
    stacked = spectras[0] if len(spectras) == 1 else PowerSpectrum.stack(spectras)
    stacked = stacked.replace(name=args.name)

    spec_cat = Catalog(field=stacked, cosmology=cosmology)
    table = spec_cat.to_dataset().with_format("arrow")[:]

    pq.write_table(table, args.output_path, compression=args.codec, use_dictionary=True)


if __name__ == "__main__":
    main()