ASKabalan/jax-fli-experiments

jax-fli experiments

Data, samples, and reference catalogs for the jax-fli forward-modelling experiments. Each experiment is exposed as one or more HuggingFace dataset configs; load a config with datasets.load_dataset.

Experiment 00 — CosmoGrid reference

A single CosmoGrid simulation (cosmo_000001) packaged as jax-fli Catalog parquet files, used as reference truth for the lensing / masked-shear experiments.

Config	Field	Shape	NSIDE	Description
00-cosmogrid-density	SphericalDensity	(56, 50331648)	2048	Density (rho = N/V, float32) lightcone, 56 per-shell parquet rows out to z <= 1.6 (DES Y3 source depth, see exp 06). One config — load by streaming: concatenating all 56 shells into a single Arrow column overflows its INT32 list-offset.
00-cosmogrid-kappa	SphericalKappaField	(4, 3145728)	512	CosmoGrid's own Stage-3 forecast convergence maps, 4 tomographic bins.
00-cosmogrid-born-des	SphericalKappaField	(4, 50331648)	2048	jax-fli Born-approximation convergence from the nside-2048 density lightcone, DES Y3 source n(z), 4 tomographic bins.
00-cosmogrid-born-s3	SphericalKappaField	(4, 50331648)	2048	jax-fli Born-approximation convergence from the same lightcone, CosmoGrid Stage-3 source n(z), 4 tomographic bins.

import jax_fli as jfli
from datasets import load_dataset
from huggingface_hub import snapshot_download
from jax_fli.io import Catalog

REPO = "ASKabalan/jax-fli-experiments"

# --- Forecast convergence (kappa), nside 512: one small config, plain (non-streaming) load ---
kappa = Catalog.from_dataset(
    load_dataset(REPO, "00-cosmogrid-kappa", split="train").with_format("numpy")
).field[0]  # SphericalKappaField (4, npix)

# --- Born convergence, nside 2048, 4 bins: one ~0.8 GB file per config, plain load ---
born_des = Catalog.from_dataset(
    load_dataset(REPO, "00-cosmogrid-born-des", split="train").with_format("numpy")
).field[0]  # SphericalKappaField (4, npix) — DES Y3 source n(z); use "00-cosmogrid-born-s3" for Stage-3

# --- Density lightcone, nside 2048, 56 shells: MUST be streamed ---
# A plain load_dataset concatenates the 56 (npix,) shell rows into one Arrow column and overflows its
# INT32 list-offset. Snapshot the per-shell parquets (+ README, which carries the config glob), then
# stream from the local dir and stack the shells into the (56, npix) lightcone.
local = snapshot_download(
    REPO,
    repo_type="dataset",
    allow_patterns=["00-cosmogrid/density/cosmogrid_density_nside2048_shell*.parquet", "README.md"],
)
shells = [
    Catalog.from_dataset(row).field[0]
    for row in load_dataset(local, "00-cosmogrid-density", split="train", streaming=True)
]
density = jfli.SphericalDensity.stack(shells)  # SphericalDensity (56, npix) float32

Experiment 01 — Resolution convergence

PM mesh-resolution study (512³–3072³, fixed box/steps/cosmology, independent IC realisation per mesh) painting 10 comoving HEALPix shells (nside 512, z≈0.017–0.35). See docs/5-experiments/01-resolution-convergence.

Four configs, each bundling all five resolutions m∈{512,1024,2048,2560,3072} as rows (the 01-resolution-density config additionally carries pencil-decomposition variants at m2560 and m3072, tagged with a _pencil suffix on each row's name):

Config	Field (per row)	Shape	Rows
01-resolution-density	SphericalDensity (DENSITY, float64)	(10, 3145728)	7 density maps (5 resolutions + 2 pencil-decomp: m2560, m3072)
01-resolution-spectra	PowerSpectrum (ANGULAR_CL, per_plane δ, full-sky)	(10, 1536)	5 raw spectra
01-resolution-deconvolved-spectra	PowerSpectrum, pixel-window-deconvolved	(10, 1536)	5 deconvolved spectra
01-resolution-perf	jax-hpc-profiler timings, 19 cols (perf_pm.parquet)	—	5 rungs

The raw perf_pm.csv, the per-rung perf markdown reports, and the m<m>.args.log launch logs sit alongside it in 01-resolution/perf/ (plain files). Rows within a parquet config are ordered by filename, not mesh, so index them by each field's mesh_size:

from datasets import load_dataset
from jax_fli.io.catalog import Catalog

REPO = "ASKabalan/jax-fli-experiments"
cat = Catalog.from_dataset(
    load_dataset(REPO, "01-resolution-spectra", split="train").with_format("numpy")
)
spectra = {int(f.mesh_size[0]): f for f in cat.field}   # {512: PowerSpectrum(10,1536), 1024: ..., ...}