ASKabalan/jax-fli-sampling

jax-fli MAP and chain outputs

Maximum a posteriori reconstructions and MCMC chains of the jax-fli forward model: 2LPT on a spherical lightcone of capped equal-volume shells, Born convergence, and a pixel likelihood on two tomographic κ maps. The notebooks in docs/3-sampling-and-inference produced the runs, and experiment 13-map-lpt2-mass-mapping draws its figures from the MAP runs. The accuracy experiments are in ASKabalan/jax-fli-experiments, and the scaling benchmarks in ASKabalan/jax-fli-scaling.

folder	content	produced by
13-LPT-MassMapping/mesh_416_DES/	MAP reconstruction of the initial conditions at 416³ from the κ maps of DES Y3 source bins 2 and 3 (1.48 galaxies/arcmin² per bin, σ_e = 0.26, 300 Adam steps)	notebook 14-LPT-MassMapping.ipynb, Part I
13-LPT-MassMapping/mesh_416_EUCLID/	the same reconstruction at Euclid IST:F noise (7.5 galaxies/arcmin² per bin, σ_e = 0.21 per component, 400 Adam steps)	notebook 14-LPT-MassMapping.ipynb, Part II
14-LPT-Sampling/	MCLMC chain over (Ω_c, σ₈) and the 128³ initial conditions (chain/), the cosmology chain with the initial conditions held fixed (chain_fixed_ic/, chain_cosmology.npz), the truth (truth/), the lensing-kernel projections of the samples (projected_ic/), summaries and figures	notebooks 16-LPT-FieldLevel-IC.ipynb and 17-LPT-FieldLevel-Cosmology.ipynb
15-inference-cosmogrid/truth/input_cg.parquet	a CosmoGrid density field (832³ in a 900 Mpc/h box), the truth of the retired inference experiment 14-inference-cosmogrid-shear	—

Each MAP run holds the static truth (true_ic, truth_kappa, observed_kappa, truth_density_lightcone, lattice_density_lightcone), eleven log-spaced frames of the optimisation in each *_evolution/ folder (IC, density lightcone, κ, and the coherence and transfer function of the IC and of κ against the truth), and metrics.json, summary.json and loss_history.npz. Every parquet is a serialised jax_fli.io.Catalog. The chain/sampling_state/ and chain_fixed_ic/warmup_state/ folders are Orbax checkpoints that resume the chains.

from pathlib import Path

from huggingface_hub import snapshot_download
from jax_fli.io import Catalog

REPO = "ASKabalan/jax-fli-sampling"
root = Path(snapshot_download(REPO, repo_type="dataset", allow_patterns=["13-LPT-MassMapping/mesh_416_DES/*.parquet"]))
truth_kappa = Catalog.from_parquet(str(root / "13-LPT-MassMapping/mesh_416_DES/truth_kappa.parquet")).field[0]