noether.data.zarr_store.benchmark

Validate and benchmark a converted ShapeNet-Car Zarr store against the .pt source.

Checks, for the test split:

  1. Equivalence — normalized per-field output of ZarrShapeNetCarDataset (full read) matches the original ShapeNetCarDataset within float16 error.

  2. Read amplification — bytes fetched for a chunk-subsampled read vs a full read, measured by instrumenting the Zarr store.

Run:

uv run python -m noether.data.zarr_store.benchmark         --pt-root /nfs-gpu/research/datasets/shapenet_car         --zarr-root /nfs-gpu/research/datasets/shapenet_car/zarr_store

Attributes

FIELDS

NORMALIZERS

Functions

check_equivalence(pt_root, zarr_root, num_samples)

Compare normalized fields of the original and Zarr datasets (full read).

check_read_amplification(zarr_root, num_volume_points)

Measure bytes fetched for a chunk-subsampled read vs a full read.

main()

Module Contents

noether.data.zarr_store.benchmark.FIELDS = ['surface_position', 'surface_pressure', 'surface_normals', 'volume_position',...
noether.data.zarr_store.benchmark.NORMALIZERS
noether.data.zarr_store.benchmark.check_equivalence(pt_root, zarr_root, num_samples)

Compare normalized fields of the original and Zarr datasets (full read).

Parameters:

  • pt_root (str)

  • zarr_root (str)

  • num_samples (int)

Return type:

None

noether.data.zarr_store.benchmark.check_read_amplification(zarr_root, num_volume_points)

Measure bytes fetched for a chunk-subsampled read vs a full read.

Parameters:

  • zarr_root (str)

  • num_volume_points (int)

Return type:

None

noether.data.zarr_store.benchmark.main()
Return type:

None