src/load_t2_material_loading_memory.py

"""Lightweight reader for T2 material-loading-memory HDF5 shards."""

from __future__ import annotations

import json
from pathlib import Path
from typing import Iterator

import h5py
import numpy as np


def load_manifest(dataset_root: str | Path) -> dict[str, object]:
    root = Path(dataset_root)
    return json.loads((root / "manifest.json").read_text(encoding="utf-8"))


def _resolve_shard(root: Path, relative_path: str) -> Path:
    candidate = root / "shards" / Path(relative_path).name
    if candidate.exists():
        return candidate
    return root.parents[1] / relative_path


def _group_to_sample(sample_id: str, group: h5py.Group) -> dict[str, object]:
    return {
        "id": sample_id,
        "case_id": str(group.attrs["case_id"]),
        "split": str(group.attrs["split"]),
        "material_model": str(group.attrs["material_model"]),
        "path_family": str(group.attrs["path_family"]),
        "parameters": json.loads(str(group.attrs["parameters_json"])),
        "metrics": json.loads(str(group.attrs["metrics_json"])),
        "history": {name: np.asarray(group[name]) for name in group.keys()},
    }


def iter_trajectories(
    dataset_root: str | Path,
    *,
    split: str | None = None,
    material_model: str | None = None,
    path_family: str | None = None,
) -> Iterator[dict[str, object]]:
    """Yield copied histories, optionally filtered by trajectory metadata."""

    root = Path(dataset_root)
    manifest = load_manifest(root)
    for shard in manifest["shards"]:
        with h5py.File(_resolve_shard(root, shard["path"]), "r") as h5:
            for sample_id in sorted(h5.keys()):
                group = h5[sample_id]
                if split is not None and str(group.attrs["split"]) != split:
                    continue
                if (
                    material_model is not None
                    and str(group.attrs["material_model"]) != material_model
                ):
                    continue
                if (
                    path_family is not None
                    and str(group.attrs["path_family"]) != path_family
                ):
                    continue
                yield _group_to_sample(sample_id, group)


def load_trajectory(
    dataset_root: str | Path, sample_id: str | int
) -> dict[str, object]:
    target = f"{int(sample_id):05d}"
    root = Path(dataset_root)
    manifest = load_manifest(root)
    for shard in manifest["shards"]:
        if shard["first_id"] <= target <= shard["last_id"]:
            with h5py.File(_resolve_shard(root, shard["path"]), "r") as h5:
                if target in h5:
                    return _group_to_sample(target, h5[target])
    raise KeyError(f"Unknown trajectory id: {target}")