omniloader.data.splits

Reproducible train/val/test index splitting.

split_indices() partitions range(n) into named splits by ratio, with an optional stratification label so each split preserves the class proportions. save_split_info() / load_split_info() persist the result (plus the seed and ratios) so the exact split can be reused across runs.

Attributes

DEFAULT_NAMES

Functions

split_indices(, seed, stratify, names, list[int]])

Partition range(n) into named, reproducible splits.

save_split_info(→ None)

Save splits (and optional metadata like seed/ratios) to JSON.

load_split_info(→ dict[str, list[int]])

Load the splits mapping saved by save_split_info().

Module Contents

omniloader.data.splits.DEFAULT_NAMES = ('train', 'valid', 'test')
omniloader.data.splits.split_indices(n: int, ratios: collections.abc.Sequence[float] = (0.8, 0.1, 0.1), seed: int = 0, stratify: collections.abc.Sequence[int] | None = None, names: collections.abc.Sequence[str] = DEFAULT_NAMES) → dict[str, list[int]][source]

Partition range(n) into named, reproducible splits.

Parameters:

  • n – Number of samples to split.

  • ratios – Relative sizes per split (need not sum to one).

  • seed – Seed for the shuffle.

  • stratify – Optional per-index class labels of length n; when given, each class is split by ratios so proportions are preserved.

  • names – Split names, aligned with ratios.

Returns:

Mapping of split name to a sorted list of indices; the splits are disjoint and cover range(n).

Raises:

ValueError – If ratios/names mismatch or stratify has the wrong length.

omniloader.data.splits.save_split_info(splits: dict[str, list[int]], path: str | pathlib.Path, meta: dict[str, Any] | None = None) → None[source]

Save splits (and optional metadata like seed/ratios) to JSON.

Parameters:

  • splits – The split mapping from split_indices().

  • path – Destination .json file.

  • meta – Optional extra fields (e.g. {"seed": 0, "ratios": [...]}).

omniloader.data.splits.load_split_info(path: str | pathlib.Path) → dict[str, list[int]][source]

Load the splits mapping saved by save_split_info().

Parameters:

path – Path to a JSON file written by save_split_info().

Returns:

The split-name to index-list mapping.