nirs4all.pipeline.analysis.splitter_config module

Splitter (CV) configuration extraction from stored pipeline configs.

Answers “which cross-validation setup did this pipeline use?” from the expanded_config JSON the workspace store persists per pipeline — the canonical step list written by WorkspaceStore.begin_pipeline().

This inference previously lived in the nirs4all-studio HTTP layer (BV-07 in its 2026-06-05 tech-debt closeout): the studio hand-parsed the library’s own storage format and hand-listed splitter names. The parsing of library-written data belongs to the library; UIs keep only their display vocabulary.

Recognition order: nirs4all’s own splitter registry, sklearn model_selection splitters (by module path or class name), then a token fallback (split/fold/loo/holdout) for custom classes.

class nirs4all.pipeline.analysis.splitter_config.SplitterConfig(splitter_class: str, reference: str, n_splits: int | None = None, shuffle: bool | None = None, random_state: int | None = None, test_size: float | None = None, group_by: str | None = None)[source]

Bases: object

Canonical CV-splitter configuration recovered from a stored pipeline.

Variables:

  • splitter_class (str) – Leaf class name (e.g. "KFold").

  • reference (str) – The raw step reference as stored.

  • n_splits (int | None) – Fold count when the splitter declares one.

  • shuffle (bool | None) – Shuffle flag when declared.

  • random_state (int | None) – Seed when declared.

  • test_size (float | None) – Test fraction when declared (shuffle/holdout splitters).

  • group_by (str | None) – Grouping column when declared (group-aware splitters).

group_by: str | None = None
n_splits: int | None = None
random_state: int | None = None
reference: str
shuffle: bool | None = None
splitter_class: str
test_size: float | None = None
nirs4all.pipeline.analysis.splitter_config.extract_splitter_config(expanded_config: Any) → SplitterConfig | None[source]

Recover the CV-splitter configuration from a stored pipeline config.

Scans the canonical step list for the first splitter step and returns its class plus canonical parameters. Returns None when the pipeline has no recognizable splitter (e.g. predict-only chains).

nirs4all.pipeline.analysis.splitter_config.is_splitter_reference(reference: str) → bool[source]

Whether a step reference denotes a CV/data splitter.

Checks nirs4all’s splitter registry, the sklearn model_selection splitter set (by class name or module path), then falls back to name tokens for custom splitters.

nirs4all.pipeline.analysis.splitter_config.parse_expanded_config_steps(expanded_config: Any) → list[Any][source]

Return the canonical step list from a stored expanded_config.

Accepts the raw column value (JSON string, dict with a pipeline key, or already-parsed list) and normalizes to a list of steps.