nirs4all.data.multimodal module

Adapt IO-owned multimodal cohorts to the Python DAG scientific host.

The adapter keeps raw source buffers in IO. It reuses the existing target, metadata and observation index interfaces, without inventing dense features.

class nirs4all.data.multimodal.MultimodalSpectroDataset(cohort: Any)

Bases: SpectroDataset

DAG host view of a nirs4all_io.MultimodalDataset.

content_hash(source_index: int | None = None, sample_rows: list[int] | None = None) → str

Fingerprint typed source buffers, optionally restricted to a cohort’s rows.

data_plan(source_ids: list[str]) → dict[str, Any]

data_schema(source_ids: list[str], sample_ids: list[str]) → dict[str, Any]

Translate the IO descriptors into DAG-ML-Data’s typed source schema.

features_sources() → int

Get number of feature sources.

groups_for_rows(rows: list[int]) → ndarray | None

headers(src: int = 0) → list[str]

Get feature headers for a source.

property independent_unit_ids: tuple[str, ...] | None

The copied explicit unit labels, independently of split groups.

is_multi_source() → bool

Check if dataset has multiple feature sources.

property n_sources: int

Get number of feature sources.

property num_features: list[int]

Fixed raw dimensions; variable-length series contribute channel counts.

property num_samples: int

Get total number of samples.

property repetition_ids: tuple[str, ...] | None

The copied explicit repetition labels, independently of observations.

source_layout(source_ids: list[str]) → dict[str, Any]

source_name(index: int) → str

split_features(splitter: Any, pool: list[int]) → ndarray

Expose row indices only to explicitly index/label/group-based splitters.

x(selector: Any, layout: Any = '2d', concat_source: bool = True, include_augmented: bool = True, include_excluded: bool = False) → Any

Get feature data with automatic augmented sample aggregation.

Parameters:

• selector – Filter criteria (partition, group, branch, etc.)

• layout – Output layout (“2d” or “3d”)

• concat_source – If True, concatenate multiple sources along feature axis

• include_augmented – If True, include augmented versions of selected samples. If False, return only base samples (origin=null). Default True for backward compatibility.

• include_excluded – If True, include samples marked as excluded. If False (default), exclude samples marked as excluded=True. Use True when transforming ALL features (e.g., preprocessing).

Returns:

Feature data array(s)

Example

>>> # Get all train samples (base + augmented)
>>> X_train = dataset.x({"partition": "train"})
>>> # Get only base train samples (for splitting)
>>> X_base = dataset.x({"partition": "train"}, include_augmented=False)
>>> # Get all features including excluded (for transformations)
>>> X_all = dataset.x({"partition": "train"}, include_excluded=True)

x_rows(sample_ints: list[int], layout: Any = '2d', concat_source: bool = True) → Any

Get feature data for an EXACT set of stored rows, addressed by their own sample int.

Unlike x(), this bypasses the two-phase base-keyed selection: each sample int (base OR augmented) addresses its own storage row directly, and rows are returned in request order. The dag-ml resolver uses it to materialize a view whose ids may include augmented children, which the base-keyed x path filters out.

Parameters:

• sample_ints – Stored-row sample ints to retrieve, in the desired output order.

• layout – Output layout (“2d” or “3d”).

• concat_source – If True, concatenate multiple sources along the feature axis.

Returns:

Feature data array(s), one row per requested sample int, in request order.

Example

>>> # Materialize a train view containing base + augmented rows, in a given order
>>> X = dataset.x_rows([0, 6, 7, 1, 8, 9])