MetadataFunctionTransformer#

class chemotools.adaptation.MetadataFunctionTransformer(func: Callable[[...], ndarray], metadata: Sequence[str] = (), validate: bool = True)[source]

Bases: DocLinkMixin, TransformerMixin, BaseEstimator

Apply a callable that consumes feature data and routed metadata.

This is useful when a preprocessing function requires per-sample or per-batch auxiliary information (e.g. reference spectra, wavelength arrays) that must be threaded through a scikit-learn Pipeline via the metadata-routing API.

Parameters:

func (Callable[..., np.ndarray]) – Function applied during transform. It must accept X as its first positional argument and each requested metadata value as a keyword argument. The function is responsible for returning a numeric, 2-D np.ndarray with the same number of samples and features as X. Use chemotools.adaptation.validation.check_metadata_function() to verify this contract on representative data.
metadata (Sequence[str], default=()) – Names of the keyword arguments requested through scikit-learn metadata routing and forwarded to func. Keys passed to transform but not listed here are ignored. Every required keyword argument of func must be listed, and the corresponding value must be supplied when calling transform. The names "X" and "y" are reserved by the estimator API and cannot be requested as metadata.
validate (bool, default=True) – If True, validate X as a numeric, 2-D array during fit and transform, and require the number of features to remain consistent. This does not validate the output of func. If False, pass X to func unchanged. In both cases, fit must be called before transform.

Variables:

n_features_in (int) – Number of features seen during fit. Only set when validate=True.

Examples

>>> import numpy as np
>>> from chemotools.adaptation import MetadataFunctionTransformer
>>>
>>> def subtract_reference(X, reference):
...     return X - reference
>>>
>>> rng = np.random.default_rng(0)
>>> X = rng.normal(size=(10, 50))
>>> reference = rng.normal(size=(1, 50))
>>>
>>> mft = MetadataFunctionTransformer(
...     func=subtract_reference, metadata=("reference",)
... )
>>> X_transf = mft.fit_transform(X, reference=reference)

Notes

To route metadata through a scikit-learn Pipeline, enable metadata routing globally with sklearn.set_config(enable_metadata_routing=True). This is not required when calling this estimator directly.

The callable signature is validated during fit, but the callable is executed only during transform. Consequently, errors that depend on metadata values or callable execution are raised during transform. Use chemotools.adaptation.validation.check_metadata_function() to validate a callable eagerly on representative inputs.