API reference

All functions accept a fitted supported estimator. See semantics for output axes, target selection, errors, and classification scales.

skgrad.supports(model: object) → bool

Return whether skgrad has an analytic backend for model.

skgrad.gradient_properties(model: object) → GradientProperties

Return computational properties of a supported model’s Jacobian.

skgrad.model_output(model: object, X: object) → NDArray[floating]

Return predictions for regressors or raw scores/logits for classifiers.

skgrad.input_gradient(model: object, X: object, target: int | None = None) → NDArray[floating]

Return gradients for one selected scalar model output.

skgrad.input_jacobian(model: object, X: object) → NDArray[floating]

Return input Jacobians with shape (samples, outputs, features).

skgrad.value_and_jacobian(model: object, X: object) → GradientResult

Return model values and analytic input Jacobians.

Values have shape (n_samples, n_outputs). Jacobians have shape (n_samples, n_outputs, n_features). Classifiers return raw decision scores or logits rather than probabilities.

class skgrad.GradientResult(values: NDArray[floating], jacobian: NDArray[floating])

Model values and their input Jacobians.

class skgrad.GradientProperties(constant_jacobian: bool, exact_quadrature_steps: int | None)

Properties that downstream gradient consumers may optimize around.

Pipeline feature-space views

skgrad.pipeline_view(model: object, *, after: str | None = None) → PipelineView

Select the inputs or an intermediate feature space of a fitted Pipeline.

after=None keeps original pipeline inputs. Otherwise give a preprocessing step name, or a nested path such as 'preprocess__scale'. The view transforms original observations with the fitted prefix and exposes the remaining predictor as view.model. Prefix steps must be supported continuous transformations. Suffix backend eligibility is checked by the caller.

class skgrad.PipelineView(source_model: Pipeline, model: object, prefix: Sequence[object], after: str | None)

A fitted predictor viewed after a named preprocessing step.

Create with skgrad.pipeline_view(). model accepts transformed inputs. Convenience gradient methods accept original pipeline inputs and differentiate in the selected space. Fitted steps are shared, not refitted or copied; do not refit the source pipeline while using this view.

get_feature_names_out(input_features: Sequence[str] | None = None) → ndarray

Names of coordinates in the selected feature space.

input_gradient(X: object, target: int | None = None) → ndarray

Differentiate in selected coordinates, accepting original inputs.

input_jacobian(X: object) → ndarray

Return all output derivatives in selected coordinates.

model_output(X: object) → ndarray

Evaluate model outputs from original pipeline inputs.

transform(X: object) → ndarray

Map original pipeline inputs to the selected feature space.

value_and_jacobian(X: object) → GradientResult

Return outputs and selected-coordinate Jacobians.