Feature builders

Bases: BaseEstimator, TransformerMixin

FeatureUnion-style composition of named transformers.

Each child is fit on the same (X, y) and outputs are concatenated column-wise.

Parameters:

Name	Type	Description	Default
transformers	list[tuple[str, TransformerMixin]] | None	List of (name, transformer) pairs. Each transformer must implement fit/transform (sklearn TransformerMixin convention).	None

Bases: BaseEstimator, TransformerMixin

Return X unchanged (the x component of phi_{z^N}(x)).

Bases: BaseEstimator, TransformerMixin

Estimate log-density log q(x | training data).

Parameters:

Name	Type	Description	Default
method	DensityMethod	"mahalanobis" — diagonal Gaussian MLE (Lee et al. 2018; Sanderink, 2026 default). "knn" — -log(d_k + eps) where d_k is the k-th NN distance. "kde" — sklearn KernelDensity with Gaussian kernel. "flow" — normalizing flow (requires pip install "deup[torch]").	'mahalanobis'
k	int	Neighbors for method="knn".	5
bandwidth	float	Bandwidth for method="kde".	1.0
var_floor	float	Minimum variance per dimension for mahalanobis (dead-ReLU dimensions).	1e-06
eps	float	Stabilizer for knn log-density.	1e-08

Bases: BaseEstimator, TransformerMixin

Estimate log predictive variance log V(x).

Parameters:

Name	Type	Description	Default
method	VarianceMethod	"ensemble" — bootstrap variance of a cloned base estimator's predictions. "gp" — GP variance (requires pip install "deup[torch]"; not yet implemented — raises at fit).	'ensemble'
estimator	Any	Base model for method="ensemble" (cloned per bootstrap replicate).	None
n_estimators	int	Number of bootstrap replicas.	10
random_state	int | None	Seed for bootstrap subsampling.	None
eps	float	Added inside log(var + eps) for numerical stability.	1e-08

Bases: BaseEstimator, TransformerMixin

k-th nearest-neighbor distance from x to the training manifold.

Parameters:

Name	Type	Description	Default
k	int	Which neighbor distance to return (1 = nearest). Useful as a covariate-shift proxy complementary to DensityFeature.	5

Bases: BaseEstimator, TransformerMixin

Binary indicator: 1 if x was in the training set, else 0.

Parameters:

Name	Type	Description	Default
atol	float	Tolerance for floating-point duplicate detection (exact match when both 0).	0.0
rtol	float	Tolerance for floating-point duplicate detection (exact match when both 0).	0.0

Bases: BaseEstimator, TransformerMixin

Local average of training residual magnitudes |y - f(x)|.

At fit(X, y) the base estimator is cloned and fit, and absolute residuals are stored. At transform(X) (inference, no y), returns the mean residual magnitude among k nearest training neighbors — a local error prior.

Parameters:

Name	Type	Description	Default
estimator	Any	Base model f (any sklearn fit/predict object). Cloned at fit.	required
k	int	Neighbors for smoothing.	5