igl.models

scikit-learn-compatible estimators: classifier, regressor, autoencoder.

igl.models.classifier.IGLClassifier

Bases: _BaseIGLEstimator[CrossEntropyLoss], ClassifierMixin

scikit-learn-compatible classifier built on :class:igl.IGLModule.

Parameters:

Name	Type	Description	Default
max_dim	int	Latent dimension d_max for Matryoshka truncation.	16
n_anchors	int | None	Anchor count. None defers to config or default 64.	None
n_scales	int | None	Scale count. None defers to config or 4.	None
operator	OperatorNameLike | None	Kernel operator. None defers to config or OperatorName.GAUSSIAN.	None
normalize	NormalizeModeLike | None	Φ-normalization mode.	None
encoder_hidden	int | tuple[int, ...] | None	Encoder hidden shorthand (int for uniform width, tuple for per-layer widths).	None
encoder_depth	int | None	Encoder depth shorthand.	None
config	IGLConfig | None	Optional :class:igl.IGLConfig. Explicit kwargs override it.	None
random_state	int | None	Optional integer seed for reproducible training.	None
validation_fraction	float | None	If set, hold out this fraction of training data for early-stopping and dimension-curve evaluation.	0.2

Attributes:

Name	Type	Description
classes_		Sorted unique training labels.
n_features_in_		Ambient input dimension seen during fit.
module_		The underlying :class:igl.IGLModule.
history_		:class:igl.TrainingHistory from the trainer.
dimension_curve_		{k: error_rate} from :func:igl.eval_dimension_curve post-fit.
effective_dimension_		Detected elbow k (the discovered d_eff).
scaler_		:class:sklearn.preprocessing.StandardScaler used on inputs.

fit(x, y)

Fit the classifier on (x, y).

predict(x)

Predict class labels for x.

predict_proba(x)

Predict class probabilities for x.

igl.models.regressor.IGLRegressor

Bases: _BaseIGLEstimator[MSELoss], RegressorMixin

scikit-learn-compatible regressor.

Supports scalar and multi-output regression. Targets y may be a 1-D array ([n_samples]) for scalar regression, or 2-D ([n_samples, n_outputs]) for multi-output. The output dimension is inferred from y in :meth:fit.

Args mirror :class:IGLClassifier except for validation_fraction.

Attributes:

Name	Type	Description
n_features_in_		Ambient input dimension seen during fit.
n_outputs_		Number of output dimensions inferred from y.
module_		The underlying :class:igl.IGLModule.
history_		:class:igl.TrainingHistory from the trainer.
dimension_curve_		{k: mse} from :func:igl.eval_dimension_curve.
effective_dimension_		Discovered d_eff.
scaler_		:class:sklearn.preprocessing.StandardScaler.

fit(x, y)

Fit the regressor on (x, y).

predict(x)

Predict targets for x. Returns 1-D if fit with scalar targets.

igl.models.autoencoder.IGLAutoencoder

Bases: _BaseIGLEstimator[MSELoss], TransformerMixin

Train an IGL model with y = x (reconstruction).

The :meth:fit method ignores its y argument and uses x as both input and target.

Attributes:

Name	Type	Description
n_features_in_		Ambient input dimension.
n_outputs_		Equal to n_features_in_.
module_		The underlying :class:igl.IGLModule.
history_,	(dimension_curve_, effective_dimension_, scaler_)	same as :class:IGLRegressor.

fit(x, y=None)

Fit the autoencoder. y is ignored; the scaled x is the target.

Reconstruction is therefore measured in the StandardScaler-scaled feature space. :meth:reconstruct inverse-transforms back to the original feature space for users who need real-space outputs.

reconstruct(x)

Inverse-transform the reconstruction back to the original feature space.

transform(x)

Return the reconstructed input (in the scaled feature space).

igl.IGLDistiller

Bases: _BaseIGLEstimator[WhitenedMSELoss]

Distill states into a Matryoshka latent under a metric geometry.

The pipeline mirrors metric-weighted distillation: inputs are centered per feature and scaled by a single scalar std; the reconstruction target is the raw state, whitened inside the loss by metric's square root so plain least squares optimizes the second-order expansion of the geometry the metric encodes. :meth:reconstruct undoes the whitening, returning states in the original space.

Parameters:

Name	Type	Description	Default
max_dim	int	Matryoshka latent width.	16
metric	Tensor | None	Symmetric PSD metric [C, C] over the state space (e.g. :func:igl.whitening.fisher_pullback), or None for the identity metric (variance-faithful distillation).	None
clamp	float	Eigenvalue floor for the metric square root.	1e-06
config	IGLConfig | None	Optional :class:igl.IGLConfig.	None
device	str | device | None	Optional torch device for fitting.	None
random_state	int | None	Seed scoping module construction and training.	None
validation_fraction	float | None	Held-out fraction for early stopping.	0.2

Attributes:

Name	Type	Description
module_		The fitted :class:igl.IGLModule (post-fit).
history_		Training history (post-fit).
dimension_curve_		Loss-versus-dimension curve in the whitened geometry (post-fit).
effective_dimension_		Elbow of the dimension curve (post-fit).
whitener_		The fitted :class:igl.whitening.TargetWhitener (post-fit).
input_mean_		Per-feature input mean [C] (post-fit).
input_std_		Scalar input std (post-fit).

fit(x, y=None)

Fit the distiller on states [N, C] (the states are the target).

Parameters:

Name	Type	Description	Default
x	NDArray[floating]	States to distill.	required
y	None	Ignored; present for sklearn API compatibility.	None

Returns:

Type	Description
IGLDistiller	self, for chaining.

project(x, *, k=None)

Map states to chart coordinates, optionally truncated to k.

Parameters:

Name	Type	Description	Default
x	NDArray[floating]	States [N, C].	required
k	int | None	Keep only the first k Matryoshka coordinates; None keeps all max_dim.	None

Returns:

Type	Description
NDArray[floating]	Coordinates [N, k or max_dim].

reconstruct(x, *, k=None)

Reconstruct states through the bottleneck, back in the original space.

Parameters:

Name	Type	Description	Default
x	NDArray[floating]	States [N, C].	required
k	int | None	Read the reconstruction from only the first k coordinates (a tight-budget read-out); None uses all max_dim.	None

Returns:

Type	Description
NDArray[floating]	Reconstructed states [N, C].