API reference

Every public symbol the library exposes. Generated from docstrings via mkdocstrings.

Top-level

The flat surface re-exported by igl:

• Configuration — EncoderConfig, KernelConfig, MatryoshkaConfig, IGLConfig.
• Building blocks — MLPEncoder, LinearEncoder, GreenKernel, IGLModule.
• Training — MatryoshkaTrainer, TrainingHistory, CrossEntropyLoss, MSELoss, direct_solve_weights, normalize_phi.
• Matryoshka & dimension discovery — UniformSampler, PowerLawSampler, eval_dimension_curve, detect_elbow, d_eff_from_curve, compare_d_eff, DimensionComparison.
• Kernel registry — Operator, register_operator, get_operator, list_operators.
• sklearn estimators — IGLClassifier, IGLRegressor, IGLAutoencoder.
• Types & enums — OperatorName, SamplingMode, NormalizeMode, NormType, ActivationType, EncoderKind, SchedulerType, LossStrategy, MatryoshkaSampler, OperatorFn, EncoderProtocol.
• Exceptions — IGLError, IGLConfigError, IGLConvergenceError, IGLDependencyError, IGLNotFittedError.

Subpackages

• igl.core — encoder, kernel, solver, trainer, losses.
• igl.kernels — operator zoo + registration API.
• igl.matryoshka — samplers + dimension curve helpers.
• igl.models — sklearn-compatible estimators.
• igl.metrics — cross-task d_eff comparison + elbow detectors.
• igl.nn — bare PyTorch IGLModule.
• igl.spd — Riemannian extension (AIRM, log-Eig, orthogonality, reconstruction).
• igl.data — synthetic data generators.
• igl.viz — optional matplotlib helpers.

Source layout

src/igl/
├── __init__.py          # flat public surface
├── config.py            # frozen dataclasses
├── exceptions.py        # IGLError + subclasses
├── types.py             # Protocols + StrEnums + Literal companions
├── core/                # encoder, kernel, solver, trainer, losses, normalization
├── kernels/             # 9 log-space operators + registry
├── matryoshka/          # samplers + dimension curve
├── models/              # sklearn estimators
├── metrics/             # d_eff comparison + elbow detectors
├── nn/                  # IGLModule
├── spd/                 # AIRM + log-Eig + orthogonality + reconstruction
├── data/                # synthetic generators
└── viz/                 # matplotlib helpers (gated behind [viz])