Laplace API¶

This page documents the public Laplace approximation wrapper used throughout Deep-UQ. The notes below explain the public control variables and workflow; the generated section contains the exact signatures and source docstrings.

Public objects¶

• LaplaceWrapper

Parameter and variable conventions¶

For hessian_structure="kron" and "full", Deep-UQ now prefers the legacy laplace-torch backend when that optional dependency is installed. Install it with:

pip install "uqdeepnn[laplace]"

Without that extra, those two structures fall back to the native Deep-UQ implementations.

Workflow expectations¶

1. train the base model to a MAP solution with ordinary optimization
2. construct LaplaceWrapper(model, ...)
3. call fit(train_loader, prior_precision=...)
4. call predict(...) or predict_uq(...)

predict_uq(...) cannot be called before fit(...).

Input and output shapes¶

• fit(...) expects an iterable of (inputs, targets) minibatches compatible with the wrapped model.
• regression predict(...) returns (mean, var) with the same trailing shape as one model prediction.
• classification predict(...) returns (probs, probs_var_or_none) with shape [batch, n_classes].

UQResult mapping¶

predict_uq(...) returns:

• regression: mean, epistemic_var, optional aleatoric_var, total_var
• classification: probs, optional probs_var, and metadata describing the chosen backend

Common preconditions and failure modes¶

• unsupported hessian_structure raises ValueError
• full curvature over subset_of_weights="all" may be rejected if full_max_params is exceeded
• calling predict(...) or predict_uq(...) before fit(...) raises RuntimeError
• regression backends must return predictive variance; otherwise predict_uq(...) raises RuntimeError

Minimal example¶

la = LaplaceWrapper(
    model,
    likelihood="regression",
    hessian_structure="block_diag",
    subset_of_weights="last_layer",
)
la.fit(train_loader, prior_precision=10.0)
uq = la.predict_uq(x_test, n_samples=32)

Related docs¶

• UQ API Conventions
• Types API
• Laplace Approximation theory

>>> la = LaplaceWrapper(model, likelihood="classification", hessian_structure="diag")
>>> la.fit(train_loader, prior_precision=1.0)
>>> probs, probs_var = la.predict(x_test)

fit(
    train_loader: Iterable,
    prior_precision: float | None = 1.0,
    **_
) -> object

predict(x: Tensor, **predict_kwargs)

predict_uq(x: Tensor, **predict_kwargs) -> UQResult

supported_hessian_structures() -> tuple[str, ...]