Laplace API¶

deepuq.methods.laplace ¶

LaplaceWrapper ¶

Fit a Laplace approximation around a MAP-trained model.

Supported Hessian structures: - diag - fisher_diag - lowrank_diag - block_diag - kron - full

Example: la = LaplaceWrapper(model, likelihood='classification', hessian_structure='diag') la.fit(dataloader, prior_precision=1.0) probs, var = la.predict(x)

predict_uq ¶

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

Return standardized uncertainty fields without changing legacy predict().