LaplaceApprox#

class torch_uncertainty.post_processing.LaplaceApprox(task, model=None, weight_subset='last_layer', hessian_struct='kron', pred_type='glm', link_approx='probit', optimize_prior_precision=True)[source]#

Laplace approximation for uncertainty estimation.

This class is a wrapper of Laplace classes from the laplace-torch library.

Parameters:

  • task (Literal['classification', 'regression']) – task type.

  • model (Module | None) – model to be converted.

  • weight_subset (str) – subset of weights to be considered. Defaults to "last_layer".

  • hessian_struct (str) – structure of the Hessian matrix. Defaults to "kron".

  • pred_type (Literal['glm', 'nn']) – type of posterior predictive, see the Laplace library for more details. Defaults to "glm".

  • link_approx (Literal['mc', 'probit', 'bridge', 'bridge_norm']) – how to approximate the classification link function for the "glm". See the Laplace library for more details. Defaults to “probit”.

  • optimize_prior_precision (bool) – whether to optimize the prior precision. Defaults to True.

References

[1] Daxberger et al. Laplace Redux - Effortless Bayesian Deep Learning. In NeurIPS 2021.