Skip to content

Types API¶

deepuq.types contains the shared container types used across the UQ APIs. The most important one is UQResult, which gives every predict_uq(...) helper a common return shape.

`UQResult` at a glance¶

Field	Type	Meaning
`mean`	`torch.Tensor`	predictive mean; for classifiers this often mirrors `probs`
`epistemic_var`	`torch.Tensor \\| None`	uncertainty due to model/posterior uncertainty
`aleatoric_var`	`torch.Tensor \\| None`	uncertainty due to observation noise
`total_var`	`torch.Tensor \\| None`	total predictive variance
`probs`	`torch.Tensor \\| None`	predictive class probabilities
`probs_var`	`torch.Tensor \\| None`	probability-space disagreement/variance
`metadata`	`dict[str, Any]`	backend, sample-count, or likelihood metadata

Usage notes¶

Regression methods should treat mean as the primary prediction and populate variance fields where available.
Classification methods should prefer probs and probs_var; mean may mirror probs for convenience.
metadata is intentionally open-ended. Use it for non-tensor method details such as n_members, n_mc, likelihood, or hessian_structure.

UQ API Conventions

deepuq.types ¶

Shared public types for Deep-UQ outputs.

UQResult `dataclass` ¶

Standardized uncertainty output container.

Fields are method-agnostic and can be partially populated depending on the inference algorithm and task type.

Parameters:

Name	Type	Description	Default
`mean`	`Tensor`	Predictive mean tensor for regression-style outputs. For classification methods this usually stores the same value as `probs` for convenience.	required
`epistemic_var`	`Tensor \| None`	Variance attributed to model or posterior uncertainty. Present for methods that expose between-sample or between-model spread.	`None`
`aleatoric_var`	`Tensor \| None`	Variance attributed to data noise or likelihood noise. Present only for methods that model observation noise explicitly.	`None`
`total_var`	`Tensor \| None`	Total predictive variance. When both epistemic and aleatoric terms are available this should be their sum.	`None`
`probs`	`Tensor \| None`	Predictive class probabilities for classification methods.	`None`
`probs_var`	`Tensor \| None`	Probability-space variance or disagreement summary for classification methods.	`None`
`metadata`	`dict[str, Any]`	Free-form method metadata such as backend, sample count, or likelihood settings.	`dict()`