mahalanobis

probly.method.mahalanobis(base: Predictor[In, Out], feature_nodes: FeatureNodes = None, input_preprocessing_eps: float = 0.0) → MahalanobisPredictor[In, Out][source]

Turn a classifier into a Mahalanobis-distance OOD detector.

Based on [LLLS18]. The final linear head is stripped to expose penultimate features, class-conditional Gaussians with a tied covariance are fitted on those features (and, optionally, on extra intermediate layers), and the per-layer Mahalanobis confidence is combined into a single out-of-distribution score by logistic regression.

The returned predictor still needs its Gaussian parameters fitted after training via fit_mahalanobis_heads (and, optionally, fit_combiner to calibrate the multi-layer weights on in- vs out-of-distribution data).

Parameters:

  • base – Base logit classifier to be transformed.

  • feature_nodes – Optional names of intermediate submodules (as returned by named_modules) whose global-average-pooled outputs provide additional feature layers for the ensemble. When None only the penultimate features are used, recovering the single-layer detector.

  • input_preprocessing_eps – Magnitude of the FGSM-style input perturbation applied at inference to sharpen the in/out separation. 0 disables preprocessing.

Returns:

The transformed Mahalanobis predictor.

Examples using probly.method.mahalanobis

Mahalanobis OOD on Two Moons

Mahalanobis OOD on Two Moons

Mahalanobis OOD on MNIST

Mahalanobis OOD on MNIST