deup.torch

Torch implementation of Direct Epistemic Uncertainty Prediction (DEUP).

Classes

ErrorPredictionHead	MLP that predicts \(\log_{10}\) of the per-sample loss.
LogDUEVariance	SVGP posterior variance as a distance-aware model-uncertainty proxy.
LogGMMDensity	Per-class GMM log-density of encoder features.
LogMAFDensity	Masked Autoregressive Flow log-density of encoder features.
LogMCDropoutVariance	MC-Dropout variance of the classifier as a model-uncertainty proxy.
StationarizingFeatureProvider	Base class for DEUP stationarizing feature providers.
TorchDEUPPredictor	Torch implementation of a DEUP predictor.
TorchDEUPRepresentation	DEUP representation backed by torch tensors.