PyTorch / TorchUncertainty integration

For deep learning workflows (Lightning training loops, vision classifiers), DEUP is available as a post-processing method in TorchUncertainty.

When to use which package

Need	Package
sklearn / LightGBM / tabular	deup (this repo)
Time-series / finance panels	deup — CrossSectionalDEUP, purged walk-forward
PyTorch / Lightning / vision	TorchUncertainty DEUP post-processor
Both	Install both; cite Lahlou et al. (2023) for the method

TorchUncertainty usage

from torch_uncertainty.post_processing import DEUP

deup = DEUP(task="classification", model=trained_model, n_folds=5)
deup.fit(calibration_dataloader)
uncertainty = deup(batch)  # g(x) >= 0

With OOD evaluation in ClassificationRoutine:

from torch_uncertainty.routines import ClassificationRoutine

baseline = ClassificationRoutine(
    num_classes=10,
    model=model,
    post_processing=deup,
    ood_criterion="deup",
    eval_ood=True,
)

Tutorial: TorchUncertainty DEUP tutorial

Upstream PR: torch-uncertainty#313

Method credit

DEUP is due to Lahlou et al. (2023, TMLR). TorchUncertainty hosts the PyTorch post-processing integration; this package remains the home for sklearn-compatible and time-series-correct DEUP.