Out-of-Distribution Detection

The OOD module provides methods for detecting when a model encounters data that is significantly different from its training distribution.

Base Class

OODDetector(model)

Abstract base class for out-of-distribution (OOD) detection methods.

Score-based Methods

MSP(model)	Maximum-Softmax-Probability (Hendrycks & Gimpel, 2017).
MaxLogit(model)	MaxLogit OOD detection (Hendrycks et al., 2019).
Energy(model[, temperature])	Energy-based score (Liu et al., NeurIPS 2020).
ODIN(model[, temperature, epsilon])	ODIN: Out-of-Distribution detector for Neural Networks (Liang et al., ICLR 2018).

Distance-based Methods

Mahalanobis(model[, layer_name])	Feature-space Mahalanobis (Lee et al., NeurIPS 2018).
KNN(model[, k, layer_name])	KNN-based OOD detection (Sun et al., NeurIPS 2022).

Training Methods

mixup_data(x, y[, alpha])	Apply mixup data augmentation.
mixup_criterion(criterion, pred, y_a, y_b, lam)	Compute mixup loss.
OutlierExposureLoss([lambda_oe])	Outlier Exposure (OE) for improved OOD detection.
EnergyRegularizedLoss([lambda_energy, margin])	Energy-based regularization for OOD detection.
CutMix([alpha])	CutMix augmentation for improved robustness and OOD detection.

Metrics

auroc(id_scores, ood_scores)	Compute AUROC for OOD detection.
fpr_at_tpr(id_scores, ood_scores[, tpr])	Compute False Positive Rate at a target True Positive Rate.
detection_accuracy(id_scores, ood_scores[, ...])	Compute detection accuracy at a given ID acceptance-rate threshold.

Visualization

plot_roc(id_scores, ood_scores[, label, ax])	Plot ROC curve for OOD detection.
score_hist(id_scores, ood_scores[, ax, bins])	Plot histogram of OOD scores for ID and OOD samples.

Utilities

compute_threshold_at_tpr(id_scores[, target_tpr])	Compute OOD score threshold that accepts target_tpr fraction of ID samples.
get_ood_predictions(scores, threshold)	Get binary OOD predictions based on threshold.
extract_features(model, data_loader[, ...])	Extract features from a specific layer of the model.