OOD Detection Metrics

Functions for computing metrics commonly used in the field of out-of-distribution (OOD) detection.

Installation

With PIP

pip install ood-metrics

With Conda

conda install -c conda-forge ood-metrics

Metrics functions

AUROC

Calculate and return the area under the ROC curve using unthresholded predictions on the data and a binary true label.

from ood_metrics import auroc

labels = [0, 0, 0, 1, 0]
scores = [0.1, 0.3, 0.6, 0.9, 1.3]

assert auroc(scores, labels) == 0.75

AUPR

Calculate and return the area under the Precision Recall curve using unthresholded predictions on the data and a binary true label.

from ood_metrics import aupr

labels = [0, 0, 0, 1, 0]
scores = [0.1, 0.3, 0.6, 0.9, 1.3]

assert aupr(scores, labels) == 0.25

FPR @ 95% TPR

Return the FPR when TPR is at least 95%.

from ood_metrics import fpr_at_95_tpr

labels = [0, 0, 0, 1, 0]
scores = [0.1, 0.3, 0.6, 0.9, 1.3]

assert fpr_at_95_tpr(scores, labels) == 0.25

Detection Error

Return the misclassification probability when TPR is 95%.

from ood_metrics import detection_error

labels = [0, 0, 0, 1, 0]
scores = [0.1, 0.3, 0.6, 0.9, 1.3]

assert detection_error(scores, labels) == 0.05

Calculate all stats

Using predictions and labels, return a dictionary containing all novelty detection performance statistics.

from ood_metrics import calc_metrics

labels = [0, 0, 0, 1, 0]
scores = [0.1, 0.3, 0.6, 0.9, 1.3]

assert calc_metrics(scores, labels) == {
    'fpr_at_95_tpr': 0.25,
    'detection_error': 0.05,
    'auroc': 0.75,
    'aupr_in': 0.25,
    'aupr_out': 0.94375
}

Plotting functions

Plot ROC

Plot an ROC curve based on unthresholded predictions and true binary labels.

from ood_metrics import plot_roc

labels = [0, 0, 0, 1, 0]
scores = [0.1, 0.3, 0.6, 0.9, 1.3]

plot_roc(scores, labels)
# Generate Matplotlib AUROC plot

Plot PR

Plot an Precision-Recall curve based on unthresholded predictions and true binary labels.

from ood_metrics import plot_pr

labels = [0, 0, 0, 1, 0]
scores = [0.1, 0.3, 0.6, 0.9, 1.3]

plot_pr(scores, labels)
# Generate Matplotlib Precision-Recall plot

Plot Barcode

Plot a visualization showing inliers and outliers sorted by their prediction of novelty.

from ood_metrics import plot_barcode

labels = [0, 0, 0, 1, 0]
scores = [0.1, 0.3, 0.6, 0.9, 1.3]

plot_barcode(scores, labels)
# Shows visualization of sort order of labels occording to the scores.