oodkit 0.1.0

Simple out-of-distribution detection tools for ML and computer vision.

OODKit

A small Python library for out-of-distribution (OOD) detection in computer vision. Fit a detector on in-distribution features, score new samples (higher score = more OOD), and evaluate or compare methods with a consistent sklearn-style API.

OODKit works on either image-classification or object-detection data, and can take you from raw images to scored samples in a few steps.

Install

pip install oodkit

OODKit includes the runtime dependencies needed for detectors, embeddings, plots, and the example workflows.

What's in the box

• Detectors (oodkit.detectors) - a family of OOD scorers (MSP, Energy, Mahalanobis, KNN, PCA / CoP / CoRP, PCAFusion, ViM, WDiscOOD) with a shared fit / score / predict contract.
• Features (oodkit.data) - a single Features(logits=..., embeddings=...) container that every detector accepts.
• Embedder (oodkit.embeddings) - helper tools that turn a dataset of images into logits/embeddings via a pretrained backbone, with disk-backed extraction so analysis can run on machines without a GPU.
• Evaluation (oodkit.evaluation) - ScoreBank aligns scores from multiple detectors with labels and metadata; metrics (evaluate, evaluate_by_class) and plots (score_distributions, rank_grid, ROC / PR / correlation) all consume a bank.
• Object-detection support - chip utilities (oodkit.data.chips, ChipDataset), image-level pooling (pool_image_scores), and oodkit.contrib.coco for COCO / COCO-O ingestion without pycocotools.

Minimal example

from oodkit.data import Features
from oodkit.detectors import Energy
from oodkit.evaluation import ScoreBank, evaluate

# You supply logits / embeddings however you like (numpy, torch, Embedder, ...).
id_feat = Features(logits=id_logits, embeddings=id_embeddings)
ood_feat = Features(logits=ood_logits, embeddings=ood_embeddings)

det = Energy()
det.fit(id_feat)

scores_id = det.score(id_feat)
scores_ood = det.score(ood_feat)

bank = ScoreBank(ood_labels=[0] * len(scores_id) + [1] * len(scores_ood))
bank.add("Energy", [*scores_id, *scores_ood])
print(evaluate(bank))  # AUROC, FPR@95, AUPR, ...

Examples

For full end-to-end runs (images to extraction to multiple detectors to comparison plots), see the example notebooks on GitHub:

• ImageNet OOD showcase - classification OOD on ImageNet vs ImageNet-O.
• COCO OOD showcase - object-detection OOD on COCO vs COCO-O (chips, image-level pooling, per-domain breakdowns, ranked chip galleries).
• Notebook guide - dataset layouts, configurable paths, and notebook-specific notes.

Development

Contributing or running from a cloned checkout? See README_DEV.md.