Vision Unlearning

A standard interface for unlearning algorithms, datasets, metrics, and evaluation in vision-related machine learning tasks.

pip install vision-unlearning

Compatible with Python 3.10 to 3.12.

---

What is Vision Unlearning?

Machine unlearning is the problem of removing the influence of specific training data from a trained model — without retraining from scratch. Vision Unlearning provides a standard, architecture-agnostic interface for the core building blocks:

• Unlearning algorithms — FADE, UCE, Munba, and others

• Datasets — standardised forget/retain splits for CIFAR-10/100, CelebA, and more

• Metrics — quantitative measures of forgetting quality, retain performance, and side effects

• Evaluator — a unified pipeline for running an algorithm and measuring all relevant metrics

• I-CARE benchmark — a comprehensive evaluation testbed for comparing methods across tasks and entities

Getting Started

• Tutorials

API Reference

• vision_unlearning.unlearner
• vision_unlearning.datasets
• vision_unlearning.metrics
• vision_unlearning.evaluator
• vision_unlearning.integrations
• vision_unlearning.utils
• vision_unlearning.benchmarks