Contrastive Unlearning for Few-shot classifiers

This is a code repository for ICLR2025 rebuttal. This repository contains code for unlearning specific classes from CLIP models while retaining performance on other classes. The key components are:

Main Files

• unlearn.py: Contains the core unlearning implementation including:

  • Contrastive unlearning loss function
  • Fine-tuning and retention mechanisms
  • Zero-shot evaluation
  • Training loops for unlearning and retention

• finetune.py: Contains the fine-tuning implementation including:

  • Dataset wrapper for image-text pairs
  • Fine-tuning training loop with contrastive loss
  • Zero-shot evaluation on test set
  • Model checkpointing to save best performing model
  • Wandb integration for experiment tracking

Requirements

• PyTorch
• Weights & Biases
• Torchvision
• TQDM
• PIL