Enabling Selective Forgetting in Generative Models
Arxiv | TMLR |

• Introduction
• Key Features
• Installation
• Prerequisites
• Quick Start and Usage
• Unlearned Models Checkpoint
• Citation

This repository implements "Adapt then Unlearn" – a framework for selective forgetting in Generative Adversarial Networks (GANs). By exploiting parameter space semantics, our method enables:

• Class-level unlearning (e.g., remove digit "7" from MNIST generators)
• Feature-level unlearning (e.g., suppress "Smiling" attribute in CelebA-HQ)

Recommended for researchers exploring machine unlearning, GAN safety, and model editing.

• ✅ Multi-Dataset Support: MNIST | CelebA-HQ | AFHQ
• ✅ Three Repulsion Loss Variants: EL2 | NL2 | IL2
• ✅ Pretrained Model Integration
• ✅ Extensible Codebase

git clone https://github.com/yourusername/adapt-then-unlearn.git
cd adapt-then-unlearn
conda create -n unlearn python=3.8
conda activate unlearn
pip install -r requirements.txt

Download required checkpoints:

# MNIST StyleGAN2
wget https://example.com/pretrained/mnist_generator.pth

# CelebA-HQ StyleGAN2
wget https://example.com/pretrained/celeba_generator.pth

# Classifiers
./scripts/download_classifiers.sh

We utilize pre-trained classifiers for the feedback-based framework.

MNIST

The checkpoint and model for the MNIST classifier are available in this repository. We used the checkpoints provided in csinva/gan-vae-pretrained-pytorch.

CelebA-HQ

For training and implementation of a classifier for classifying facial attributes in the CelebA dataset, we referred to rgkannan676/Recognition-and-Classification-of-Facial-Attributes.

MNIST (Digit Class):

python stylegan2_ewc.py --exp "7" --iter 2000 \
  --g_ckpt ./pretrained/mnist_generator.pth --size 32

CelebA-HQ (Facial Feature):

python stylegan2_ewc_train.py --exp "Smiling" --iter 5000 \
  --gan_ckpt ./pretrained/celeba_generator.pth

MNIST (Class Removal):

python train_different_losses.py --expt "7" \
  --list_of_models adapted_model1.pth adapted_model2.pth \
  --size 32 --ckpt ./pretrained/mnist_generator.pth \
  --iter 3000 --repel --gamma 1.2 --loss_type "l2"

CelebA-HQ (Feature Removal):

python unlearn_main.py --expt "Smiling" \
  --list_of_models classifier1.pth classifier2.pth \
  --ckpt ./pretrained/celeba_generator.pth \
  --iter 5000 --repel_loss --gamma 0.8 --loss_type "exp"

Parameters:

• expt: Experiment name.
• list_of_models: Paths to the negative adapted models.
• ckpt: Path to the pre-trained GAN.
• iter: Number of iterations.
• repel_loss: Flag to use repulsion loss.
• gamma: Value of the scaling factor for the loss function.
• loss_type: Type of repulsion loss function ("exp", "l2", or "reci").

Feature List of CelebA-HQ

Note Regarding the Type of Repulsion Loss Function:

• For $\mathcal{L}_{repulsion}^{{EL2}}$, use args.loss_type="exp"
• For $\mathcal{L}_{repulsion}^{{NL2}}$, use args.loss_type="l2"
• For $\mathcal{L}_{repulsion}^{{IL2}}$, use args.loss_type="reci"

Theory details in our paper.

Access pre-trained unlearned models via the following link:

Unlearned Models Folder

The folder contains subdirectories for CelebA-HQ, MNIST, and AFHQ, each containing (unlearnt) models that can be used for evaluation and comparison.

If you use this code in your research, please cite our paper:

@article{
tiwary2025adapt,
title={Adapt then Unlearn: Exploring Parameter Space Semantics for Unlearning in Generative Adversarial Networks},
author={Piyush Tiwary and Atri Guha and Subhodip Panda and Prathosh AP},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2025},
url={https://openreview.net/forum?id=jAHEBivObO},
note={}
}