examples_unlearning

Sparse Unlearn CIFAR10 with PyTorch

This is modified and sourced from NeurIPS 2023 spotlight paper repo Model Sparsity Can Simplify Machine Unlearning from MSU OPTML Group.

Environment

# create and prepare a virtual environment
conda create -n modelsmith python=3.9
conda activate modelsmith
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

Training to get a pretrained model

• Predefined models to train: ResNet18, ResNet34, ResNet50, ResNet101, ResNet152, VGG11, VGG13, VGG16, VGG19

# Start training with:
python train.py --arch=ResNet18 --epochs=100

Sparse Unlearning

# Sparse Unlearning the pretrained model with:
python main_forget.py

# You will see retain accuracy, forget accuracy, validation acc, and test acc in the final output