prompt-wise

Worst-case forget set on prompt-wise unlearning

This is the official repository for Worst-case forget set on prompt-wise unlearning.

Requirements

This repo is based on the environment ip2p.

Install the environment using the following command:

conda env create -f environment.yaml

Download model and UnlearnCanvas dataset

1. Model
2. UnlearnCanvas dataset

Scripts

1. Utilize ckpt_path and data_path to specify the model and dataset path. Employ ESD with SignSGD to identify the worst-case forget set on prompt-wise unlearning.

python stable_diffusion/train-scripts/select-pair.py --train_method=xattn --SignESD --lr 1e-5 --w_lr 100.0 --devices 0,0 --output_dir results/select-pair --ckpt_path {the path of origin model} --data_path {the path of dataset}

2. Employ ESD with AdamW to unlearn the worst-case forget set. Subsequently, utilize the unlearned model to generate associated images.

python stable_diffusion/train-scripts/evaluate_selection.py --train_method=xattn --lr 3e-7 --devices 0,0 --output_dir results/evaluation --ckpt_path {the path of origin model} --data_path {the path of dataset} --w_path {path of the selection weight}

If random_choice, the original model will unlearn on a random forget set.