EC-523 Project: Enhancing the SCRUB Model

Introduction

This repository contains the re-implementation and enhancement of the SCalable Remembering and Unlearning unBound (SCRUB) model. The SCRUB model represents a novel approach in [briefly describe the main function or advantage of the model]. This project aims to [state the objectives of your enhancements or what aspects of the SCRUB model you focused on].

Prerequisites

To install the prerequsites librarys run

pip install -r requirements.txt

Usage

Obtaining the Models

To download and set up the pre-trained and re-trained SCRUB models in your local environment, run:

python3 train_checkpoints.py

Performing Unlearning

After obtaining the pre-trained model, use the following commands to perform unlearning:

SCRUB Method:

python3 main.py --checkpoints --sgda_learning_rate 0.006 --sub_sample 0.0 --unlearning_method "SCRUB" 

Finetuning Method:

python3 main.py --checkpoints --sgda_learning_rate 0.1 --unlearning_method "finetuning" 

Negative Gradient Method:

python3 main.py --checkpoints --sgda_learning_rate 0.006 --unlearning_method "negative" --sgda_epochs 5

Hyperparameter Tuning

Use the following scripts to tune hyperparameters:

run_lr.sh
run_subsampling.sh
run_temp.sh

Reviewing Results

All results can be reviewed in the checkpoints.ipynb notebook.