README.md

Experiments on Gaussian Mixture Models

This folder contains the code and scripts for the Gaussian mixture model (GMM) experiments.

Quick Start

We give an example of MCMC unlearning with SGLD. Other experiment scripts can be found in ./scripts.

Step 1: Train the GMM on the full dataset

python train.py \
    --cpu \
    --gmm-kk 4 \
    --gmm-std 1 \
    --optim sgld \
    --batch-size 64 \
    --burn-in-steps 4000 \
    --eval-freq 100 \
    --lr 4 \
    --lr-decay-exp -0.5005 \
    --ifs-scaling 1 \
    --ifs-iter-T 32 \
    --ifs-samp-T 5 \
    --ifs-iter-bs 64 \
    --ifs-rm-bs 4 \
    --ifs-kill-num 0 \
    --samp-num 4000 \
    --save-dir ./exp_data/gmm/sgld \
    --save-name full

Step 2: Remove datums from the trained GMM

python forget.py \
    --cpu \
    --gmm-kk 4 \
    --gmm-std 1 \
    --optim sgld \
    --batch-size 64 \
    --burn-in-steps 2000 \
    --eval-freq 100 \
    --lr 4 \
    --lr-decay-exp -0.5005 \
    --ifs-scaling 1 \
    --ifs-iter-T 32 \
    --ifs-samp-T 5 \
    --ifs-iter-bs 64 \
    --ifs-rm-bs 4 \
    --ifs-kill-num 800 \
    --resume-path ./exp_data/gmm/sgld/full-model.pkl \
    --save-dir ./exp_data/gmm/sgld \
    --save-name forget