UW-Madison-Lee-Lab/GenLabel_official

ICML 2022 Paper entitled "GenLabel: Mixup Relabeling using Generative Models": https://arxiv.org/pdf/2201.02354.pdf

GenLabel

This repo contains demo PyTorch reimplementations of GenLabel on both synthetic & real datasets.

Dependencies

Tested stable dependencies:

• Ubuntu 18.04.5
• python 3.7.10 (Anaconda)
• PyTorch 1.10.0+cpu
• numpy 1.20.3
• scikit-learn 0.24.2
• scipy 1.6.2
• matplotlib 3.4.2
• openml 0.12.2
• pandas 1.3.3
• jupyter 1.0.0

An alternative way to set up the environment is running the following command with our requirements.txt.

pip install -r requirements.txt

OpenML

Please follow the instruction in OpenML/ directory

2D cube

Clean validation

Vanilla training

python 2dcube_model.py --mixup_setting 0 --lr 0.1 --epoch 40 --validate cln --num_sample 20

Mixup

python 2dcube_model.py --mixup_setting 1 --lr 0.1 --epoch 40 --validate cln --num_sample 20

Mixup + GenLabel

python 2dcube_model.py --mixup_setting 2 --lr 0.1 --epoch 40 --lam 1 --validate cln --num_sample 20

3D cube

Clean validation

Vanilla training

python 3dcube_model.py --mixup_setting 0 --lr 0.1 --epoch 40 --validate cln --num_sample 20

Mixup

python 3dcube_model.py --mixup_setting 1 --lr 0.1 --epoch 40 --validate cln --num_sample 20

Mixup + GenLabel

python 3dcube_model.py --mixup_setting 2 --lr 0.1 --epoch 40 --lam 0.8 --validate cln --num_sample 20

9-class Gaussian

For 9-class Gaussian dataset, please find details in this notebook.

Moon

Clean validation

Vanilla training

python Syn_model.py --mixup_setting 0 --dataset moon --lr 0.1 --epoch 100 --validate cln --num_sample 1000

Mixup

python Syn_model.py --mixup_setting 1 --dataset moon --lr 0.1 --epoch 100 --validate cln --num_sample 1000

Mixup + GenLabel

python Syn_model.py --mixup_setting 2 --dataset moon --lr 0.1 --epoch 100 --lam 1 --bw 0.2 --validate cln --num_sample 1000

Circle

Clean validation

Vanilla training

python Syn_model.py --mixup_setting 0 --dataset circle --lr 0.1 --epoch 100 --validate cln --num_sample 1000

Mixup

python Syn_model.py --mixup_setting 1 --dataset circle --lr 0.1 --epoch 100 --validate cln --num_sample 1000

Mixup + GenLabel

python Syn_model.py --mixup_setting 2 --dataset circle --lr 0.1 --epoch 100 --lam 0.8 --bw 0.2 --validate cln --num_sample 1000

Two-circle

Clean validation

Vanilla training

python Syn_model.py --mixup_setting 0 --dataset twocircle --lr 0.1 --epoch 100 --validate cln --num_sample 1000

Mixup

python Syn_model.py --mixup_setting 1 --dataset twocircle --lr 0.1 --epoch 100 --validate cln --num_sample 1000

Mixup + GenLabel

python Syn_model.py --mixup_setting 2 --dataset twocircle --lr 0.1 --epoch 100 --lam 1 --bw 0.1 --validate cln --num_sample 1000