mybirth0407/learning-without-prejudices

Implementation of Learning without Prejudices: Continual Unbiased Learning via Benign and Malignant Forgetting (ICLR 2023)

Learning without Prejudices: Continual Unbiased Learning via Benign and Malignant Forgetting (ICLR 2023)

Official Implementation of Learning without Prejudices (LwP), accepted at ICLR 2023.

@inproceedings{
jeon2023learning,
title={Learning without Prejudices: Continual Unbiased Learning via Benign and Malignant Forgetting},
author={Myeongho Jeon and Hyoje Lee and Yedarm Seong and Myungjoo Kang},
booktitle={International Conference on Learning Representations},
year={2023},
url={https://openreview.net/forum?id=gfPUokHsW-}
}

Prerequisites

Our implementations are conducted on following settings

python==3.9
torch==1.7.0
torchvision==0.8.2

Make Dataset

• Colored Biased MNIST

# generate train and test set with beta 0.85
python make_dataset.py --data mnist-biased --split both --biased_ratio 0.85

Train

# Baseline
python train.py --exp_name Exp --epoch 20 --biased_r 0.85 --method base 

# Learning without Prejudices
python train.py --exp_name Exp --epoch 20 --biased_r 0.85 --method lwp

# Learning without Prejudices (Replay version)
python train.py --exp_name Exp --epoch 20 --biased_r 0.85 --method lwp --replay

Evaluation

python eval.py --exp_path <path to experiment>

Project Structure

This project relies on the following folder organisation:

├───dataset                             # folder of dataset
|   ├───MNIST-BIASED                    
|       ├───MNIST                       # original MNIST dataset
|       └───MNIST-BIASED-BG-Biased*     # Biased MNIST dataset generated by make_dataset.py
|   ├───MNIST-CELEBA                    
|       ├───CelebA-HQ.py                # data loader 
|       ├───celeba_train_[TASK].txt     # train dataset for [TASK]
|       └───celeba_test.txt             # test dataset
│   ├───loader.py                       # data loader
│   └───mnist.py                        # Biased MNIST dataset
├───model
|   ├───lwp                             # LwP
|   |   ├───gen                         # feature generator
|   |   |   ├───__init__.py
|   |   |   ├───generator.py
|   |   |   ├───wgan.py
|   |   ├───ssl                         # self-supervised learninge
|   |   |   ├───__init__.py
|   |   |   ├───base.py
|   |   |   ├───byol.py
|   |   ├───__init__.py
|   |   ├───lwp_model.py                # LwP
|   ├───buffer.py                          
|   ├───module.py
|   ├───simple_convnet.py               # backbone network
├───exps                                # saved weights and logs in training
├───train.py                            # training code
├───eval.py                             # evalulation code
├───make_dataset.py                     # make Colored Biased MNIST
├───option.py                           
├───traininer.py                        
└───utils.py