M2m: Imbalanced Classification via Major-to-minor Translation

This repository contains code for the paper "M2m: Imbalanced Classification via Major-to-minor Translation" by Jaehyung Kim*, Jongheon Jeong* and Jinwoo Shin.

Dependencies

• python3
• pytorch >= 1.1.0
• torchvision
• tqdm

Scripts

Please check out run.sh for all the scripts to reproduce the CIFAR-10-LT results reported.

Training procedure of M2m

1. Train a baseline network g for generating minority samples

python train.py --no_over --ratio 100 --decay 2e-4 --model resnet32 --dataset cifar10 \
--lr 0.1 --batch-size 128 --name 'ERM' --warm 200 --epoch 200   

2. Train another network f using M2m with the pre-trained g

python train.py -gen -r --ratio 100 --decay 2e-4 --model resnet32 --dataset cifar10 \
--lr 0.1 --batch-size 128 --name 'M2m' --beta 0.999 --lam 0.5 --gamma 0.9 \
--step_size 0.1 --attack_iter 10 --warm 160 --epoch 200 --net_g ./checkpoint/pre_trained_g.t7 

We also provide a pre-trained ResNet-32 model of g at checkpoint/erm_r100_c10_trial1.t7, so one can directly use M2m without pre-training as follows:

python train.py -gen -r --ratio 100 --decay 2e-4 --model resnet32 --dataset cifar10 \
--lr 0.1 --batch-size 128 --name 'M2m' --beta 0.999 --lam 0.5 --gamma 0.9 \
--step_size 0.1 --attack_iter 10 --warm 160 --epoch 200 --net_g ./checkpoint/erm_r100_c10_trial1.t7