Influence-balanced Loss for Imbalanced Visual Classification (ICCV, 2021)

This is the official implementation of Influence-balanced Loss for Imbalanced Visual Classification in PyTorch. The code heavily relies on LDAM-DRW.

Requirements

All codes are written by Python 3.7, and ‘requirements.txt’ contains required Python packages. To install requirements:

pip install -r requirements.txt

Dataset

Create ‘data/’ directory and download original data in the directory to make imbalanced versions.

  • Imbalanced CIFAR. The original data will be downloaded and converted by imbalancec_cifar.py.
  • Imbalanced Tiny ImageNet. Download the data first, and convert them by imbalance_tinyimagenet.py.
  • The paper also reports results on iNaturalist 2018. We will update the code for iNaturalist 2018 later.

Training

We provide several training examples:

CIFAR

  • CE baseline (CIFAR-100, long-tailed imabalance ratio of 100)

python cifar_train.py --dataset cifar100 --loss_type CE --train_rule None --imb_type exp --imb_factor 0.01 --epochs 200 --num_classes 100 --gpu 0
  • IB (CIFAR-100, long-tailed imabalance ratio of 100)

python cifar_train.py --dataset cifar100 --loss_type IB --train_rule IBReweight --imb_type exp --imb_factor 0.01 --epochs 200 --num_classes 100 --start_ib_epoch 100 --gpu 0
  • IB + CB (CIFAR-100, long-tailed imabalance ratio of 100)

python cifar_train.py --dataset cifar100 --loss_type IB --train_rule CBReweight --imb_type exp --imb_factor 0.01 --epochs 200 --num_classes 100 --start_ib_epoch 100 --gpu 0
  • IB + Focal (CIFAR-100, long-tailed imabalance ratio of 100)

python cifar_train.py --dataset cifar100 --loss_type IBFocal --train_rule IBReweight --imb_type exp --imb_factor 0.01 --epochs 200 --num_classes 100 --start_ib_epoch 100 --gpu 0

Tiny ImageNet

  • CE baseline (long-tailed imabalance ratio of 100)

python tinyimage_train.py --dataset tinyimagenet -a resnet18 --loss_type CE --train_rule None --imb_type exp --imb_factor 0.01 --epochs 100 --lr 0.1  --num_classes 200
  • IB (long-tailed imabalance ratio of 100)

python tinyimage_train.py --dataset tinyimagenet -a resnet18 --loss_type IB --train_rule IBReweight --imb_type exp --imb_factor 0.01 --epochs 100 --lr 0.1  --num_classes 200 --start_ib_epoch 50

Citation

If you find our paper and repo useful, please cite our paper

GitHub

https://github.com/pseulki/IB-Loss