Class-Wise Difficulty-Balanced Loss for Solving Class-Imbalance

---

by Saptarshi Sinha, Hiroki Ohashi and Katsuyuki Nakamura

This repository contains the official implementation of the paper 'Class-Wise Difficulty-Balanced Loss for Solving Class-Imbalance', which was accepted for Oral presentation at ACCV, 2020. (paper) (arXiv)

The code has been organized under 3 folders namely 'CIFAR-LT', 'EGTEA' and 'ImageNet-LT' that represents the 3 datasets that we have used in our paper.

Requirements

---

The environment required to successfully implement our paper mainly needs

- Python >= 3.6
- PyTorch == 1.5.0
- Opencv-python == 4.1.2
- Pillow
- PyYaml

Experiments on CIFAR-LT

---

cd CDB-loss/CIFAR-LT

Please download the CIFAR-100 data and extract it in ./data/.

To start training on CIFAR-LT using our CDB-CE loss,

python cifar_train.py --class_num 100 --imbalance 200 --loss_type CDB-CE --tau 1.5 --n_gpus 1

Use -- class_num 100 for CIFAR100-LT. Select the amount of imbalance you want to inject in the dataset by using -- imbalance 200. Note -- imbalance 1 means no imbalance will be injected.

To evaluate the best model on the balanced test set,

python cifar_test.py --saved_model_path saved_model/best_cifar100_imbalance200.pth --class_num 100 --n_gpus 1

Select the appropriate saved model and evaluate. The output reads like

Test Accuracy is 0.3740

Experiments on ImageNet-LT

---

For ImageNet-LT, we build our codes on the code from classifier-balancing. To reproduce results of classifier-balancing, please follow this.

Download the dataset ImageNet2014. Accordingly change the data_root in CDB-loss/ImageNet-LT/main.py.

cd CDB-loss/ImageNet-LT

To train a ResNet10 on ImageNet-LT using our CDB-CE loss,

python main.py --cfg ./config/ImageNet_LT/feat_uniform_with_CDBloss.yaml

In the config file, you can change the value of tau for our loss function. To evaluate the final model on the test set,

python main.py --test --model_dir ./logs/ImageNet_LT/models/resnet10_uniform_cdbce

Experiments on EGTEA

---

Download the trimmed action clips and annotations for EGTEA dataset. Extract the frames using CDB-loss/EGTEA/data/extract_frames.py and save the frames under extracted_frames. The folder structure should look like this.

datasets
|
|
|__EGTEA
    |
    |__  extracted_frames
          |
          |___  OP01-R01-PastaSalad
          |      |
          |      |___  OP01-R01-PastaSalad-1002316-1004005-F024051-F024101     
          |      |              |__ 000000.jpg       
          |      |              |__ 000001.jpg
          |      |              |__  ...
          |      |___  OP01-R01-PastaSalad-1004110-1021110-F024057-F024548
          |      |                   ...
          |                         
          |____  OP01-R02-TurkeySandwich
          |      |
          |      |___   OP01R02-TurkeySandwich-102320-105110-F002449-F002529
          |                                   ...
          |   ...

The train/val/test splits used for our experiments are provided under CDB-loss/EGTEA/data.

cd CDB-loss/EGTEA

Download resnext-101-kinetics.pth from here and save it under ./pretrained_weights/ .

To train a 3D-ResNeXt101 on EGTEA dataset using CDB-CE loss,

python EGTEA_train.py --data_root ~/datasets/EGTEA/extracted_frames --loss_type CDB-CE --tau 1.5 --n_gpus 2

Provide absolute path to extracted_frames as data_root.

To evaluate the final model,

python EGTEA_test.py --data_root ~/datasets/EGTEA/extracted_frames --trained_model ./models/best_model.pth --n_gpus 2

Citations

---

If you use our code, please consider citing our paper as

@InProceedings{Sinha_2020_ACCV,
author={Sinha, Saptarshi and Ohashi, Hiroki and Nakamura, Katsuyuki},
title={Class-Wise Difficulty-Balanced Loss for Solving Class-Imbalance},
booktitle={Proceedings of the Asian Conference on Computer Vision (ACCV)},
month={November},
year={2020}
}

For queries, contact at saptarshi.sinha.hx@hitachi.com