Home

Awesome

Classifier-Balancing

This repository contains code for the paper:

Decoupling Representation and Classifier for Long-Tailed Recognition
Bingyi Kang, Saining Xie,Marcus Rohrbach, Zhicheng Yan, Albert Gordo, Jiashi Feng, Yannis Kalantidis
[OpenReview] [Arxiv] [PDF] [Slides] [@ICLR]
Facebook AI Research, National University of Singapore
International Conference on Learning Representations (ICLR), 2020

Abstract

The long-tail distribution of the visual world poses great challenges for deep learning based classification models on how to handle the class imbalance problem. Existing solutions usually involve class-balancing strategies, e.g., by loss re-weighting, data re-sampling, or transfer learning from head- to tail-classes, but all of them adhere to the scheme of jointly learning representations and classifiers. In this work, we decouple the learning procedure into representation learning and classification, and systematically explore how different balancing strategies affect them for long-tailed recognition. The findings are surprising: (1) data imbalance might not be an issue in learning high-quality representations; (2) with representations learned with the simplest instance-balanced (natural) sampling, it is also possible to achieve strong long-tailed recognition ability with relative ease by adjusting only the classifier. We conduct extensive experiments and set new state-of-the-art performance on common long-tailed benchmarks like ImageNet-LT, Places-LT and iNaturalist, showing that it is possible to outperform carefully designed losses, sampling strategies, even complex modules with memory, by using a straightforward approach that decouples representation and classification.

 

<p align="center"> <img src='./assets/tau_norm.png' width=800> </p> &nbsp;

If you find this code useful, consider citing our work:

@inproceedings{kang2019decoupling,
  title={Decoupling representation and classifier for long-tailed recognition},
  author={Kang, Bingyi and Xie, Saining and Rohrbach, Marcus and Yan, Zhicheng
          and Gordo, Albert and Feng, Jiashi and Kalantidis, Yannis},
  booktitle={Eighth International Conference on Learning Representations (ICLR)},
  year={2020}
}

Requirements

The code is based on https://github.com/zhmiao/OpenLongTailRecognition-OLTR.

Dataset

Change the data_root in main.py accordingly.

Representation Learning

  1. Instance-balanced Sampling
python main.py --cfg ./config/ImageNet_LT/feat_uniform.yaml
  1. Class-balanced Sampling
python main.py --cfg ./config/ImageNet_LT/feat_balance.yaml
  1. Square-root Sampling
python main.py --cfg ./config/ImageNet_LT/feat_squareroot.yaml
  1. Progressively-balancing Sampling
python main.py --cfg ./config/ImageNet_LT/feat_shift.yaml

Test the joint learned classifier with representation learning

python main.py --cfg ./config/ImageNet_LT/feat_uniform.yaml --test 

Classifier Learning

  1. Nearest Class Mean classifier (NCM).
python main.py --cfg ./config/ImageNet_LT/feat_uniform.yaml --test --knn
  1. Classifier Re-training (cRT)
python main.py --cfg ./config/ImageNet_LT/cls_crt.yaml --model_dir ./logs/ImageNet_LT/models/resnext50_uniform_e90
python main.py --cfg ./config/ImageNet_LT/cls_crt.yaml --test
  1. Tau-normalization

Extract fatures

for split in train_split val test
do
  python main.py --cfg ./config/ImageNet_LT/feat_uniform.yaml --test --save_feat $split
done

Evaluation

for split in train val test
do
  python tau_norm.py --root ./logs/ImageNet_LT/models/resnext50_uniform_e90/ --type $split
done
  1. Learnable weight scaling (LWS)
python main.py --cfg ./config/ImageNet_LT/cls_lws.yaml --model_dir ./logs/ImageNet_LT/models/resnext50_uniform_e90
python main.py --cfg ./config/ImageNet_LT/cls_lws.yaml --test

Results and Models

ImageNet_LT

iNaturalist 2018

Places_LT

To test a pretrained model:
python main.py --cfg /path/to/config/file --model_dir /path/to/model/file --test

License

This project is licensed under the license found in the LICENSE file in the root directory of this source tree (here). Portions of the source code are from the OLTR project.