jiequancui / reslt

ResLT: Residual Learning for Long-tailed Recognition (TPAMI 2022)

ResLT: Residual Learning for Long-tailed Recognition (TPAMI 2022)

This repository contains the implementation code for paper:
Residual Learning for Long-tailed Recognition https://arxiv.org/abs/2101.10633

Updates

We further verifty the proposed ResLT is complementary to ensemble-based methods. Equipped with RIDEResNeXt, our model achieves better results. All experiments are conducted without knowledge distillation for fair comparison. For RIDE, we use their public code and train 180 epochs.

ImageNet-LT

Model	Top-1 Acc	Download	log
RIDEResNeXt(3 experts)	55.1	-	log
RIDEResNeXt-ResLT(3 experts)	57.6	model	log

Inaturalist 2018

Model	Top-1 Acc	Download	log
RIDEResNeXt(3 experts)	70.8	-	log
RIDEResNeXt-ResLT(3 experts)	72.9	model	log

Overview

In this paper, we proposed a residual learning method to address long-tailed recognition, which contains a Residual Fusion Module and a Parameter Specialization Mechanism. With extensive ablation studies, we demonstrate the effectiveness of our method.

Get Started

ResLT Training

For CIFAR, due to the small data size, different experimental environment can have a big difference. To achieve the reported results, you may need to slightly tune the $\alpha$.

bash sh/CIFAR100/CIFAR100LT_imf0.01_resnet32sx1_beta0.9950.sh

For ImageNet-LT,

bash sh/X50.sh

For iNaturalist 2018,

bash sh/R50.sh

Results and Models

ImageNet-LT

Model	Download	log
ResNet-10	model	log
ResNeXt-50	model	log
ResNeXt-101	model	log

iNatualist 2018

Model	Download	log
ResNet-50	model	log

Places-LT

Model	Download	log
ResNet-152	-	-

Acknowledgements

This code is partly based on the open-source implementations from offical PyTorch examples and LDAM-DRW.

Contact

If you have any questions, feel free to contact us through email ([email protected]) or Github issues. Enjoy!

BibTex

If you find this code or idea useful, please consider citing our works:

@ARTICLE{9774921,
  author={Cui, Jiequan and Liu, Shu and Tian, Zhuotao and Zhong, Zhisheng and Jia, Jiaya},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={ResLT: Residual Learning for Long-Tailed Recognition}, 
  year={2023},
  volume={45},
  number={3},
  pages={3695-3706},
  doi={10.1109/TPAMI.2022.3174892}}

@ARTICLE{10130611,
  author={Cui, Jiequan and Zhong, Zhisheng and Tian, Zhuotao and Liu, Shu and Yu, Bei and Jia, Jiaya},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence}, 
  title={Generalized Parametric Contrastive Learning}, 
  year={2023},
  volume={},
  number={},
  pages={1-12},
  doi={10.1109/TPAMI.2023.3278694}}


@inproceedings{cui2021parametric,
  title={Parametric contrastive learning},
  author={Cui, Jiequan and Zhong, Zhisheng and Liu, Shu and Yu, Bei and Jia, Jiaya},
  booktitle={Proceedings of the IEEE/CVF international conference on computer vision},
  pages={715--724},
  year={2021}
}
  
@article{cui2022region,
  title={Region Rebalance for Long-Tailed Semantic Segmentation},
  author={Cui, Jiequan and Yuan, Yuhui and Zhong, Zhisheng and Tian, Zhuotao and Hu, Han and Lin, Stephen and Jia, Jiaya},
  journal={arXiv preprint arXiv:2204.01969},
  year={2022}
  }
  
@article{zhong2023understanding,
  title={Understanding Imbalanced Semantic Segmentation Through Neural Collapse},
  author={Zhong, Zhisheng and Cui, Jiequan and Yang, Yibo and Wu, Xiaoyang and Qi, Xiaojuan and Zhang, Xiangyu and Jia, Jiaya},
  journal={arXiv preprint arXiv:2301.01100},
  year={2023}
}

reslt's Issues

NO UTILS

the company is Smartmore?

Can you share the code about how to use mixup in reslt?

Thanks for the open source, but how to use mixup in reslt confuses me.

Directly randomly sample a batch of data in, and we do not manually do anything specific to the long tail data ?

Am I correct ?

When can the relevant code be uploaded?

Hello, I want to study your related papers. When can the relevant code be uploaded?

请问源代码什么时候可以展示出来？

When can the source code be displayed?

Why 1x1 Conv can be replaced with BN layer?

Greetings
Really nice work, but I have a small question.
You mentioned in the code comments that # 1x1 conv can be replaced with more light-weight bn layer, I wonder why they are equivalent? Actually I think conv and bn are totally different.

Gamma

Thanks for your great work.

I'm confused about the gamma hyperparameter. Could you explain the function of gamma for me? Why do you just set it on the ImageNet-lt and iNaturalist datasets, not Cifar?

Variables are not defined in the code

Sorry to bother you that

Code is incomplete, train_sampler is not defined in ImageNetTrain， And Can you explain root_path，dataset，data_path to me

Looking forward to your reply

About the classifiers

Thanks for proposing such interesting method. I am wondering three branches have diffierent classifiers or same?
Furthermore, do you compare with mulit-experts methods as [1]?
[1] LONG-TAILED RECOGNITION BY ROUTING DIVERSE DISTRIBUTION-AWARE EXPERTS, 2021 ICLR.

1x1conv Block

Thanks for your great work.

I'm confused about the site of the 1x1conv Block. Could you explain why the 1x1conv block before the avgpool during training and behind the avgpool during testing?