imbalanced-semi-self

Rethinking the Value of Labels for Improving Class-Imbalanced Learning

This repository contains the implementation code for paper:
Rethinking the Value of Labels for Improving Class-Imbalanced Learning
Yuzhe Yang, and Zhi Xu
34th Conference on Neural Information Processing Systems (NeurIPS), 2020
[Website] [arXiv] [Paper] [Slides] [Video]

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

@inproceedings{yang2020rethinking,
  title={Rethinking the Value of Labels for Improving Class-Imbalanced Learning},
  author={Yang, Yuzhe and Xu, Zhi},
  booktitle={Conference on Neural Information Processing Systems (NeurIPS)},
  year={2020}
}

Overview

In this work, we show theoretically and empirically that, both semi-supervised learning (using unlabeled data) and self-supervised pre-training (first pre-train the model with self-supervision) can substantially improve the performance on imbalanced (long-tailed) datasets, regardless of the imbalanceness on labeled/unlabeled data and the base training techniques.

Semi-Supervised Imbalanced Learning: Using unlabeled data helps to shape clearer class boundaries and results in better class separation, especially for the tail classes.

Self-Supervised Imbalanced Learning: Self-supervised pre-training (SSP) helps mitigate the tail classes leakage during testing, which results in better learned boundaries and representations.

Installation

Prerequisites

• Download CIFAR & SVHN dataset, and place them in your data_path. Original data will be converted by imbalance_cifar.py and imbalance_svhn.py
• Download ImageNet & iNaturalist 2018 dataset, and place them in your data_path. Long-tailed version will be created using train/val splits (.txt files) in corresponding subfolders under imagenet_inat/data/
• Change the data_root in imagenet_inat/main.py accordingly for ImageNet-LT & iNaturalist 2018

Dependencies

• PyTorch (>= 1.2, tested on 1.4)
• yaml
• scikit-learn
• TensorboardX

Code Overview

Main Files

• train_semi.py: train model with extra unlabeled data, on CIFAR-LT / SVHN-LT
• train.py: train model with (or without) SSP, on CIFAR-LT / SVHN-LT
• imagenet_inat/main.py: train model with (or without) SSP, on ImageNet-LT / iNaturalist 2018
• pretrain_rot.py & pretrain_moco.py: self-supervised pre-training using Rotation prediction or MoCo

Main Arguments

• --dataset: name of chosen long-tailed dataset
• --imb_factor: imbalance factor (inverse value of imbalance ratio \rho in the paper)
• --imb_factor_unlabel: imbalance factor for unlabeled data (inverse value of unlabel imbalance ratio \rho_U)
• --pretrained_model: path to self-supervised pre-trained models
• --resume: path to resume checkpoint (also for evaluation)

Getting Started

Semi-Supervised Imbalanced Learning

Unlabeled data sourcing

CIFAR-10-LT: CIFAR-10 unlabeled data is prepared following this repo using the 80M TinyImages. In short, a data sourcing model is trained to distinguish CIFAR-10 classes and an “non-CIFAR” class. For each class, images are then ranked based on the prediction confidence, and unlabeled (imbalanced) datasets are constructed accordingly. Use the following link to download the prepared unlabeled data, and place in your data_path:

• Unlabeled dataset for CIFAR-10-LT from TinyImages

SVHN-LT: Since its own dataset contains an extra part with 531.1K additional (labeled) samples, they are directly used to simulate the unlabeled dataset.

Note that the class imbalance in unlabeled data is also considered, which is controlled by --imb_factor_unlabel (\rho_U in the paper). See imbalance_cifar.py and imbalance_svhn.py for details.

Semi-supervised learning with pseudo-labeling

To perform pseudo-labeling (self-training), first a base classifier is trained on original imbalanced dataset. With the trained base classifier, pseudo-labels can be generated using

python gen_pseudolabels.py --resume <ckpt-path> --data_dir <data_path> --output_dir <output_path> --output_filename <save_name>

We provide generated pseudo label files for CIFAR-10-LT & SVHN-LT with \rho=50, using base models trained with standard cross-entropy (CE) loss:

• Generated pseudo labels for CIFAR-10-LT with \rho=50
• Generated pseudo labels for SVHN-LT with \rho=50

To train with unlabeled data, for example, on CIFAR-10-LT with \rho=50 and \rho_U=50

python train_semi.py --dataset cifar10 --imb_factor 0.02 --imb_factor_unlabel 0.02

Self-Supervised Imbalanced Learning

Self-supervised pre-training (SSP)

To perform Rotation SSP on CIFAR-10-LT with \rho=100

python pretrain_rot.py --dataset cifar10 --imb_factor 0.01

To perform MoCo SSP on ImageNet-LT

python pretrain_moco.py --dataset imagenet --data <data_path>

Network training with SSP models

Train on CIFAR-10-LT with \rho=100

python train.py --dataset cifar10 --imb_factor 0.01 --pretrained_model <path_to_ssp_model>

Train on ImageNet-LT / iNaturalist 2018

python -m imagenet_inat.main --cfg <path_to_ssp_config> --model_dir <path_to_ssp_model>

Results and Models

All related data and checkpoints can be found via this link. Individual results and checkpoints are detailed as follows.

Semi-Supervised Imbalanced Learning

CIFAR-10-LT

Model	Top-1 Error	Download
CE + D_U@5x (\rho=50 and \rho_U=1)	16.79	ResNet-32
CE + D_U@5x (\rho=50 and \rho_U=25)	16.88	ResNet-32
CE + D_U@5x (\rho=50 and \rho_U=50)	18.36	ResNet-32
CE + D_U@5x (\rho=50 and \rho_U=100)	19.94	ResNet-32

SVHN-LT

Model	Top-1 Error	Download
CE + D_U@5x (\rho=50 and \rho_U=1)	13.07	ResNet-32
CE + D_U@5x (\rho=50 and \rho_U=25)	13.36	ResNet-32
CE + D_U@5x (\rho=50 and \rho_U=50)	13.16	ResNet-32
CE + D_U@5x (\rho=50 and \rho_U=100)	14.54	ResNet-32

Test a pretrained checkpoint

python train_semi.py --dataset cifar10 --resume <ckpt-path> -e

Self-Supervised Imbalanced Learning

CIFAR-10-LT

• Self-supervised pre-trained models (Rotation)

  Dataset Setting	\rho=100	\rho=50	\rho=10
  Download	ResNet-32	ResNet-32	ResNet-32

• Final models (200 epochs)

  Model	\rho	Top-1 Error	Download
  CE(Uniform) + SSP	10	12.28	ResNet-32
  CE(Uniform) + SSP	50	21.80	ResNet-32
  CE(Uniform) + SSP	100	26.50	ResNet-32
  CE(Balanced) + SSP	10	11.57	ResNet-32
  CE(Balanced) + SSP	50	19.60	ResNet-32
  CE(Balanced) + SSP	100	23.47	ResNet-32

CIFAR-100-LT

• Self-supervised pre-trained models (Rotation)

  Dataset Setting	\rho=100	\rho=50	\rho=10
  Download	ResNet-32	ResNet-32	ResNet-32

• Final models (200 epochs)

  Model	\rho	Top-1 Error	Download
  CE(Uniform) + SSP	10	42.93	ResNet-32
  CE(Uniform) + SSP	50	54.96	ResNet-32
  CE(Uniform) + SSP	100	59.60	ResNet-32
  CE(Balanced) + SSP	10	41.94	ResNet-32
  CE(Balanced) + SSP	50	52.91	ResNet-32
  CE(Balanced) + SSP	100	56.94	ResNet-32

ImageNet-LT

• Self-supervised pre-trained models (MoCo)
  [ResNet-50]

• Final models (90 epochs)

  Model	Top-1 Error	Download
  CE(Uniform) + SSP	54.4	ResNet-50
  CE(Balanced) + SSP	52.4	ResNet-50
  cRT + SSP	48.7	ResNet-50

iNaturalist 2018

• Self-supervised pre-trained models (MoCo)
  [ResNet-50]

• Final models (90 epochs)

  Model	Top-1 Error	Download
  CE(Uniform) + SSP	35.6	ResNet-50
  CE(Balanced) + SSP	34.1	ResNet-50
  cRT + SSP	31.9	ResNet-50

Test a pretrained checkpoint

# test on CIFAR-10 / CIFAR-100
python train.py --dataset cifar10 --resume <ckpt-path> -e

# test on ImageNet-LT / iNaturalist 2018
python -m imagenet_inat.main --cfg <path_to_ssp_config> --model_dir <path_to_model> --test

Acknowledgements

This code is partly based on the open-source implementations from the following sources: OpenLongTailRecognition, classifier-balancing, LDAM-DRW, MoCo, and semisup-adv.

Contact

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