Learning to Bootstrap for Combating Label Noise

This repo is the official implementation of our paper "Learning to Bootstrap for Combating Label Noise".

Citation

If you use this code for your research, please cite our paper "Learning to Bootstrap for Combating Label Noise".

@misc{zhou2022learning,
      title={Learning to Bootstrap for Combating Label Noise}, 
      author={Yuyin Zhou and Xianhang Li and Fengze Liu and Xuxi Chen and Lequan Yu and Cihang Xie and Matthew P. Lungren and Lei Xing},
      year={2022},
      eprint={2202.04291},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Requirements

Python >= 3.6.4
Pytorch >= 1.6.0
Higher = 0.2.1
Tensorboardx = 2.4.1

Training

First, please create a folder to store checkpoints by using the following command.

mkdir checkpoint

CIFAR-10

To reproduce the results on CIFAR dataset from our paper, please follow the command and our hyper-parameters.

First, you can adjust the corruption_prob and corruption_type to obtain different noise rates and noise type.

Second, the reweight_label indicates you are using the our L2B method. You can change it to baseline or mixup.

python  main.py  --arch res18 --dataset cifar10 --num_classes 10 --exp L2B --train_batch_size  512 \
 --corruption_prob 0.2 --reweight_label  --lr 0.15  -clipping_norm 0.25  --num_epochs 300  --scheduler cos \
 --corruption_type unif  --warm_up 10  --seed 0  

CIFAR-100

Most of settings are the same as CIFAR-10. To reproduce the results, please follow the command.

python  main.py  --arch res18 --dataset cifar100 --num_classes 100 --exp L2B --train_batch_size  256  \
--corruption_prob 0.2 --reweight_label  --lr 0.15  --clipping_norm 0.80  --num_epochs 300  --scheduler cos \
--corruption_type unif  --warm_up 10  --seed 0 \ 

ISIC2019

On the ISIC dataset, first you should download the dataset by following command.

Download ISIC dataset as follows:
wget https://isic-challenge-data.s3.amazonaws.com/2019/ISIC_2019_Training_Input.zip
wget https://isic-challenge-data.s3.amazonaws.com/2019/ISIC_2019_Training_GroundTruth.csv \

Then you can reproduce the results by following the command.

python main.py  --arch res50  --dataset ISIC --data_path isic_data/ISIC_2019_Training_Input --num_classes 8 
--exp L2B  --train_batch_size 64  --corruption_prob 0.2 --lr 0.01 --clipping_norm 0.80 --num_epochs 30 
--temperature 10.0  --wd 5e-4  --scheduler cos --reweight_label --norm_type softmax --warm_up 1 

Clothing-1M

First, the num_batch and train_batch_size indicates how many training images you want to use (we sample a balanced training data for each epoch).

Second, you can adjust the num_meta to sample different numbers of validation images to form the metaset. We use the whole validation set as metaset by default.

The data_path is where you store the data and key-label lists. And also change the data_path in the line 20 of main.py. If you have issue for downloading the dataset, please feel free to contact us.

Then you can reproduce the results by following the command.

python main.py --arch res18_224 --num_batch 250 --dataset clothing1m \
--exp L2B_clothing1m_one_stage_multi_runs  --train_batch_size 256  --lr 0.005  \
--num_epochs 300  --reweight_label  --wd 5e-4 --scheduler cos   --warm_up 0 \
--data_path /data1/data/clothing1m/clothing1M  --norm_type org  --num_classes 14 \ 
--multi_runs 3 --num_meta 14313

Contact

Yuyin Zhou

• email: [email protected]

Xianhang Li

• email: [email protected]

If you have any question about the code and data, please contact us directly.