This is the code for the paper: Confidence Scores Make Instance-dependent Label-noise Learning Possible Antonin Berthon, Bo Han, Gang Niu, Tongliang Liu, Masashi Sugiyama
To be presented at ICML 2021.
If you find this code useful in your research then please cite
@inproceedings{berthon2021confidence,
title={Confidence scores make instance-dependent label-noise learning possible},
author={Berthon, Antonin and Han, Bo and Niu, Gang and Liu, Tongliang and Sugiyama, Masashi},
booktitle={ICML},
pages={825--836},
year={2021}
}
All the experiments were ran on NVIDIA Tesla K80 GPUs from Google Colab notebooks.
As described in the paper, we experiment on a synthetic dataset as well as CIFAR10 and SVHN by generating some instance-dependant label noise. Please contact berthon.antonin[at]gmail[dot]com to receive a download link.
You can run the experiments on synthetic, CIFAR10 and SVHN experiments on the following Google Colab Notebook.
An example of training of ILFC and benchmark models on Clothing1M can be found in the following Google Colab Notebook.
Train ILFC on the Clothing1M dataset:
python experiments.ILFC_clothing.py --seed 123 --import_data_path <path to clothing1m dataset> \
--noisy_model_epochs 0 --mom_decay_start 10 --warm_start 3 --bs 64 --method mean \
--model resnet18 --nb_epoch 60 --noisy_model resnet18 --lr 0.0001
--train_limit 2000 --optim "Adam" \
--result_dir <results export path> \
--model_export <model export path>
# Optional
# --noisy_model_import <path to naive model checkpoint>
# --noisy_model_export <path to naive model checkpoint>
Train a benchmark method on the Clothing1M dataset:
python experiments.COMP_clothing.py --seed 1 --import_data_path <path to clothing&m dataset>" \
--nb_epoch 10 --lr 0.0001 --comp_model <One of {"MAE", "LQ", "F"}> \
--mom_decay_start 5 --bs 64 --train_limit 6000 \
--result_dir <results export path> \
--model_export <model export path>
# --model_import <model import path>