lijichang / lnl-nce

A pytorch implementation for "Neighborhood Collective Estimation for Noisy Label Identification and Correction", which is accepted by ECCV2022.

LNL-NCE

A pytorch implementation for "Neighborhood Collective Estimation for Noisy Label Identification and Correction", accepted by ECCV2022. More details of this work can be found in our paper: Arxiv or ECCV2022.

Installation

Refer to DivideMix.

Model training

(1) To run training on CIFAR-10/CIFAR-100 with different noise modes (namely sym or asym) and various noise ratios (namely 0.20, 0.50, 0.80, 0.90, etc.),

CUDA_VISIBLE_DEVICES=0 python ./cifar/main.py --dataset cifar10 --num_class 10 --batch_size 128 --data_path ./data/cifar-10/ --r 0.50 --noise_mode sym --remark exp-ID

CUDA_VISIBLE_DEVICES=0 python ./cifar/main.py --dataset cifar100 --num_class 100 --batch_size 128 --data_path ./data/cifar-100/ --r 0.50 --noise_mode sym --remark exp-ID

(2) To run training on Webvision-1.0,

CUDA_VISIBLE_DEVICES=0,1,2 python ./webvision/main.py --data_path ./data/webvision/ --remark exp-ID

Citation

If you consider using this code or its derivatives, please consider citing:

@inproceedings{li2022neighborhood,
  title={Neighborhood Collective Estimation for Noisy Label Identification and Correction},
  author={Li, Jichang and Li, Guanbin and Liu, Feng and Yu, Yizhou},
  booktitle={European Conference on Computer Vision},
  pages={128--145},
  year={2022},
  organization={Springer}
}

Contact

Please feel free to contact the first author, namely Li Jichang, with an Email address [email protected], if you have any questions.

lnl-nce's Issues

label consistency regularization for unlabeled data cifar100

Hi,
It would be great if you can tell me why loss of unlabeled data should be calculated with 'cat[inputs_us, inputs_xs]' in cifar100,but not in other dataset such as cifar10.I didn't find the explanation for that in paper.
Thanks!
The part of the code is:

    # label consistency regularization for unlabeled data
    if (args.dataset == 'cifar100') and ((args.r==0.2) or (args.r==0.5)):
        all_inputs, all_labels, all_masks = torch.cat([inputs_us, inputs_xs], dim=0), torch.cat([labels_u,  labels_x], dim=0), torch.cat([mask_u, mask_x], dim=0)
        all_logits = net(all_inputs)
        Lu = (F.cross_entropy(all_logits, all_labels, reduction='none') * all_masks.float()).mean()
    else:
        logits_us = net(inputs_us)
        Lu = (F.cross_entropy(logits_us, labels_u, reduction='none') * mask_u.float()).mean()

Code release

Your work is really interesting and impressive!

And I wonder when the code will be released?

How to run on LNL-NCE custom datasets ?

I would like to know how to run LNL-NCE on custom datasets, the repository code seems very specific to datasets, have you made LNL-NCE into a tool?

Thanks.

Symmetric 20% reproducibility

Hi. First of all, I really want to say I am really inspired by your awesome work!

However, while reproducing the results from the paper I've encountered an issue.

With all the hyperparameters set same as the paper I can't reproduce the same result with 20% symmetric noise.

The paper says the best results is 96.2±0.09, but what I get consistently is around 95.2.

Is there something I missed? I would really appreciate your response, thanks!