ydkim1293/nlnl-negative-learning-for-noisy-labels

NLNL: Negative Learning for Noisy Labels

NLNL-Negative-Learning-for-Noisy-Labels

Pytorch implementation for paper NLNL: Negative Learning for Noisy Labels, ICCV 2019

Paper: https://arxiv.org/abs/1908.07387

Requirements

• python3
• pytorch
• matplotlib

Generating noisy data

python3 noise_generator.py --noise_type val_split_symm_exc

Start training

Simply run sh file: run.sh

GPU=0 setting='--dataset cifar10_wo_val --model resnet34 --noise 0.2 --noise_type val_split_symm_exc'
CUDA_VISIBLE_DEVICES=$GPU python3 main_NL.py $setting
CUDA_VISIBLE_DEVICES=$GPU python3 main_PL.py $setting --max_epochs 720
CUDA_VISIBLE_DEVICES=$GPU python3 main_pseudo1.py $setting --lr 0.1 --max_epochs 480 --epoch_step 192 288
CUDA_VISIBLE_DEVICES=$GPU python3 main_pseudo2.py $setting --lr 0.1 --max_epochs 480 --epoch_step 192 288

Citation

@inproceedings{kim2019nlnl,
  title={Nlnl: Negative learning for noisy labels},
  author={Kim, Youngdong and Yim, Junho and Yun, Juseung and Kim, Junmo},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision},
  pages={101--110},
  year={2019}
}

nlnl-negative-learning-for-noisy-labels's Issues

tabular data/new datasets

Hi,
thanks for sharing your implementation. I have two questions about it:

1. Does it also work on tabular data?
2. Is the code tailored to the datasets used in the paper or can one apply it to any data?

Thanks!

Implementation Problem for NLNL loss

Hi,

Thanks for this great implementation. When I ran this code, I found that the implementation of the NLNL loss might exist a problem that will cause loss to be Nan. It might be caused by this line:

NLNL-Negative-Learning-for-Noisy-Labels/main_NL.py

Line 193 in 921dec4

labels = labels*0 - 100

This line will make all labels be -100. Would you mind checking this issue?
Best Regards,
Hongxin

The loss function doesn't match the one mentioned in Sec 3 of paper ?

Hello ,I'm very interested in your work and trying to reproduce your results.
Q1:
I found the loss function in (line 200,main_NL.py)
( (loss+loss_neg) / (float((labels>=0).sum())+float((labels_neg[:,0]>=0).sum())) ).backward()
It seems that it dosen't match (Eq.2) or any loss function mentioned in (Section 3)?

Q2:
'main_NL.py','main_PL.py','main_pseudo1.py','main_pseudo2.py' are very similar to each other. Could you summarize the differeces and provide more comments or instructions of the code ?

Thanks a lot!

Request to put trained models on CIFAR10/CIFAR100/Fashion MNIST

Hi there,

I am very curious to see the results shown in the paper, it seems impressive! Can you provide trained models weights so that I can use it directly for inference?

You can email me on [email protected]

Waiting for your reply.

-Ishan

这篇论文这么水？？都没有人用

有人试了下，可以区分噪声图片。

question on running error

Hello,
I followed the instructions but received error says

FileNotFoundError: [Errno 2] No such file or directory: 'logs/cifar10_wo_val_resnet34_val_split_symm_exc_20/checkpoint_epoch1439.pth.tar'

is there anywhere I should go to download the file?

Thank you so much for your time!

1

Hello. In your paper, you mentioned using negtive labels that are diffent from the true labels to train the network, but how can we get these true labels under the unsupervised situation?

How to select the samples after NL?

After NL, the next step is SelNL, but i have problems with "py > 1/c". According to my understanding, the ideal condition is that the
network will output a low probability corresponding to a complementary label after NL. If we select the samples with output probability over 1/c, did we select the data that can not be effectively splitted after NL? So what exactly the 'py' means? Or what the confidence 'py' represents?