kthyeon/fine_official

NeurIPS 2021, "Fine Samples for Learning with Noisy Labels"

[Official] FINE Samples for Learning with Noisy Labels

This repository is the official implementation of "FINE Samples for Learning with Noisy Labels" paper presented in NeurIPS 2021. New version of previous repository https://github.com/jaychoi12/FINE. Future code modifications and official developments will take place here. Thanks to the contributors in the previous repo.

• Paper, NeurIPS 21, FINE Samples for Learning with Noisy Labels, [Arxiv][OpenReview]

Reference Codes

We refer to some official implementation codes

• https://github.com/bhanML/Co-teaching
• https://github.com/LiJunnan1992/DivideMix
• https://github.com/shengliu66/ELR

Requirements

• This codebase is written for python3 (used python 3.7.6 while implementing).
• To install necessary python packages, run pip install -r requirements.txt.

Training

Sample-Selection Approaches and Collaboration with Noise-Robust loss functions

• Most code strucutres are similar with the original implementation code in https://github.com/bhanML/Co-teaching and https://github.com/shengliu66/ELR.
• If you want to train the model with FINE, move to the folder dynamic_selection and run the bash files by following the README.md.

Semi-Supervised Approaches

• Most codes are similar with the original implementation code in https://github.com/LiJunnan1992/DivideMix.
• If you want to train the model with FINE (f-dividemix), move to the folder dividemix and run the bash files by following the README.md in the dividemix folder.

Results

You can reproduce all results in the paper with our code. All results have been described in our paper including Appendix. The results of our experiments are so numerous that it is difficult to post everything here. However, if you experiment several times by modifying the hyperparameter value in the .sh file, you will be able to reproduce all of our analysis.

Contact

• Jongwoo Ko : [email protected]
• Taehyeon Kim : [email protected]

License
This project is licensed under the terms of the MIT license.

Acknowledgements

This work was supported by Institute of Information & communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) [No.2019-0-00075, Artificial Intelligence Graduate School Program (KAIST)] and [No. 2021-0-00907, Development of Adaptive and Lightweight Edge-Collaborative Analysis Technology for Enabling Proactively Immediate Response and Rapid Learning].

fine_official's Issues

non-exist functions

Hi, thanks for sharing your implementation.
In your implementation, threr are some code error.
The functions ,get_out_list and get_singular_value_vector, used in FINE_official/dynamic_selection/traintools/gtrobustlosstrain.py is not really so non-existent.
Can you give me details about these functions?

Motivation of the method

Hello,

I have read your paper and find it very interesting. However, I may have some confusion about your method. If I understood correctly, the first eigenvector represents the latent distribution of a class, which is similar as the function of a prototype. And I also saw some methods utilize the similarity between a sample and the class-prototype to select clean samples. I would like to know what is the advantage of using the eigenvector over prototypes.

Thanks.

tabular data/ noisy instances/ new datasets

Hi,
thanks for sharing your implementation. I have some 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?
3. Is it possible to identify the noisy instances (return the noisy IDs or the clean set)?

Thanks!

Could you provide hyperparameters?

I found some hyperparameters are different from the description in the paper, "All hyper-parameters settings are the same with [25], even for the clean probability threshold.". For example, the threshold in the code is 0.6, but in DivideMix the threshold is 0.5. The number of warmup is also different from DivideMix. Could you provide hyperparameters for reproducing the results?