https://github.com/vandit15/Class-balanced-loss-pytorch

Pytorch implementation of the paper "Class-Balanced Loss Based on Effective Number of Samples"
https://github.com/vandit15/Class-balanced-loss-pytorch

computer-vision cvpr2019 deep-learning loss-functions pytorch

Last synced: 5 months ago
JSON representation

Pytorch implementation of the paper "Class-Balanced Loss Based on Effective Number of Samples"

• Host: GitHub
• URL: https://github.com/vandit15/Class-balanced-loss-pytorch
• Owner: vandit15
• License: mit
• Created: 2019-08-31T07:34:08.000Z (about 6 years ago)
• Default Branch: master
• Last Pushed: 2024-02-18T05:57:18.000Z (over 1 year ago)
• Last Synced: 2024-08-04T03:12:00.152Z (about 1 year ago)
• Topics: computer-vision, cvpr2019, deep-learning, loss-functions, pytorch
• Language: Python
• Size: 157 KB
• Stars: 775
• Watchers: 11
• Forks: 120
• Open Issues: 16
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# Class-balanced-loss-pytorch

Pytorch implementation of the paper

[Class-Balanced Loss Based on Effective Number of Samples](https://arxiv.org/abs/1901.05555) presented at CVPR'19.

[Yin Cui](https://ycui.me/), Menglin Jia, [Tsung-Yi Lin](https://vision.cornell.edu/se3/people/tsung-yi-lin/)(Google Brain), [Yang Song](https://ai.google/research/people/author38270)(Google), [Serge Belongie](http://blogs.cornell.edu/techfaculty/serge-belongie/)

## Dependencies

- Python (>=3.6)

- Pytorch (>=1.2.0)

## Review article of the paper

[Medium Article](https://medium.com/@vandit_15/handling-class-imbalanced-data-using-a-loss-specifically-made-for-it-6e58fd65ffab?source=friends_link&sk=ac09ea6061990ead2a2f90e3767ae91f)

## How it works

It works on the principle of calculating effective number of samples for all classes which is defined as:

![alt-text](https://github.com/vandit15/Class-balanced-loss-pytorch/blob/master/samples.png)

Thus, the loss function is defined as:

![alt-text](https://github.com/vandit15/Class-balanced-loss-pytorch/blob/master/loss.png)

Visualisation for effective number of samples

![alt-text](https://github.com/vandit15/Class-balanced-loss-pytorch/blob/master/image.png "Visualisation for effective number of samples")

## References

[official tensorflow implementation](https://github.com/richardaecn/class-balanced-loss)