https://github.com/takashiishida/comp
[NeurIPS 2017] [ICML 2019] Code for complementary-label learning
https://github.com/takashiishida/comp
deep-learning machine-learning
Last synced: 9 months ago
JSON representation
[NeurIPS 2017] [ICML 2019] Code for complementary-label learning
- Host: GitHub
- URL: https://github.com/takashiishida/comp
- Owner: takashiishida
- License: mit
- Created: 2019-05-06T00:40:59.000Z (about 7 years ago)
- Default Branch: master
- Last Pushed: 2024-01-21T09:51:32.000Z (over 2 years ago)
- Last Synced: 2025-04-11T02:56:52.638Z (about 1 year ago)
- Topics: deep-learning, machine-learning
- Language: Python
- Homepage:
- Size: 3.21 MB
- Stars: 48
- Watchers: 1
- Forks: 17
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# Complementary-Label Learning
This repository gives the implementation for *complementary-label learning* from the ICML 2019 paper [1], the ECCV 2018 paper [2], and the NeurIPS 2017 paper [3].
## Requirements
- Python 3.6
- numpy 1.14
- PyTorch 1.1
- torchvision 0.2
## Demo
The following demo will show the results with the MNIST dataset. After running the code, you should see a text file with the results saved in the same directory. The results will have three columns: epoch number, training accuracy, and test accuracy.
```bash
python demo.py -h
```
#### Methods and models
In `demo.py`, specify the `method` argument to choose one of the 5 methods available:
- `ga`: Gradient ascent version (Algorithm 1) in [1].
- `nn`: Non-negative risk estimator with the max operator in [1].
- `free`: Assumption-free risk estimator based on Theorem 1 in [1].
- `forward`: Forward correction method in [2].
- `pc`: Pairwise comparison with sigmoid loss in [3].
Specify the `model` argument:
- `linear`: Linear model
- `mlp`: Multi-layer perceptron with one hidden layer (500 units)
## Reference
1. T. Ishida, G. Niu, A. K. Menon, and M. Sugiyama.
**Complementary-label learning for arbitrary losses and models**.
In *ICML 2019*.
[[paper]](https://arxiv.org/abs/1810.04327)
2. Yu, X., Liu, T., Gong, M., and Tao, D.
**Learning with biased complementary labels**.
In *ECCV 2018*.
[[paper]](https://arxiv.org/abs/1711.09535)
3. T. Ishida, G. Niu, W. Hu, and M. Sugiyama.
**Learning from complementary labels**.
In *NeurIPS 2017*.
[[paper]](https://arxiv.org/abs/1705.07541)
If you have any further questions, please feel free to send an e-mail to: ishida at ms.k.u-tokyo.ac.jp.