SCARCE

This repository is the official implementation of the paper "Learning with Complementary Labels Revisited: The Selected-Completely-at-Random Setting Is More Practical" and technical details of this approach can be found in the paper.

Requirements:

• Python 3.6.13
• numpy 1.19.2
• Pytorch 1.7.1
• torchvision 0.8.2
• pandas 1.1.5
• scipy 1.5.4

Arguments:

• lr: learning rate
• bs: training batch size
• ds: dataset
• me: method name, i.e. SCARCE
• mo: model
• wd: weight decay
• gpu: the gpu index
• ep: training epoch number
• bs: training batch size
• op: optimizer
• gen: generation process of complementary labels
• run_times: random running times

Demo:

python main.py -ds mnist -gen random -me SCARCE -mo mlp -op adam -lr 1e-3 -wd 1e-5 -bs 256 -ep 200 -seed 0 -gpu 0

Citation

@inproceedings{wang2024learning,
    author = {Wang, Wei and Ishida, Takashi and Zhang, Yu-Jie and Niu, Gang and Sugiyama, Masashi},
    title = {Learning with complementary labels revisited: The selected-completely-at-random setting is more practical},
    booktitle = {Proceedings of the 41st International Conference on Machine Learning},
    year = {2024}
}