The official implementation of 'Provable Multi-instance Deep AUC Maximization with Stochastic Pooling', ICML2023
Here are some dependencies (with the version for the experiments results reported in the paper): torch==1.9.0, numpy==1.17.4, CUDA version:12.0 on NVIDIA Quadro RTX 6000 card.
This is the code that runs MIDAM with stochastic softmax or attention pooling on benchmark datasets, such as MUSK1&2, Fox, Tiger, Elephant, Breast Cancer, etc.
CUDA_VISIBLE_DEVICES=0 python3 experiment.py --dataset=Fox --loss=MIDAM-att --momentum=0.1 --seed=123 --lr=1e-2 --bag_batch_size=4
CUDA_VISIBLE_DEVICES=0 python3 experiment.py --dataset=Fox --loss=MIDAM-smx --momentum=0.1 --seed=123 --lr=1e-2 --bag_batch_size=4
Please make sure you have the data on the data folder (some smaller datasets are already included). Please refer to the experiment.py and datasets.py for how to load the data.
@inproceedings{zhu2023provable,
title={Provable Multi-instance Deep AUC Maximization with Stochastic Pooling},
author={Dixian Zhu and Bokun Wang and Zhi Chen and Yaxing Wang and Milan Sonka and Xiaodong Wu and Tianbao Yang},
booktitle={Proceedings of the 40th International Conference on Machine Learning},
year={2023}
}