Official implementation for "LFME: A Simple Framework for Learning from Multiple Experts in Domain Generalization" in NeurIPS 2024.
Simple combination of two baselines (i.e. CE and MSE losses)
logits = self.predict(all_x)
loss_ce = F.cross_entropy(logits, all_y)
all_y_one_hot = F.one_hot(all_y, num_classes=logits.shape[1]).float()
loss_mse = F.mse_loss(logits, all_y_one_hot)
#alpha can be value, personally recommend it within the range of [0.01, 10]
loss = loss_ce + alpha*loss_mse
If you find this work useful, please consider citing it.
@inproceedings{chen2024lfme,
title={LFME: A Simple Framework for Learning from Multiple Experts in Domain Generalization},
author={Chen, Liang and Zhang, Yong and Song, Yibing and Shen, Zhiqiang and Liu, Lingqiao},
booktitle={NeurIPS},
year={2024}
}
Please contact me via email ([email protected]) if your have any questions regarding this project.