Relational Surrogate Loss Learning (ReLoss)

Official implementation for paper "Relational Surrogate Loss Learning" in International Conference on Learning Representations (ICLR) 2022.

By Tao Huang, Zekang Li, Hua Lu, Yong Shan, Shusheng Yang, Yang Feng, Fei Wang, Shan You, Chang Xu.

Usage

Install ReLoss

pip install git+https://github.com/hunto/ReLoss.git

Or install for development:

git clone https://github.com/hunto/ReLoss
cd ReLoss
pip install -e .

Train models with ReLoss

All the inputs and outputs of ReLoss are the same as the original loss.

• classification

  from reloss.cls import ReLoss
  loss_fn = ReLoss()

• human pose estimation

  from reloss.pose import ReLoss
  loss_fn = ReLoss(heatmap_size=(64, 48))

• non-autoregressive neural machine translation

  The loss should be used in fairseq framework. You can add it into the criterions.

Train ReLoss

You can train your own ReLoss, please see example/train_reloss/README.md for instructions.

Citation

@inproceedings{
huang2022relational,
title={Relational Surrogate Loss Learning},
author={Tao Huang and Zekang Li and Hua Lu and Yong Shan and Shusheng Yang and Yang Feng and Fei Wang and Shan You and Chang Xu},
booktitle={International Conference on Learning Representations},
year={2022},
url={https://openreview.net/forum?id=dZPgfwaTaXv}
}