This repo is the official implementation of "Sliceformer: Make Multi-head Attention as Simple as Sorting in Discriminative Tasks".
We provide the Sliceformer code for various tasks, including the LRA benchmark, image classification, text classification, and molecular property prediction.
The LRA directory contains the code for the Long-Range Arena benchmark.
The sliceformer directory covers tasks related to image classification, which includes datasets such as CIFAR-10, CIFAR-100, MNIST, and the Dogs vs. Cats dataset. It also includes text classification tasks with the IMDB dataset.
The Graphormer directory is dedicated to tasks involving molecular property prediction, specifically on the PCQM4M-LSC dataset.
For more detailed information, please refer to the README files in each of these three directories.
If you find out work useful, please cite our paper at:
@misc{yuan2023sliceformer,
title={Sliceformer: Make Multi-head Attention as Simple as Sorting in Discriminative Tasks},
author={Shen Yuan and Hongteng Xu},
year={2023},
eprint={2310.17683},
archivePrefix={arXiv},
primaryClass={cs.LG}
}