This is the PyTorch implementation of SeqBoat 🚤 proposed in our paper. This repository is based on MEGA and the fairseq package v0.9.0.

• [Nov. 26] Added a standalone CIFAR-10 training script of SeqBoat for quickstart!
• [Nov. 5] Released training scripts for enwik8 and added a standalone implementation of SeqBoat here!
• [Sep. 21] Our paper is accepted by NeurIPS 2023!
• [July 18] Released training scripts for LRA and Speech Commands.

1. The compress and extract operators for Sparse Modular Activation (SMA) are implemented in fairseq/modules/seqboat_utils.py with the functions compress_seq and extract respectively.
2. SeqBoat layer is implemented in fairseq/modules/seqboat_unit.py.

This repository requires Python 3.8+ and Pytorch 1.11+.

# Install from this repo
pip install -e .

For faster training, install NVIDIA's apex library following fairseq.

The easiest way to get started is to run the standalone_cifar.py script. This script trains a simple SeqBoat model on CIFAR-10:

python standalone_cifar.py --prenorm

• Long Range Arena
• Speech Classification
• Language Modeling

We also provide the training and testing scripts for each of the tasks in the experiment directory.

If you find our work useful, please consider citing:

@inproceedings{ren2023sparse,
  title={Sparse Modular Activation for Efficient Sequence Modeling},
  author={Liliang Ren and Yang Liu and Shuohang Wang and Yichong Xu and Chenguang Zhu and ChengXiang Zhai},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
  year={2023},
  url={https://openreview.net/forum?id=TfbzX6I14i}
}

SeqBoat is under MIT license. The license also applies to model checkpoints.

Liliang Ren (liliang3@illinois.edu)