SPADE: State Space Augmented Transformer

This PyTorch package implements the language modeling experiments in Efficient Long Sequence Modeling via State Space Augmented Transformer . For Hugging Face Transformers-style implementation for fine-tuning experiments, refer to this repo.

Dependencies

The package runs on PyTorch v1.11.0 with CUDA 11.3. Note that installation of CUDA and cuDNN are required.
Our implementation requires fairseq v0.11+.

Installation

Download fairseq. Place spade-modules/ in the downloaded directory.
- If training of S4 is desirable (see instructions), replace fairseq/trainer.py with the provided trainer.py file. Specifically, we modified the trainer.py/_build_optimizer() function to add separate S4 parameter groups.
Run pip install -e . to install fairseq locally.
Run pip install -r requirements.txt to install dependencies.
Our implementation requires computing the Cauchy kernel. Note that spade-modules/s4.py largely depends on this file. There are two possible ways:
- (Recommended) Install PyKeOps using pip install pykeops.
- Download the extensions/ folder from this repo, and place the downloaded folder to spade-modules/extensions/. Run python setup.py install within spade-modules/extensions/cauchy/ to install the CUDA kernel.

Instructions

Follow the instructions here to pre-process wikitext-103 data. Run bash run_lm.sh to train a SPADE model.
Empirically, we can initialize the S4 module and then freeze its parameters during training. We find that training the S4 parameters only provides marginal performance gain.
Learning rates of S4 parameters are specified using the --s4-lr argument, which is a dictionary containing three keys: A, B and dt. Each key can use three types of values:
- null means the parameter uses the same learning rate as model parameters;
- 0.0 means the parameter is not trained;
- float (>0.0) specifies a learning rate.

Notes

Contact Information

For personal communication related to this package, please contact Simiao Zuo (simiaozuo@gatech.edu), Xiaodong Liu (xiaodl@microsoft.com), or Jian Jiao (jian.jiao@microsoft.com).

Contributing

This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.

When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.

This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact opencode@microsoft.com with any additional questions or comments.

Trademarks

This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.

Reference

Please cite the following paper if you use this package:

@article{zuo2022efficient,
  title={Efficient Long Sequence Modeling via State Space Augmented Transformer},
  author={Zuo, Simiao and Liu, Xiaodong and Jiao, Jian and Charles, Denis and Manavoglu, Eren and Zhao, Tuo and Gao, Jianfeng},
  journal={arXiv preprint arXiv:2212.08136},
  year={2022}
}

microsoft/EfficientLongSequenceModeling