This repo contains the pytorch implementation of our draft COMPACTER: Efficient Low-Rank Hypercomplex Adapter Layers. This repo additionally contains the implementation for the recent parameter-efficient finetuning methods as well.

python setup.py install

We provide the example scripts to run each model in the paper in seq2seq/scripts folder with their config files in seq2seq/configs. To run the models, please do cd seq2seq and run:

• Full-finetuning (T5):

  bash scripts/baseline.sh
  

• AdapterDrop:

  bash scripts/adapters_drop.sh
  

• Adapters:

  bash scripts/adapters.sh
  

• Low-rank adapters (uses rank-1 approximation for each adapter weight):

  bash scripts/low_rank_adapters.sh
  

• Pfeiffer-Adapters:

  bash scripts/pfeiffer_adapters.sh
  

• BitFit:

  bash scripts/bitfit.sh
  

• Compacter++:

  bash scripts/compacter++.sh
  

• Compacter:

  bash scripts/compacter.sh
  

• PHM-Adapters:

  bash scripts/phm_adapters.sh
  

• Intrinsic-SAID:

  bash scripts/intrinsic_said.sh
  

• Prompt tuning-R (Prompt tuning with random initialization):

  bash scripts/prompt_tuning_random_init.sh
  

• Prompt tuning-T (Prompt tuning with initialization from language model's vocabulary):

  bash scripts/prompt_tuning_tokens_init.sh
  

If you find this repo useful, please cite our work:

@inproceedings{karimi2021parameterefficient,
  title={Parameter-efficient Multi-task Fine-tuning for Transformers via Shared Hypernetworks},
  author={Karimi Mahabadi, Rabeeh and Ruder, Sebastian and Dehghani, Mostafa and Henderson, James},
  booktitle={Annual Meeting of the Association for Computational Linguistics},
  year={2021}
}

To implement the intrinsic-SAID method, we used the codes from the following paper. If using this method, please also consider citing this work:

@inproceedings{aghajanyan-etal-2021-intrinsic,
    title = {Intrinsic Dimensionality Explains the Effectiveness of Language Model Fine-Tuning},
    author = {Aghajanyan, Armen  and Gupta, Sonal  and Zettlemoyer, Luke},
    publisher = {Association for Computational Linguistics},
    year = {2021}
}

To implement parameterized hypercomplex layers, we use the implementation of the following work, if using PHM-adapters/Compacter/Compacter++ please also consider citing this work:

@article{le2021parameterized,
  title={Parameterized hypercomplex graph neural networks for graph classification},
  author={Le, Tuan and Bertolini, Marco and No{\'e}, Frank and Clevert, Djork-Arn{\'e}},
  journal={arXiv preprint arXiv:2103.16584},
  year={2021}
}

Hope this repo is useful for your research. For any questions, please create an issue or email [email protected] or [email protected], and I will get back to you as soon as possible.