m2mKD

This repository contains the checkpoints for m2mKD: Module-to-Module Knowledge Distillation for Modular Transformers.

Released checkpoints

For the usage of the checkpoints listed below, please refer to the instructions provided on our GitHub repo.

• nac_scale_tinyimnet.pth/nac_scale_imnet.pth: NAC model with a scale-free prior trained using m2mKD.
• vmoe_base.pth: V-MoE-Base model trained using m2mKD.
• FT_huge: a directory containing DeiT-Huge teacher modules for NAC model training.
• nac_tinyimnet_students: a directory containing NAC student modules for Tiny-ImageNet.

Acknowledgement

Our implementation is mainly based on Deep-Incubation.

Citation

If you use the checkpoints, please cite our paper:

@misc{lo2024m2mkd,
    title={m2mKD: Module-to-Module Knowledge Distillation for Modular Transformers}, 
    author={Ka Man Lo and Yiming Liang and Wenyu Du and Yuantao Fan and Zili Wang and Wenhao Huang and Lei Ma and Jie Fu},
    year={2024},
    eprint={2402.16918},
    archivePrefix={arXiv},
    primaryClass={cs.LG}
}