This repo contains our implementation for our research "Fed-EE: Federating Heterogeneous ASR Models using Early-Exit Architectures." The paper was accepted at the ENLSP (2023) workshop, hosted by NeurIPS, 2023. Also our under-review paper in Computer Speech & Language journal "Federating Dynamic Models using Early-Exit Architectures for Automatic Speech Recognition on Heterogeneous Clients"
client.py : Basic code of a client with fixed speakers
script.py : Training and Evaluation loop (based on client.py)
inference.py: Performs the inference on both dev and test sets (called by script.py)
dynamic_client.py: Code of a client with dynamic speakers
dynamic_script.py: Training and evaluation loop (based on dyamic_client.py)
conf.py: Model and training configuration parameters
dataset.py: Dataloaders and loading functions for Librispeech and TedLium-3 datasets
tedlium_dataset.py: Advanced dataloader for TedLium-3
client.py (it runs a single process that simulates multiple client.py)
inference.py -- model model_name --round round_number (it evaluates the saved central model on the dev and test sets)
More results with in-depth discussion are found in our research paper.
@inproceedings{nawar2023fed,
title={Fed-EE: Federating Heterogeneous ASR Models using Early-Exit Architectures},
author={Nawar, Mohamed Nabih Ali Mohamed and Falavigna, Daniele and Brutti, Alessio},
booktitle={Proceedings of 3rd Neurips Workshop on Efficient Natural Language and Speech Processing},
year={2023}
}
@misc{ali2024federatingdynamicmodelsusing,
title={Federating Dynamic Models using Early-Exit Architectures for Automatic Speech Recognition on Heterogeneous Clients},
author={Mohamed Nabih Ali and Alessio Brutti and Daniele Falavigna},
year={2024},
eprint={2405.17376},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2405.17376},
}
- We acknowledge the support of the PNRR project FAIR - Future AI Research (PE00000013), under the NRRP MUR program funded by the NextGenerationEU.