‘DA-DPFL’: Dynamic Aggregation & Decentralized Personalized Federated Learning

Running Environment:

Structure (description of folders)

data

• the folder to store the data

fedml_api (interface for multiple modules)

• data_preprocessing: processing tools

• model: the folder to store the model (parent models) and training tools

• standalone: the folder to store the different algorithms

  • client.py: the functions executed in client such as training and pruning

  • api.py: the functions executed in the server, i.e. aggregation, and the whole logic in algorithm.

  • model_trainer.py: store the class of model for corresponding algorithm

• utils: the folder to store the tools for logging, FLOPS computation, etc.

fedml_core

• the folder to store the core functions of FedML

fedml_xxx

• main running interface for different baselines, where .sh files stored.

Example to use the code:

Change the directory to the root of the your project

replace work_dir

/nfs/da-dpfl/

in config.yaml with the root of your project

/your_path_to_project/

Install dependencies and setup the permissions

pip3 install -r requirements.txt

sh setup_permission.sh

Run CIFAR10 experiments

Format - sh /your_directory/algorithm_name/data_name.sh

- sh /your_path_to_fedml/fedml_dadpfl/cifar10.sh
- sh /your_path_to_fedml/fedml_dispfl/cifar10.sh

Acknowledgements (Codes)

• https://github.com/diaoenmao/Pruning-Deep-Neural-Networks-from-a-Sparsity-Perspective
• https://github.com/rong-dai/DisPFL
• https://github.com/liboyue/beer

Citation

If you find this work useful, please consider citing:

@ARTICLE{11060892,
  author={Long, Qianyu and Wang, Qiyuan and Anagnostopoulos, Christos and Bi, Daning},
  journal={IEEE Transactions on Neural Networks and Learning Systems}, 
  title={Decentralized Personalized Federated Learning Based on a Conditional “Sparse-to-Sparser” Scheme}, 
  year={2025},
  volume={},
  number={},
  pages={1-15},
  keywords={Training;Costs;Computational modeling;Adaptation models;Convergence;Data models;Servers;Network topology;Topology;Federated learning;Decentralized federated learning (DFL);model pruning;personalized FL (PFL);sparsification},
  doi={10.1109/TNNLS.2025.3580277}}