EricLoong/feddip

The official code for ICDM2023 paper: ' FedDIP: Federated Learning with Extreme Dynamic Pruning and Incremental Regularization'

FedDIP

How to use?

1. All functions (procedures) used on clients are stored in 'client.py', including different training functions for different algorithms.
2. Functions for the server to execure are stored in ‘fl_functions’.
3. Utilization tools are stored in 'functions_new.py'.
4. 'exp_args' include all arguments for experiments.
5. 'fl_server' stores the FedDIP algorithm. If no 'prune' and 'sparse', it reduces to Fedavg.

How to run experiments? (Example: FedDIP and FedDST; Definition of arguments refers to exp_args.py)

Notice: Use os.chdir() to change the directory to the folder of your project or use absolute path to run the experiments.

1. Run FedDIP on CIFAR10 with AlexNet

FedDIP: -python fl_server.py --epochs 1000 --num_users 50 --local_ep 5 --amount_sparsity 0.95 --model alexnet --dataset cifar10 --frac 0.1 --parallel 0 --num_workers 2 --reconfig_interval 5 --init_sparsity 0.5 --sparse --prune

2. Run FedDST on CIFAR100 with ResNet18

FedDST: -python feddst.py --feddst --epochs 1000 --num_users 50 --local_ep 5 --amount_sparsity 0.90 --model resnet18 --dataset cifar100 --frac 0.1 --parallel 0 --num_workers 2 --reconfig_interval 20

3. Change differnt PRIFIX.py files to run different baseline algorithms according to their names.

Acknowledgements

Some code is modified based on the following repositories:

1. https://github.com/rong-dai/DisPFL
2. https://github.com/jiangyuang/PruneFL
3. https://github.com/namhoonlee/snip-public

Citation

If you find this work useful, please consider citing:

@inproceedings{long2023feddip,
  title={Feddip: Federated learning with extreme dynamic pruning and incremental regularization},
  author={Long, Qianyu and Anagnostopoulos, Christos and Parambath, Shameem Puthiya and Bi, Daning},
  booktitle={2023 IEEE International Conference on Data Mining (ICDM)},
  pages={1187--1192},
  year={2023},
  organization={IEEE}
}