This repo holds the source code and scripts for reproducing the key experiments of our paper:

"Personalized Federated Learning via ADMM with Moreau Envelope".

Author: Shengkun Zhu, Jinshan Zeng, Sheng Wang, Yuan Sun, Zhiyong Peng.

The default values for various parameters parsed to the experiment are given in options.py. Details are given on some of those parameters:

• framework: two personalied federated learnign frameworks, i.e., FLAME and pFedMe.

• num_users: number of users.

• q: number of data shards of each user.

• model: SVM, MLP, MLR, CNN for choices.

• dataset: four datasets for choices.

• strategy: client selection strategy.

• frac_candidates: fraction of clients candidates, c/m in our paper.

• frac: fraction of clients, s/m in our paper.

• optimizer: type of optimizer, default sgd.

• momentum: sgd momentum, default 0.

• epoches: number of communication rounds.

• local_ep: number of local iterations.

• local_bs: local batch size.

• lr: learning rate.

• mu: hyperparameter in regularization term.

• Lambda: hyperparameter in Moreau envelope.

• rho: hyperparameter in penalty term.

• iid: data distribution, 0 for non iid.

• seed: random seed.

• eta: learning rate for global model in pFedMe.