README.md

Personalized federated learning with theoretical guarantees: A model-agnostic meta-learning approach [NIPS 2020]

PyTorch Implementation of Per-FedAvg.

For simulating Non-I.I.D scenario, the dataset is split by labels and each client has only two classes of data.

It’s just a toy demo for demonstrating algorithm, so take it easy. 🤣

Note that I have recently released a benchmark of federated learning that includes this method and many ohter baselines. Welcome to check my benchmark and star it! 🤗

Requirements

torch~=1.10.2
path~=16.4.0
numpy~=1.21.2
fedlab~=1.1.4
torchvision~=0.11.3
rich~=12.2.0

pip install -r requirements.txt

Preprocess dataset

MNIST and CIFAR-10 is prepared.🌟

cd data; python preprocess.py

The way of preprocessing is adjustable, more details in each dataset folder's preprocess.py.

Run the experiment

Before run the experiment, please make sure that the dataset is downloaded and preprocessed already.

It’s so simple.🤪

python main.py

Hyperparameters

--global_epochs: Num of communication rounds. Default: 200

--local_epochs: Num of local training rounds. Default: 4

--pers_epochs: Num of personalization rounds (while in evaluation phase). Default: 1

--dataset: Name of experiment dataset. Default: mnist

--fraction: Percentage of training clients in all alients. Default: 0.9

--client_num_per_round: Num of clients that participating training at each communication round. Default: 5

--alpha: Learning rate $\alpha$ in paper. Default: 0.01

--beta: Learning rate $\beta$ in paper. Default: 0.001

--gpu: Non-zero value for using CUDA; 0 for using CPU. Default: 1

--batch_size: Batch size of client local dataset. Default: 40.

--eval_while_training: Non-zero value for performing evaluation while in training phase. Default: 1

--valset_ratio: Percentage of validation set in client local dataset. Default: 0.1

--hf: Non-zero value for performing Per-FedAvg(HF); 0 for Per-FedAvg(FO). Default: 1

--seed: Random seed for init model parameters and selected clients. Default: 17

--log: Non-zero value for recording experiment's output to a .html file in ./log. Default: 0