Federated-Averaging-PyTorch

This repo is the implementation of FedAvg

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

Requirements

path~=16.4.0

torch~=1.10.2

numpy~=1.21.2

fedlab~=1.1.4

torchvision~=0.11.3

tqdm~=4.62.3

pip install -r requirements.txt

It’s so simple.

MNIST, CIFAR-10, Synthetic are supported. Have fun!🤪

python main.py

--comms_round: Num of communication rounds. Default: 40

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

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

--test_round: Num of round for final evaluation. Default: 1

--local_lr: Learning rate for client local model updating. Default: 0.05

--batch_size: Batch size of client local dataset.

--global_lr: Learning rate for server model updating. Default: 1.0

--cuda: True for using GPU. Default: True

--epochs: Num of local epochs in client training. Default: 5

--model: Structure of model. Must be mlp or cnn. Default: cnn

Algorithm	Global Loss	Localized Loss	Global Acc	Localized Acc
FedAvg	`1.7883`	`0.5261`	`90.00%`	`96.40%`

Localization means the model additionally train for 10 local epochs at the final evaluation phase, which is for adapting client’s local dataset.