This is a implemention of FedAvg in paper Communication-Efficient Learning of Deep Networks from Decentralized Data.
At first, you should creat two dirs called 'logs' and 'checkpoints', then you can cd into 'src', and run FedAvg on iid cifar10 and mnist
python3 federated_main.py --model=cnn --dataset=cifar --iid=1 --epochs=300 --lr=0.01 --local_ep=5 --local_bs=32
python3 federated_main.py --model=cnn --dataset=mnist --iid=1 --epochs=100 --lr=0.01 --local_ep=5 --local_bs=32
noniid cifar10 and mnist:
python3 federated_main.py --model=cnn --dataset=cifar --iid=0 --epochs=300 --lr=0.01 --local_ep=5 --local_bs=32
python3 federated_main.py --model=cnn --dataset=mnist --iid=0 --epochs=100 --lr=0.01 --local_ep=5 --local_bs=32
For best test acc：
| - | CIFAR10 | MNIST |
|---|---|---|
| IID | 68.70% | 98.87% |
| Non-IID | 68.11% | 98.05% |
test acc curve(for cifar10 both in iid and noniid setting):
