Does Federated Learning Really Need Backpropagation?

Here is the official implementation of the model BAFFLE in paper Does Federated Learning Really Need Backpropagation?

Setup

Install Package Dependencies

python >= 3.6
torch >= 1.2.0
torchvision >= 0.4.0
wandb
ast
torch_ema

BAFFLE For MNIST with IID Participation ($C=10,\sigma=10^{-4}$)

Evaluate BAFFLE with Different K

# For LeNet
python main.py --K 100 --dataset MNIST --epochs 20 --lr 1e-2 --optimizer Adam --ema --normalization NoNorm 
# For WRN-10-2
python main.py --K 100 --net-name wideresnet --depth 10 --width 2 --dataset MNIST --epochs 20 --lr 1e-2 --optimizer Adam  --ema

Evaluate BAFFLE with Different Guidelines

# w/o EMA
python main.py --K 100 --dataset MNIST --epochs 20 --lr 1e-2 --optimizer Adam --normalization NoNorm
# ReLU 
python main.py --K 100 --dataset MNIST --epochs 20 --lr 1e-2 --optimizer Adam --normalization NoNorm --ema --activation "ReLU"
# SELU
python main.py --K 100 --dataset MNIST --epochs 20 --lr 1e-2 --optimizer Adam --normalization NoNorm --ema --activation "SELU"
# Central Scheme
python main.py --K 50 --fd-format center  --dataset MNIST --epochs 20 --lr 1e-2 --optimizer Adam  --normalization NoNorm --ema 

Evaluate BAFFLE with Epoch-level Communications

# One Epoch Communication
python main.py --K 100 --fedopt FedAvg --dataset MNIST --epochs 20 --lr 1e-2 --optimizer Adam  --ema --normalization NoNorm