FedGMA

This is the code of paper Communication-Efficient Generalized Neuron Matching for Federated Learning

Requirements

System requirements:

• Openssh client & server
• Cmake & build essentials
• CUDA runtime >= 11.2
• Tmux >=3.2a

Python requirements:

• Python == 3.9
• Install all requirements with pip install -r requirements.txt.

Usage

First edit workers.txt to configure the GPU resource allocation. Each line should have the format of <server|client> localhost <N>. <N> specifies the GPU index (starting from 0). The first line must be server localhost <N>, and the remaining lines must start with client. The total number of client lines must be greater than the total number of clients to start.

This is an example to run the code:

python -um code.launcher.client_server \
    +a fedavg +n 8 +sa -md.n unordered-resnet18 \
    -dt.r.d cifar10 -fl.bt model -fedavg.eslr 0.04 \
    -tr.bs 128 -tr.s.n reduceonplateau \
    -tr.o.sgd.wghtdcy 5e-4 -tr.e 10000 -tr.lr 0.05

This command starts a tmux session that manages all the processes of the experiment. To query all the available hyperparameters, run:

python -um code.launcher.node.node -l.a <algorithm>.server -h

where <algorithm> is the algorithm to run (i.e. fedavg|fednova|fedgma). The datasets will be downloaded if they are not already present after the experiment starts.

The processes are managed by tmux. Use tmux attach to attach to the tmux session and observe the progress of the experiment. Experiment logs will also be outputted to logs/<date_time>.

Results

Accuracies on CIFAR-10/100

CIFAR10	CIFAR100

Accuracy -- Data Curve on CIFAR10

VGG9	ResNet18	InceptionNet3

Citation

TBD