This is the source code of the paper "FedGA: Federated Learning with Gradient Alignment for Error Asymmetry Mitigation". Please refer to the pdf file for the convergence analysis and proof.

All the experiments were conducted on a Linux machine with the following specifications:

• OS: Red Hat Enterprise Linux 8.5 (Ootpa)
• CPU: Intel(R) Xeon(R) Platinum 8360Y CPU @ 2.40GHz

The code was developed using the following versions of the software:

• Python: 3.11.9

The following packages are required to run the code:

• datasets 2.21.0
• python-box 7.0
• scipy 1.12.0
• torchvision 0.15.2
• torchmetrics 0.10.1
• wandb 0.17.7

Always cd to the root directory of this project before running the experiments.

Sync the dependencies with uv:

uv sync

Activate the virtual environment before running the experiments:

source .venv/bin/activate

To run a short example, use the following command:

python fedlearn.py

To run a specific experiment on dataset SVHN with Dirichlet parameter 0.1 based on GA algorithm, you can supply the configuration as command line argument:

python fedlearn.py DATA.NAME SVHN DATA.IB.ALPHA 0.1 FL.ALG FedGA

or, make a copy of the default config file config_default and modify it and run the following command:

python fedlearn.py --cfg configs/your_config.yaml

To reproduce the results in the paper, run the following command:

python fedlearnParallel.py --cfg configs/config_MNIST.yaml
python fedlearnParallel.py --cfg configs/config_SVHN.yaml
python fedlearnParallel.py --cfg configs/config_CIFAR10.yaml