Awesome

FedMix: Approximation of Mixup under Mean Augmented Federated Learning

Paper: arxiv.org/abs/2107.00233

Authors: Tehrim Yoon, Sumin Shin, Sung Ju Hwang, Eunho Yang

Abstract: Federated learning (FL) allows edge devices to collectively learn a model without directly sharing data within each device, thus preserving privacy and eliminating the need to store data globally. While there are promising results under the assumption of independent and identically distributed (iid) local data, current state-of-the-art algorithms suffer from performance degradation as the heterogeneity of local data across clients increases. To resolve this issue, we propose a simple framework, Mean Augmented Federated Learning (MAFL), where clients send and receive averaged local data, subject to the privacy requirements of target applications. Under our framework, we propose a new augmentation algorithm, named FedMix, which is inspired by a phenomenal yet simple data augmentation method, Mixup, but does not require local raw data to be directly shared among devices. Our method shows greatly improved performance in the standard benchmark datasets of FL, under highly non-iid federated settings, compared to conventional algorithms.

About this baseline

What’s implemented: The code in this directory implements two of the Federated Strategies mentioned in the paper: NaiveMix and FedMix

Datasets: CIFAR10, CIFAR100, FEMNIST

Hardware Setup: Experiments in this baseline were run on 2x Nvidia Tesla V100 16GB.

Contributors: DevPranjal

Experimental Setup

Task: Image Classification

Model: Models used are modified versions of existing known models and are descirbed in Appendix B (Experimental Details). For CIFAR10 and CIFAR100 datasets, the authors use a modified version of VGG, while LeNet-5 is used for the FEMNIST dataset.

Dataset: The datasets used (CIFAR10, CIFAR100 FEMNIST) incorporate data heterogenity in diferent ways to simulate a non-iid setting. For CIFAR10 and CIFAR100 experiments, data is allocated such that each client has data from only a selected number of randomly chosen classes. For the FEMNIST experiment, data is allocated such that each client has data from only one writer, resulting in 200-300 samples per client on average.

Training Hyperparameters:

Environment Setup

# set local python version via pyenv
pyenv local 3.10.6

# then fix that for poetry
poetry env use 3.10.6

# then install poetry env
poetry install

# activate the environment
poetry shell

Running the Experiments

To setup the directory for FEMNIST experiments, run:

git clone https://github.com/TalwalkarLab/leaf.git
cp -r leaf/data/femnist fedmix/
cp -r leaf/data/utils fedmix/

To run an experiment with default hyperparameters (as mentioned in the paper), execute:

python -m fedmix.main +experiment={dataset}_{strategy}
# dataset can be: cifar10, cifar100, femnist
# strategy can be: fedavg, naivemix, fedmix

To run custom experiments: create a new .yaml file in conf/experiment and override the default hyperparameters.

Expected Results

To get all results, run:

# femnist experiments
python -m fedmix.main --multirun +experiment=femnist_fedavg,femnist_naivemix,femnist_fedmix

# cifar10 experiments
python -m fedmix.main --multirun +experiment=cifar10_fedavg,cifar10_naivemix,cifar10_fedmix

# cifar100 experiments
python -m fedmix.main --multirun +experiment=cifar100_fedavg,cifar100_naivemix,cifar100_fedmix

To generate table (similar to Table 1 of paper) and figures (similar to Figure 2 of paper), run the above commands and then:

# table
python -m fedmix.generate_media --output_type=table --output_path="_static/table.png" --input_directory="results/"

# femnist learning curve
python -m fedmix.generate_media --output_type=figure --output_path="_static/femnist.png" --input_directory="results/" --dataset_name="femnist"

# cifar10 learning curve
python -m fedmix.generate_media --output_type=figure --output_path="_static/cifar10.png" --input_directory="results/" --dataset_name="cifar10"

# cifar100 learning curve
python -m fedmix.generate_media --output_type=figure --output_path="_static/cifar100.png" --input_directory="results/" --dataset_name="cifar100"