Usage

Setup

See the Pipfile for python environment requirements.

pipenv update

Dataset partitioning

The dataset partitioning is inspired by Federated Learning on Non-IID Data Silos: An Experimental Study.

. ./shell/fog_partitions.sh {fog_partitions} {client_partitions}
. ./shell/client_partitions.sh {fog_partitions} {client_partitions}

• fog_partitions: dataset partitioning for fog servers
• client_partitions: dataset partitioning for clients

There are three variants for partitioning.

• iid: each participant has 10 labels
• noniid-label2: each participant has 2 labels
• noniid-label1: each participant has 1 label

Run Federated Learning

To simulate the large cohorts of clients, we leverage Ray following the code examples provided by Flower.

. ./run_fedfog.sh {fog_partitions} {client_partitions}
. ./run_fml.sh {fog_partitions} {client_partitions} {alpha} {beta}
. ./run_f2mkd.sh {fog_partitions} {client_partitions} {alpha} {beta}

alpha and beta are FL strategy-specific hyperparameters.