Simple Federated Learning with Flower

• By Seok-Ju Hahn ([email protected])

Requirements

flwr>=0.17
torch>=1.10
torchvision>=0.9
PyTDC>=0.3
rdkit-pypi>=2021.9.2.1
scikit-learn>=1.0.1

or you can simply execute: pip install -r requirments.txt

Run (example)

python3 server.py --dataset_name mnist --iid \
-N 100 -K 0.1 -R 500 -B 10 -E 5 \
--test_fraction 0 \
--seed 951023

Configurations

• --dataset_name: (required) which data to use for federated learning: {MNIST | CIFAR10 | TOX21}
• -N or --num_clients: total number of clients participating in federated learning
• -K or --fraction: fraction of participating clients at each round
• -B or --batch_size: batch size for client-side update/evaluation
• -E or --num_epochs: number of local epochs required for client-side update
• -R or --num_rounds: number of total rounds
• -S or --num_shards: number of resulting shards used for splitting dataset in non-IID setting (valid only if --iid argument is not passed)
• --iid: wheter to split data in an IID manner; if not used, the dataset is split into pathological non-IID setting propsed in (McMahan et al., 2017)
• test_fraction: fraction of testset for each client (set to zero if only centralized evluation (i.e. server-side evaluation)) is required)
• --data_path: path to store dataset
• --seed: random seed

Datasets

Tox21

Dataset Description:: Tox21 is a data challenge which contains qualitative toxicity measurements for 7,831 compounds on 12 different targets, such as nuclear receptors and stree response pathways.

Task Description: Binary classification. Given a drug SMILES string, predict the toxicity in a specific assay.

Command: sh commands/run_tox.sh

References: Tox21 Challenge.

MNIST

Dataset Description:: The MNIST database of handwritten digits has a training set of 60,000 examples, and a test set of 10,000 examples with the size of 28x28.

Task Description: Multiclass classification. Given a gray-scale handwritten digit image, predict its label (0-9).

Command: sh commands/run_mnist.sh

References: MNIST Database

CIFAR10

Dataset Description:: The CIFAR-10 dataset consists of 60000 32x32 color images in 10 classes, with 6000 images per class. There are 50,000 training images and 10,000 test images.

Command: sh commands/run_cifar.sh

Task Description: Multiclass classification. Given a 3-channel image, predict its label (airplane, automobile, bird, cat, deer, dog, frog, horse, ship, truck).

References: CIFAR-10 Dataset