zxc-phy / datafl

Online Federated Learning on Distributed Unknown Data Using UAVs

This repo is the pytorch implementation of paper "Online Federated Learning on Distributed Unknown Data Using UAVs".

Prerequisites

nvidia-smi	# NVIDIA GeForce RTX 4090
nvcc -V 	# cuda-12.1
python -V	# Python 3.10.10
pip list	# torch 2.0.0+cu118, torchvision 0.15.1+cu118, GPy 1.10.0

The estimation of the reward function requires the GPy module. (https://sheffieldml.github.io/GPy/)

Preparing dataset and model

You can specify the datasets, models and other hyperparameters in ./sim/utils/options.py. For any new dataset, you first need to download the dataset and run ./sim/data/partition.py to construct a distributed dataset. To use a specific model on a new dataset, you can add settings in the ./sim/model folder and modify ./sim/model/build_models.py to add the model. Currently used models include ResNet-18, CNNs, AlexNet and so on.

OFL-UD^2

To execute the command, just run python ./OFL.py, and you will get the change of test accuracy during training. If you need to visualize the output, you can find the corresponding csv file in the ./save folder and plot it. All baselines are stored in the ./baselines folder. The parameters in the experiment are placed in options.py, and you can use the default parameters or modify them at runtime by specifying methods such as -d cifar10.