Skip to content

lx10077/fedavgpy

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

6 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

On the Convergence of FedAvg on Non-IID Data

This repository contains the codes for the paper

On the Convergence of FedAvg on Non-IID Data

Our paper is a tentative theoretical understanding towards FedAvg and how different sampling and averaging schemes affect its convergence.

Our code is based on the codes for FedProx, another federated algorithm used in heterogeneous networks.

Usage

  1. First generate data by the following code. Here generate_random_niid is used to generate the dataset named as mnist unbalanced in our paper, where the number of samples among devices follows a power law. generate_equal is used to generate the dataset named as mnist balanced where we force all devices to have the same amount of samples. More non-iid distributed datasets could be found in FedProx.

    cd fedpy
    python data/mnist/generate_random_niid.py
    python data/mnist/generate_equal.py
    python data/synthetic/generate_synthetic.py
    
  2. Then start to train. You can run a single algorithm on a specific configuration like

    python main.py --gpu --dataset $DATASET --clients_per_round $K --num_round $T --num_epoch $E --batch_size $B --lr $LR --device $device --seed $SEED --model $NET --algo $ALGO  --noaverage --noprint
    

Notes

  • There are three choices for $ALGO, namely fedavg4 (containning the Scheme I and II), fedavg5 (for the original scheme) and fedavg9 (for the Transformed Scheme II).

  • If you don't want to use the Scheme I (where we sample device acccording to $p_k$ and simply average local parameters), please add --noaverage.

  • If you want to mute the printed information, please use --noprint.

  1. Once the trainning is started, logs that containning trainning statistics will be automatically created in result/$DATASET. Each run has a unique log file name in this way year-month-day-time_$ALGO_$NET_wn10_tn100_sd$SEED_lr$LR_ep$E_bs$B_a/w, for example,

    2019-11-24T12-05-13_fedavg4_logistic_wn10_tn100_sd0_lr0.1_ep5_bs64_a
    
  2. During the trainning, you visualize the process by running either of the following

 tensorboard --logdir=result/$DATASET
 tensorboard --logdir=result/$DATASET/$LOG
 # For example
 tensorboard --logdir=result/mnist_all_data_0_equal_niid/
 tensorboard --logdir=result/mnist_all_data_0_equal_niid/2019-11-24T12-05-13_fedavg4_logistic_wn10_tn100_sd0_lr0.1_ep5_bs64_a
  1. All the codes we used to draw figures are in plot/. You can find some choices of hyperparameters in both our paper and the scripts in plot/.

Dependency

Pytorch = 1.0.0

numpy = 1.16.3

matplotlib = 3.0.0

tensorboardX

About

On the Convergence of FedAvg on Non-IID Data

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages