Skip to content

bibikar/feddst

Repository files navigation

Federated Dynamic Sparse Training (FedDST)

This experiment code accompanies the paper

Sameer Bibikar, Haris Vikalo, Zhangyang Wang, and Xiaohan Chen, "Federated Dynamic Sparse Training: Computing Less, Communicating Less, Yet Learning Better," 2021.

Dependencies

  • Python 3.6 or greater
  • PyTorch, torchvision
  • tqdm

Run git submodule init followed by git submodule update to download the dataset code we use.

Examples

Experiment Command line
FedAvg on CIFAR-10 python3 dst.py --dataset cifar10 --sparsity 0.0
FedProx on CIFAR-10 (mu = 1) python3 dst.py --dataset cifar10 --sparsity 0.0 --prox 1
FedDST on CIFAR-10 (S=0.8, alpha=0.01, R_adj=15) python3 dst.py --dataset cifar10 --sparsity 0.8 --readjustment-ratio 0.01 --rounds-between-readjustments 15
FedDST on CIFAR-100 (S=0.5, alpha=0.01, R_adj=10) python3 dst.py --dataset cifar100 --sparsity 0.5 --readjustment-ratio 0.01 --distribution dirichlet --beta 0.1
FedDST+FedProx on CIFAR-10 (S=0.8, alpha=0.01, R_adj=15, mu=1) python3 dst.py --dataset cifar10 --sparsity 0.8 --readjustment-ratio 0.01 --rounds-between-readjustments 15 --prox 1
RandomMask on MNIST (S=0.8) python3 dst.py --dataset mnist --sparsity 0.8 --readjustment-ratio 0.0
PruneFL on MNIST python3 prunefl.py --dataset mnist --rounds-between-readjustments 50 --initial-rounds 1000

About

Federated Dynamic Sparse Training

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages