ProgFed: Effective, Communication, and Computation Efficient Federated Learning by Progressive Training

[Paper (ICML 2022)] [Project Page]

Citation

@inproceedings{wang2022progfed,
  title={ProgFed: Effective, Communication, and Computation Efficient Federated Learning by Progressive Training},
  author={Wang, Hui-Po and Stich, Sebastian and He, Yang and Fritz, Mario},
  booktitle={International Conference on Machine Learning},
  pages={23034--23054},
  year={2022},
  organization={PMLR}
}

Quick Start

To quickly reproduce the main result of our work, prepare the datasets and follow the instructions below.

Modify the --data_path argument in main.py accordingly (or specify it in the scripts)
Run scripts

cd script-fed
# run end-to-end baselines
bash exec_baseline.sh
# run ProgFed
bash exec_prog.sh

Installation

Pytorch >= 1.8.0
torchvision >= 0.9.1
idx2numpy >= 1.2.3

Datasets

CIFAR-100

Reproduce Results

In script-fed/, we provide exec_baseline.sh and exec_fed.sh for reproducing the experiments with an end-to-end baseline model and with our progressive training algorithm. We also offer detailed usage as follows.

Basci Usage

python main.py -cfg <path-to-configuration-file>

settings-fed contains various configurations. They are categorized by architectures, e.g., convnet, resnet18, and vgg16.

ProgFed vs. End-to-End Baseline

# ProgFed
python main.py -cfg settings-fed/resnet18/cifar100-prog-warmup25-375u.yaml
# end-to-end baseline
python main.py -cfg settings-fed/resnet18/cifar100-baseline.yaml

Compression

We show that ProgFed is compatible with sparsification and quantization. The configuration files specify the number of bits for linear quantization and sparsification ratio, e.g., cifar100-baseline-l8+s25.yaml stands for 8-bit linear quantization along with 25% sparsification.

ProgFed

├── settings-fed/resnet18
    ├── cifar100-prog-warmup25-375u-l8.yaml
    ├── cifar100-prog-warmup25-375u-l4.yaml
    ├── cifar100-prog-warmup25-375u-l2.yaml
    ├── cifar100-prog-warmup25-375u-s25.yaml
    ├── cifar100-prog-warmup25-375u-s10.yaml
    ├── cifar100-prog-warmup25-375u-l8+s25.yaml
    └── cifar100-prog-warmup25-375u-l8+s10.yaml

Baseline

├── settings-fed/resnet18
    ├── cifar100-baseline-s25.yaml
    ├── cifar100-baseline-s10.yaml
    ├── cifar100-baseline-l2.yaml
    ├── cifar100-baseline-l4.yaml
    ├── cifar100-baseline-l8.yaml
    ├── cifar100-baseline-l8+s10.yaml
    └── cifar100-baseline-l8+s25.yaml

Optimization

ProgFed is compatible with several advanced federated optimizers, e.g, FedProx and FedAdam.

# FedProx
python main.py -cfg settings-fed/resnet18/fedprox-cifar100-prog-warmup25-375u.yaml
# FedAdam
python main.py -cfg settings-fed/resnet18/fedadam-cifar100-prog-warmup25-375u.yaml

Ablation Study

We provide an ablation study on two critical parameters of ProgFed, namely the number of stages (S) and the number iterations (t). Note that the last setting always refers to the one used in the main paper, i.e., S=4.

Number of stages (S)

├── settings-fed/resnet18
    ├── cifar100-prog-warmup25-S2.yaml
    ├── cifar100-prog-warmup25-S2.yaml
    ├── cifar100-prog-warmup25-S2.yaml
    ├── cifar100-prog-warmup25-S2.yaml
    └── cifar100-prog-warmup25-375u.yaml

More settings

├── settings-fed/resnet18
    ├── cifar100-prog-warmup25-S2-2000t.yaml
    ├── cifar100-prog-warmup25-S2-2500t.yaml
    ├── cifar100-prog-warmup25-S3-4000t.yaml
    ├── cifar100-prog-warmup25-S5-4000t.yaml
    ├── cifar100-prog-warmup25-S8-4000t.yaml
    ├── cifar100-prog-warmup100-S2.yaml
    ├── cifar100-prog-warmup200-S2.yaml
    ├── cifar100-prog-warmup25-375u-4000t.yaml
    └── cifar100-prog-warmup25-375u.yaml

sisaman/ProgDP