Exponential Graph is Provably Efficient for Decentralized Deep Training

This code repository is for the paper Exponential Graph is Provably Efficient for Decentralized Deep Training to be appeared in NeurIPS 2021. If you feel this work helps in your research, please consider cite

Bicheng Ying, Kun Yuan, Yiming Chen, Hanbin Hu, Pan Pan, and Wotao Yin, Exponential Graph is Provably Efficient for Decentralized Deep Training. Advances in Neural Information Processing Systems (NeurIPS), 2021.

@inproceedings{Ying_2021_ExpoGraph,
    title = {Exponential Graph is Provably Efficient for Decentralized Deep Training},
    author = {Bicheng Ying and Kun Yuan and Yiming Chen and Hanbin Hu and Pan Pan and Wotao Yin},
    booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
    year = 2021
}

Pre-requisites

This code was trained and tested with

1. Python 3.7.5
2. PyTorch 1.4.0
3. Torchvision 0.5.0
4. tqdm 4.62.3
5. tensorboardX 2.4
6. bluefog 0.3.0

You can also use the Bluefog docker image for testing.

Static and One-peer exponential graphs comparison on CIFAR-10

Here gives a simple code snippet for running the CIFAR-10 experiments on a decentralized network of size 4 using the one-peer exponential graph. You can find the script for testing in run.sh as well.

$ bfrun -np 4 python train_cifar10.py --epochs 200 --model resnet56 --batch-size 128 --base-lr 0.01

Expected performance on CIFAR-10 experiments

Method	Accuracy
Static exponential graph	91.21%
One-peer exponential graph	91.84%