This repository contains the codes of the MG-WFBP (Merged-Gradient Wait-Free BackPropagation) paper submitted to IEEE TPDS. This version works well with PyTorch, and a preliminary version was presented at the conference IEEE INFOCOM 2019, which was originally implemented on B-Caffe: https://github.com/hclhkbu/B-Caffe. As PyTorch becomes much popular than Caffe, you are recommended to use this repository with the MG-WFBP algorithm.

• Python 2 or 3
• PyTorch-0.4.+
• OpenMPI-4.0.+
• Horovod-0.14.+

git clone https://github.com/HKBU-HPML/MG-WFBP.git
cd MG-WFBP
pip install -r requirements.txt
dnn=resnet20 nworkers=4 ./dist_mpi.sh

Assume that you have 4 GPUs on a single node and everything works well, you will see that there are 4 workers running at a single node training the ResNet-20 model with the Cifar-10 data set using the MG-WFBP algorithm.

• S. Shi, X.-W. Chu, and B. Li, “MG-WFBP: Merging Gradients Wisely for Efficient Communication in Distributed Deep Learning,” Under review (Extension of the following conference version).
• S. Shi, X.-W. Chu, and B. Li, “MG-WFBP: Efficient Data Communication for Distributed Synchronous SGD Algorithms,” IEEE INFOCOM 2019, Paris, France, May 2019. PDF

• Deep speech: https://github.com/SeanNaren/deepspeech.pytorch
• PyTorch examples: https://github.com/pytorch/examples