0

PowerSGD: Practical Low-Rank Gradient Compression for Distributed Optimization

A new power iteration based low-rank gradient compressor significantly reduces training times for convolutional networks and LSTMs while achieving test performance on par with standard SGD.

Year
2019
Venue
powersgd-practical-low-rank-gradient-1
Authors
3
Hosting
Abstract onlyARXIV-DEFAULT

Cite

Notes

Only stored in your browser.

Attribution

Abstract & full text
arxiv.org/abs/1905.13727v3ARXIV-DEFAULT
TL;DR
Semantic Scholar
Attribution policy →

Abstract

We study gradient compression methods to alleviate the communication bottleneck in data-parallel distributed optimization. Despite the significant attention received, current compression schemes either do not scale well or fail to achieve the target test accuracy. We propose a new low-rank gradient compressor based on power iteration that can i) compress gradients rapidly, ii) efficiently aggregate the compressed gradients using all-reduce, and iii) achieve test performance on par with SGD. The proposed algorithm is the only method evaluated that achieves consistent wall-clock speedups when benchmarked against regular SGD with an optimized communication backend. We demonstrate reduced training times for convolutional networks as well as LSTMs on common datasets. Our code is available at https://github.com/epfml/powersgd.

Authors

3