Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

GradiVeQ: Vector Quantization for Bandwidth-Efficient Gradient Aggregation in Distributed CNN Training #16

Open
ryoherisson opened this issue Jun 4, 2021 · 0 comments
Assignees

Comments

@ryoherisson
Copy link
Contributor

ryoherisson commented Jun 4, 2021

一言でいうと

Ring All-Reduce(RAR)上で勾配を量子化する手法を提案.

論文リンク

https://proceedings.neurips.cc/paper/2018/file/cf05968255451bdefe3c5bc64d550517-Paper.pdf

著者/所属機関

Mingchao Yu, Zhifeng Lin, Krishna Narra, Songze Li, Youjie Li, Nam Sung Kim, Alexander Schwing, Murali Annavaram, Salman Avestimehr
(University of Southern California, University of Illinois at Urbana Champaign)

投稿日付(yyyy/MM/dd)

2018/12/31

概要

Ring All-Reduce(RAR)のようなリング型の集約プロトコルと併用した場合に勾配を圧縮する手法を提案.
圧縮には,PCAを通じて得たCNNの勾配間の線形相関を用いる.
線形な圧縮手法であるため,各ノードでの解凍・圧縮のオーバーヘッドを取り除くことができる.
目立った精度劣化なしに,勾配集約時間と学習時間を短縮.

RAR
スクリーンショット 2021-06-04 13 06 33

新規性・差分

  • 畳み込み層の勾配の線形相関を踏まえた圧縮手法

手法

スクリーンショット 2021-06-04 13 21 21

スクリーンショット 2021-06-04 13 27 48

結果

スクリーンショット 2021-06-04 13 03 27

コメント

@ryoherisson ryoherisson self-assigned this Jun 4, 2021
@ryoherisson ryoherisson changed the title [WIP]GradiVeQ: Vector Quantization for Bandwidth-Efficient Gradient Aggregation in Distributed CNN Training GradiVeQ: Vector Quantization for Bandwidth-Efficient Gradient Aggregation in Distributed CNN Training Jun 4, 2021
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

No branches or pull requests

1 participant