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

Quantized compressive sampling of stochastic gradients for efficient communication in distributed deep learning #15

Open
mitsuhiko-nozawa opened this issue Jun 3, 2021 · 0 comments

Comments

@mitsuhiko-nozawa
Copy link
Contributor

mitsuhiko-nozawa commented Jun 3, 2021

一言でいうと

既存の圧縮手法における、勾配の総分散の増加、圧縮率の上限、暗黙的なバイアスの追加といった問題を解決する、Quantized Compression Sampling を提案する

論文リンク

https://ojs.aaai.org/index.php/AAAI/article/view/5706

著者/所属機関

Afshin Abdi, Faramarz Fekri
School of Electrical and Computer Engineering
Georgia Institute of Technology, Atlanta, GA

投稿日付(yyyy/MM/dd)

AAAI
2020/04/03

概要

量子化を行うと、いくつかの問題がある

  • 量子化によって生じるノイズにより、勾配の総分散が増加し、通常収束するはいパラで収束しなくなる、また学習係数を落とす必要があり、イテレーション数が増える
  • バイアスを受ける量子化手法では収束が保証されない
  • 小さい勾配が0に丸められ、変化しなくなる
    (a) ある学習係数での最終的な収束率
    (b), (c) ある学習係数でのイテレーションごとの収束率
    image
    これらの問題を解決する、エラーフィードバックというものがある。
    これは量子化前後でのエラーを各ノードで保存しておく。

新規性・差分

量子化よって生じる問題点を指摘し(概要)、それを解決する手法を提案

手法

n次元の勾配をk次元にマッピングし、それを再構築する行列T(n, k)を考える。元の勾配をg((n, 1), マップ後の勾配をv(k, 1)とする。すなわちv = Tg
Tはアダマール行列の上からk行のHと、対角成分に-1か1をランダムにもつ対角行列Rの積
image
その後、gのQuantized Compressive Samplingは以下で表される、sは量子化スケール係数、eは(-1/2, 1/2)の一様分布.
image
このv^から、E[g^]=gとはるような変換を求めたい。
image
かくかくしかじかで、分散の大きさとマップ後の次元k(通信量)のトレードオフになる。
image
image
アルゴリズム、勾配をk次元に変換した後量子化を行い、通信、その後変換行列で再構築。
同じシードを使うことで完全に復元できる前提
ランダム要素を加えることで、R^nの固定されたk次元部分空間に射影されることを防ぐ?
image

結果

学習の収束は問題がないようだが、肝心の通信時間に関する実験結果がなかった
FC, Lenet, CifarNet, Alexnetで実験

提案手法は圧縮もしているけど速い
image

ノード数を増やした時の精度
image

収束率
image

コメント

@mitsuhiko-nozawa mitsuhiko-nozawa self-assigned this Jun 4, 2021
@mitsuhiko-nozawa mitsuhiko-nozawa changed the title [WIP] Quantized compressive sampling of stochastic gradients for efficient communication in distributed deep learning Quantized compressive sampling of stochastic gradients for efficient communication in distributed deep learning 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