3
0

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?

More than 3 years have passed since last update.

torch.bmmのスピード改善について

Posted at

何をしたいか

結論

  • torch.bmmが結構早くなった
  • 特に、GPUで内積を計算するだけなら要素積+和で計算するより早い

CPU

計算方法 t(内積) 別記事
要素積+和 43.4 µs 26.4 µs
torch.bmm 199 µs 964 µs

GPU

計算方法 t(内積) t(内積+微分) t(微分)
要素積+和 18.7 µs 262 µs 243.3 µs
torch.bmm 17.9 µs 336 µs 318.1 µs

GPU(別記事)

計算方法 t(内積) t(内積+微分) t(微分)
要素積+和 25.9 µs 163 µs 137.1 µs
torch.bmm 608 µs 1.13 ms 552 µs

環境

  • python: 3.7.2
  • pytorch: 1.6.0

実行したスクリプト

  • 以下はjupyter notebookにて実行
準備
import torch
a = torch.randn(500, 500, dtype=torch.float, device='cpu')
b = torch.randn(500, 500, dtype=torch.float, device='cpu')

c = torch.randn(500, 500, dtype=torch.float, device='cuda')
d = torch.randn(500, 500, dtype=torch.float, device='cuda')

e = torch.randn(500, 500, dtype=torch.float, device='cuda', requires_grad=True)
f = torch.randn(500, 500, dtype=torch.float, device='cuda', requires_grad=True)
同じ演算
# 下の2つの演算は同じです。
(a*b).sum(1, keepdim=True)
torch.bmm(a.unsqueeze(1), b.unsqueeze(2)).squeeze(2))
cpuで演算(要素積+和)
%timeit (a*b).sum(1, keepdim=True)
43.4 µs ± 3.35 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
cpuで演算(bmm)
%timeit torch.bmm(a.unsqueeze(1), b.unsqueeze(2)).squeeze(2)
199 µs ± 2.61 µs per loop (mean ± std. dev. of 7 runs, 10000 loops each)
gpuで演算(要素積+和)
%timeit (c*d).sum(1, keepdim=True)
18.7 µs ± 39.3 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
gpuで演算(bmm)
%timeit torch.bmm(c.unsqueeze(1), d.unsqueeze(2)).squeeze(2)
17.9 µs ± 33.9 ns per loop (mean ± std. dev. of 7 runs, 100000 loops each)
gpuで演算+微分(要素積+和)
%timeit (e*f).sum(1, keepdim=True).sum().backward()
262 µs ± 840 ns per loop (mean ± std. dev. of 7 runs, 1000 loops each)
gpuで演算+微分(bmm)
%timeit torch.bmm(e.unsqueeze(1), f.unsqueeze(2)).squeeze(2).sum().backward()
336 µs ± 270 ns per loop (mean ± std. dev. of 7 runs, 1000 loops each)

コメント

  • ここからソースコードを追いかけたかったが、masterと大幅に違って追いかけるのが大変だったので諦めた

その他関連ページ

3
0
0

Register as a new user and use Qiita more conveniently

  1. You get articles that match your needs
  2. You can efficiently read back useful information
  3. You can use dark theme
What you can do with signing up
3
0

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?