8
4

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 5 years have passed since last update.

RAdvent Calendar 2018

Day 4

Rでtensor train decomposition してみた

Last updated at Posted at 2018-12-05

#はじめに
Tensor train decomposition というテンソル分解の手法があります。あまりメジャーじゃないのかRで実装されている例をみないので作ってみました。

#テンソルトレイン分解

TT.R
TT <- function(x,r)
{
D <- dim(x)
n<- length(D)
LIST <- rep(list(NA),n)
for (i in c(1:(n-1)))
{
if (i==1) {l0 <- 1} else {l0<- r[i-1]}
l1 <- r[i]
dim(x) <- c(D[i]*l0,prod(D[(i+1):n]))
SVD <- svd(x)
LIST[[i]] <- SVD$u[,1:l1]
dim(LIST[[i]]) <- c(l0,D[i],l1)
x<- t(SVD$u[,1:l1]) %*% x
}
LIST[[n]] <- t(x)
dim(LIST[[1]]) <- c(D[1],r[1])
return(LIST)
}

#実行例
使い方は簡単です。テンソル$\mathcal{Z} \in \mathbb{R}^{N_1 \times N_2 \times \cdots \times N_m}$であるとします。
$r= (R_1, R_2, \cdots , R_{m-1})$を用意します。$R_i$はテンソル分解の次元をあたえる整数列です。
以下は$\mathcal{Z} \in \mathbb{R}^{3 \times 4 \times 5}, R_1=2,R_2=3$の場合の実行例です。

> set.seed(0)
> Z <- array(rnorm(3*4*5),c(3,4,5))
> dim(Z)
[1] 3 4 5
> r <- c(2,3)
> LIST <- TT(Z,r)
> LIST
[[1]]
           [,1]       [,2]
[1,] 0.33405782  0.4717883
[2,] 0.94144541 -0.1250656
[3,] 0.04567177 -0.8727969

[[2]]
, , 1

            [,1]        [,2]        [,3]       [,4]
[1,] -0.02676980  0.68746071  0.02735584 -0.6839789
[2,]  0.08867982 -0.08374744 -0.04396587  0.2032114

, , 2

            [,1]     [,2]       [,3]      [,4]
[1,] 0.005461026 0.161148  0.1228488 0.3503506
[2,] 0.094666990 0.485986 -0.2761150 0.7174820

, , 3

           [,1]       [,2]       [,3]       [,4]
[1,] 0.06102372 -0.3423556  0.6592200 -0.1576585
[2,] 0.41660602 -0.3899808 -0.2696078  0.1459710


[[3]]
           [,1]       [,2]       [,3]
[1,] -0.3594280  2.7946508 -1.4622505
[2,] -0.3689526 -0.3154154 -1.3315321
[3,] -1.7488794 -1.4190110 -0.6076388
[4,]  1.9922802 -2.1448302 -1.2577524
[5,]  3.2735897  0.8185284  0.1302124

> dim(LIST[[1]])
[1] 3 2
> dim(LIST[[2]])
[1] 2 4 3
> dim(LIST[[3]])
[1] 5 3


これだけです。テンソルトレイン分解は
$$
z_{i_1i_2,\cdots,i_m} \sim \sum_{r_1=1}^{R_1} \sum_{r_2=1}^{R_2} \cdots \sum_{r_{m-1}=1}^{R_{m-1}} G(i_1,r_1) G(r_1,i_2,r_2) \cdots G(r_{m-2},i_{m-1},r_{m-1}) G(i_m,r_{m-1})
$$
という分解をします。$G(i_1,r_1) \in \mathbb{R}^{N_1 \times R_1},G(i_m,r_{m-1}) \in \mathbb{R}^{N_m \times R_{m-1}}, G(r_{k-1}, i_k, r_k) \in \mathbb{R}^{R_{k-1} \times N_k \times R_k}$です。
この実行例ではLIST[[1]]に$G(i_1,r_1) \in \mathbb{R}^{N_1 \times R_1}$が、LIST[[2]]に$G(r_1,i_2,r_2) \in \mathbb{R}^{R_1 \times N_2 \times R_2}$が、LIST[[3]]に$G(i_3,r_2) \in \mathbb{R}^{N_3 \times R_2}$が、それぞれ入っています。

#参考文献
SIAM J. Sci. Comput., 33(5), 2295–2317. (23 pages)
Tensor-Train Decomposition
I. V. Oseledets
https://doi.org/10.1137/090752286

8
4
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
8
4

Delete article

Deleted articles cannot be recovered.

Draft of this article would be also deleted.

Are you sure you want to delete this article?