repeat large number of vector elements

2 views (last 30 days)
Yufan Miao
Yufan Miao on 20 Feb 2013
Edited: Jan on 27 Dec 2015
Hi, I have a vector storing only unique values: v = (0, 1, 2), and another vector about frequency of the unique values: c = (3, 2, 4). Now I want to create a new vector with repeated values as follows:
v2 = (0, 0, 0, 1, 1, 2, 2, 2, 2)
Since both vectors v and c are very large, how to efficiently derive v2 in matlab. Many thanks.
  1 Comment
Jan
Jan on 20 Feb 2013
Please specify "very large" explicitly: Some users tell 1000 elements large, others need billions for this term.

Sign in to comment.

Answers (5)

Jan
Jan on 20 Feb 2013
Edited: Jan on 20 Feb 2013
This is a run-length encoding. A standard decoding:
v = [0, 1, 2];
c = [3, 2, 4];
d = cumsum(c);
index = zeros(1, d(end));
index(d(1:end-1)+1) = 1;
index(1) = 1;
index = cumsum(index);
result = v(index);

Azzi Abdelmalek
Azzi Abdelmalek on 20 Feb 2013
Edited: Azzi Abdelmalek on 21 Feb 2013
v=[0 1 2];
c=[3,2,4];
out=cell2mat(arrayfun(@(x) repmat(v(x),1,c(x)),1:numel(v),'un',0))
EDIT
i1=0;
out=zeros(1,sum(c));
for k=1:numel(v)
i0=i1+1;
i1=i0+c(k)-1;
out(i0:i1)=v(k)*ones(1,c(k));
end
  5 Comments
Azzi Abdelmalek
Azzi Abdelmalek on 20 Feb 2013
It seems that ones(1,n)*v is faster then repmat
Youssef  Khmou
Youssef Khmou on 20 Feb 2013
right, 0.0054s with repmat, and 0.0027s with ones(1,n)

Sign in to comment.


Youssef  Khmou
Youssef Khmou on 20 Feb 2013
Edited: Youssef Khmou on 20 Feb 2013
for i=1:length(c)
V{i}=(i-1)*ones(c(i),1);
end
f=cell2mat(V(:))'

Jan
Jan on 21 Feb 2013
Edited: Jan on 21 Feb 2013
Some timings (R2009a/64/Win7/Core2Duo):
x = rand(1, 10000);
n = randi([0,10], 1, 10000);
tic; for k=1:100; r=runlength(x,n); end; toc
% With runlength() is a function with the following contents:
% Azzi's ARRAYFUN approach:
r = cell2mat(arrayfun(@(k) repmat(x(k), 1, n(k)), 1:numel(x), 'un', 0))
Elapsed time is 48.976947 seconds.
% Azzi's approach from the comment section: [EDITED start]
Elapsed time is 2.674559 seconds.
% Azzi's approach from the comment section with slight modification:
i1 = 0;
result = zeros(1, sum(c));
for k = 1:numel(v)
i0 = i1+1;
i1 = i0 + c(k) - 1;
result(i0:i1) = v(k); % Without "*ones(1,c(k))" !!
end
Elapsed time is 1.656987 seconds. [EDITED end]
% Youssef' approach:
V = cell(1, length(n));
for k = 1:length(n)
V{k} = x(k) * ones(n(k), 1); % With x(k) instead of (i-1)
end
r = cell2mat(V(:))'
Elapsed time is 1.422626 seconds
% The CUMSUM method from my former answer:
Elapsed time is 0.083854 seconds.
% An equivalent MEX implementation:
Elapsed time is 0.025775 seconds.
Conclusions:
  • ARRAYFUN and the anonymous function is not efficient compared to a simple loop. Therefore I suggest to avoid this combination strictly.
  • Collecting the single parts in a cell is a good idea. Then CELL2MAT can be replaced by the faster FEX: Cell2Vec : 0.97 sec
  • Creating the large temporary vectors cumsum(x) and index seem to be not as bad as I was afraid. The C-MEX is not a dramatic enhancement anymore.
  2 Comments
Azzi Abdelmalek
Azzi Abdelmalek on 21 Feb 2013
You missed the code hiden in the comment.
Jan
Jan on 21 Feb 2013
Edited: Jan on 21 Feb 2013
Yes, Azzi, I haven't seen it. It is much faster, especially when the multiplication with ONES is omitted. A good improvement compared to ARRAYFUN!
I do not understand why this is slower than collecting the partial vectors in a cell array at first. It looks much more efficient.

Sign in to comment.


irina mihai
irina mihai on 6 Dec 2015
Jan, Azzi Thanks for your help ! I used the cumsum function and the v(index) idea which i didn't know. I have a very large database (over 10^6 lines) and I was looking for something like this. Brilliant !
  1 Comment
Jan
Jan on 6 Dec 2015
Edited: Jan on 27 Dec 2015
See repelem in modern Matlab versions and FEX: RunLength, which is about twice as fast as repelem.

Sign in to comment.

Tags

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!