how to avoid for loop to increase speed

2 次查看（过去 30 天）
Bertil Veenstra
Bertil Veenstra 2023-1-4
评论： Bertil Veenstra 2023-1-6
I would like to calculate the mean value of vector A for every sample as defined in vector B. So, the first point of the resulting vector C is the mean of the first 9 (=B(1)) datapoints of A. The second point of C is the mean of the following 10 (=B(2) datapoints of A. Etc. The following code works, but takes time when processing large vectors:
A=rand(91,1); % vector with 91 random samples
B=[9,10,10,8,11,10,10,9,10,6]; % vector with the number of samples
C= zeros(length(B),1); % preallocate C
first=1;
for i = 1: length(C)
last = first+B(i)-1;
interval=(first:last);
C(i) = mean(A(interval));
first =last+1;
end
Is there a way to use B in an index of A, instead of using this for loop?
  2 个评论
Matt J
Matt J 2023-1-4
Ran in:
Shouldn't sum(B) equal length(A)?
A=rand(91,1); % vector with 91 random samples
B=[9,10,10,8,11,10,10,9,10,6]; % vector with the number of samples
sum(B)
ans = 93
length(A)
ans = 91
Bertil Veenstra
Bertil Veenstra 2023-1-4
Yes, you are right. They should be the same length

请先登录，再进行评论。

请先登录，再回答此问题。

采纳的回答

Matt J
Matt J 2023-1-4
编辑：Matt J 2023-1-5
A=rand(91,1); % vector with 91 random samples
B=[9,10,10,8,11,10,10,9,10,6]; % vector with the number of samples
G=repelem(1:numel(B), B);
n=min(numel(A),numel(G));
G=G(:); A=A(:);
C=accumarray(G(1:n),A(1:n))./accumarray(G(1:n),1); %NOTE: faster than accumarray(G,A,[],@mean)
  4 个评论
显示 2更早的评论
Matt J
Matt J 2023-1-5
编辑：Matt J 2023-1-5
Ran in:
Better to avoid two ACCUMARRAY() calls and specify the function instead:
That would be quite a bit slower, unfortunately. I very deliberately avoided it for that reason:
A=rand(900000,1); % vector with 91 random samples
B=ones(1,90000)*10; % vector with the number of samples
assert(sum(B)==numel(A))
G=repelem(1:numel(B), B);
n=min(numel(A),numel(G));
G=G(:); A=A(:);
tic;
C=accumarray(G(1:n),A(1:n))./accumarray(G(1:n),1);
toc
Elapsed time is 0.022540 seconds.
tic;
C=accumarray(G(1:n),A(1:n),[],@mean); % simpler and faster
toc
Elapsed time is 0.333534 seconds.
Stephen23
Stephen23 2023-1-5
"That would be quite a bit slower, unfortunately. I very deliberately avoided it for that reason:"
Aah, that is a shame. Perhaps a code comment would help to make that choice clear.

请先登录，再进行评论。

更多回答（2 个）

Mathieu NOE
Mathieu NOE 2023-1-4
hello
had to change A to 93 samples so that it matches with sum(B) = 93
see my suggestion below. On this small data we can see an iùprovement of factor 4
Elapsed time is 0.004598 seconds (your code)
Elapsed time is 0.001204 seconds. (my code)
delta = 1.0e-14 *
0
-0.0222
-0.0111
-0.0333
-0.0555
0.1110
0.0777
-0.0222
-0.0444
-0.0222
wonder if that is going to be even better for larger vectors ?
A=rand(93,1); % vector with 93 random samples
B=[9,10,10,8,11,10,10,9,10,6]; % vector with the number of samples
C= zeros(length(B),1); % preallocate C
first=1;
tic
for i = 1: length(C)
last = first+B(i)-1;
interval=(first:last);
C(i) = mean(A(interval));
first =last+1;
end
toc
% alternative code
tic
As = cumsum(A(:));
Bs = cumsum(B(:));
As = As(Bs);
Cs = [As(1); diff(As)]./B(:);
toc
delta = C - Cs
plot(C,'-*b')
hold on
plot(Cs,'dr')
hold off
  7 个评论
显示 5更早的评论
Matt J
Matt J 2023-1-6
编辑：Matt J 2023-1-6
Ran in:
@Stephen23 already showed you earlier that you can use accumarray to apply any function to the blocks.
A = randi(5,1,9300); % vector with random integers ranging from 1 to 5
B=ones(1,930)*10; % vector the number of samples
tic
C= zeros(length(B),1); % preallocate C
first=1;
for i = 1: length(C)
last = first+B(i)-1;
interval=(first:last);
C(i) = mode(A(interval));
first =last+1;
end
toc
Elapsed time is 0.032246 seconds.
tic
G=repelem((1:numel(B)),B);
C=accumarray(G(:),A(:),[],@mode);
toc
Elapsed time is 0.019723 seconds.
It is to be expected that speed-up is more modest, unfortunately. Accumarray isn't as well optimized for arbitrary functions.
Bertil Veenstra
Bertil Veenstra 2023-1-6
Thanks you Matt for this code. It doesns't seem to run faster than the for loop, but it did learn me more about writing and understanding matlab code.

请先登录，再进行评论。

Matt J
Matt J 2023-1-6
编辑：Matt J 2023-1-6
Ran in:
And instead of the mean, I am interested in the mode.
This method should offer speed-up for a generic function, provided that it can ignore NaNs and provided the blocks don't vary too greatly in length:
B = randi([5,10],1,9300);
A = randi(5,1,sum(B));
discrepancy = max( abs(loopMethod(A,B)-altMethod(A,B)),[],'all')
discrepancy = 0
timeit(@() loopMethod(A,B))
ans = 0.1231
timeit(@() altMethod(A,B))
ans = 0.0027
function C=loopMethod(A,B)
C= zeros(1,length(B)); % preallocate C
first=1;
for i = 1: length(C)
last = first+B(i)-1;
interval=(first:last);
C(i) = mode(A(interval));
first =last+1;
end
end
function C=altMethod(A,B)
bmax=max(B);
I=(1:bmax)'<=B;
T=nan(size(I));
T(I)=A(:);
C=mode(T,1);
end
  1 个评论
Bertil Veenstra
Bertil Veenstra 2023-1-6
Hi Matt, this works nice. On larger files I get an 3-4 fold increase in speed. Smart solution. Thanks for your time.

请先登录，再进行评论。

类别

在 Help Center 和 File Exchange 中查找有关 Matrix Indexing 的更多信息

标签

产品


版本

R2017b

Community Treasure Hunt

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

Start Hunting!

Translated by