Making for loop and if statment and increasing speed using matlab / octave

Question

I'm trying to see if I can improve the speed of a for loop and an if condition statement. Basically it does a lookup on non repeating key values into an array and gets the value from another column.

If I run 100000 values it takes about 13 seconds see code below. Is there a way to make this more efficient? Ps i'm using octave 3.8.1 which works with matlab

%test if lookup statment
clear all, clc,  tic, clf; 

num_to_test=100000 %amount of numbers to test
a1=(1:1:num_to_test)';
a2=(a1.*num_to_test);
array=[a1,a2]; %array where values are stored

lookupval=(randperm(num_to_test,num_to_test/2)/4)'; %lookup these random values of non repeating integers and floats and get another value

amp=[];
freq=[];
found_array=[];
notfound_array=[];

for ii=1:1:rows(lookupval)
    if (find(lookupval(ii)==array(:,1)))  %if you find a lookup value in array
        %disp('found');
        [row,col] = find(lookupval(ii) == array(:,1));
        amp=[amp;array(row,2)];
        freq=[freq;array(row,1)];
        found_array=[freq,amp];

    else %add lookup value to another array and make amp value zero

        notfound_arraytmp=[lookupval(ii),0];
        notfound_array=[notfound_array;notfound_arraytmp];
    endif
end
comb_array=[found_array;notfound_array];
sort_comb_array=sortrows(comb_array,1); %sort array by first col incrementing

fprintf('\nfinally Done-elapsed time -%4.4fsec- or -%4.4fmins- or -%4.4fhours-\n',toc,toc/60,toc/3600);

Divakar · Accepted Answer · 2015-02-02 17:02:48Z

2

Approach #1

This could be really efficient with ismember -

lookupval = sort(lookupval);                     %// Do sorting at the start
sort_comb_array = [lookupval zeros(size(lookupval))]; %// Setup output array
[idA,idB] = ismember(array(:,1),lookupval);           %// Get matching IDs
sort_comb_array(idB(idA),2) = array(idA,2);  %// Index into second column
                                   %// of array and get corresponding values

Approach #2

I would thrown in my favorite bsxfun too, but for such huge datasizes of 100,000, its memory inefficiency could make it slower -

lookupval = sort(lookupval);
sort_comb_array = [lookupval zeros(size(lookupval))];
[idA,idB] = find(bsxfun(@eq,array(:,1),lookupval(:).')); %//'# Get matching IDs
sort_comb_array(idB,2) = array(idA,2);

edited Feb 2, 2015 at 17:02

answered Feb 2, 2015 at 16:49

Divakar

222k19 gold badges273 silver badges374 bronze badges

Sign up to request clarification or add additional context in comments.

Comments

nkjt · Accepted Answer · 2015-02-02 16:36:14Z

2

Several issues but the main one is probably that you don't preallocate - appending like this: amp=[amp;array(row,2)]; is generally slow in MATLAB. You don't need a loop here, though.

Let's start with a simple array, A:

And our lookup values are [2 6 3 9 7]; We want our output to show these lookup values, sorted, in the first column, and the second column to be either the values from the second column of A (where they exist) or zero.

lookup = sort(lookup);
output = zeros(length(lookup),2);
output(:,1) = lookup;
[c a b ] = intersect(A(:,1),lookup);
output(b,2) = A(a,2);

The output is:

edited Feb 2, 2015 at 16:36

answered Feb 2, 2015 at 14:45

nkjt

7,8179 gold badges24 silver badges28 bronze badges

8 Comments

Rick T Over a year ago

Thanks for the help. I ran your code online and the output doesn't match up (it makes all the lookup values 0) did I miss something? here's a link to what I ran ideone.com/bxMhig

Divakar Over a year ago

@RickT I'd be interested to see how the other ismember based solution performs on Octave for this problem!

Rick T Over a year ago

@Divakar Sure testing and comparing them now

Rick T Over a year ago

@Divakar I posted the results as an answer

Divakar Over a year ago

@RickT Make sure to clear out memory before the start of a new approach and sufficient number of iterations to make the timings around or more than 1sec.

|

Community · Accepted Answer · 2017-05-23 12:06:31Z

Purely from an efficiency standpoint, I would rewrite the for loop as follows:

m = 0;                          % number of omitted values
n = 0;                          % number of found values
for ii=1:1:rows(lookupval)

    [row,col] = find(lookupval(ii) == array(:,1));

    if ~isempty(row)  %if you find a lookup value in array
        %disp('found');
        n=n+1;
        amp(n)=array(row,2);
        freq(n)=;array(row,1);
        found_array=[freq,amp];

    else %add lookup value to another array and make amp value zero
        m=m+1;
        notfound_array(2*m-1:2*m)=[lookupval(ii);0];
    endif
end

This saves you a find call by using its output directly rather than recomputing it when find returns a position, and grows the arrays in a more efficient way (as shown in this question).

Rick T · Accepted Answer · 2015-02-02 18:43:14Z

This is a test Divakar suggest I do to see the speed it takes octave 3.8.1 to run this. Results are below along with the code.

1) Using ismember with 2,000,000 is faster but uses more memory
-elapsed time -0.2306sec- or -0.0038mins-
Total is 15000001 elements using 106000008 bytes

2) Using intersect with 2,000,000 is slower but uses less memory.
-elapsed time -0.3057sec- or -0.0051mins-
Total is 11749047 elements using 93992376 bytes

3) Using bskfun with 100,000 produces an error: out of memory or dimensions too large for octave's index type

First test results:

clear all, clc,  tic, clf; 
num_to_test=2000000 %amount of numbers to test
a1=(1:1:num_to_test)';
a2=(a1.*num_to_test);
array=[a1,a2]; %array where values are stored

lookupval=(randperm(num_to_test,num_to_test/2)/4)'; %lookup these random vaules of intergers and floats and get another value

lookupval = sort(lookupval);
sort_comb_array = [lookupval zeros(size(lookupval))];
[idA1,idB1] = ismember(array(:,1),lookupval);
sort_comb_array(idB1(idA1),2) = array(idA1,2);

fprintf('\nfinally Done-elapsed time -%4.4fsec- or -%4.4fmins- or -%4.4fhours-\n',toc,toc/60,toc/3600);

whos

>>>num_to_test =  2000000
>>>


   finally Done-elapsed time -0.2306sec- or -0.0038mins- or -0.0001hours-
>>>Variables in the current scope:

   Attr Name                  Size                     Bytes  Class
   ==== ====                  ====                     =====  ===== 
        a1              2000000x1                   16000000  double
        a2              2000000x1                   16000000  double
        array           2000000x2                   32000000  double
        idA1            2000000x1                    2000000  logical
        idB1            2000000x1                   16000000  double
        lookupval       1000000x1                    8000000  double
        num_to_test           1x1                          8  double
        sort_comb_array 1000000x2                   16000000  double

Total is 15000001 elements using 106000008 bytes
========================================================================

Second test results:

clear all, clc,  tic, clf; 

num_to_test=2000000 %amount of numbers to test
a1=(1:1:num_to_test)';
a2=(a1.*num_to_test);
array=[a1,a2]; %array where values are stored

lookupval=(randperm(num_to_test,num_to_test/2)/4)'; %lookup these random vaules of intergers and floats and get another value

lookupval = sort(lookupval);
output = zeros(length(lookupval),2);
output(:,1) = lookupval;
[c a b ] = intersect(array(:,1),lookupval);
output(b,2) =array(a,2);

fprintf('\nfinally Done-elapsed time -%4.4fsec- or -%4.4fmins- or -%4.4fhours-\n',toc,toc/60,toc/3600);

whos

>>>num_to_test =  2000000
>>>
finally Done-elapsed time -0.3057sec- or -0.0051mins- or -0.0001hours-
>>>Variables in the current scope:

   Attr Name              Size                     Bytes  Class
   ==== ====              ====                     =====  ===== 
        a            250005x1                    2000040  double
        a1          2000000x1                   16000000  double
        a2          2000000x1                   16000000  double
        array       2000000x2                   32000000  double
        b            250005x1                    2000040  double
        c            250005x1                    2000040  double
        lookupval   1000000x1                    8000000  double
        num_to_test       1x1                          8  double
        output      1000000x2                   16000000  double

Total is 11750016 elements using 94000128 bytes


=======================================================================

Collectives™ on Stack Overflow

Making for loop and if statment and increasing speed using matlab / octave

4 Answers 4

Comments

8 Comments

Comments

Comments

Your Answer

Linked

Hot Network Questions

Collectives™ on Stack Overflow

4 Answers 4

Comments

8 Comments

Comments

Comments

Your Answer

Sign up or log in

Post as a guest

Linked

Related