find() along a given dimension of a matrix

I have an n-by-m matrix, A. I would like to find the last element in each row of A satisfying some condition. Intuitively, I'm looking for a way to use the find() command along a given dimension of A like the following syntax:

k = find(X,n,direction,dimension)

where the syntax is the same as the current find() command, except there is an additional argument that tells MATLAB to perform the search along a specific dimension of the array X. And I would use it like

k = find(A < 5, 1, 'last', 2);

This would say, return the column index (dimension 2) of the last 1 element of each row of A satisfying the condition A < 5. I would expect k to be a column vector of length size(A,1) with the i-th element of k giving the desired column index.

Right now, I'm just using find() on each row inside of a for-loop:

for i = 1:size(A,1)
    k(i) = find(A(i,:) < 5, 1, 'last');
end

but I'm trying to figure out a way to do this without a for-loop. For a small matrix the for loop wouldn't be a problem, but my matrix, A, has millions of rows, so the for-loop has millions of iterations and is pretty slow. Any suggestions?

[jgg]

This will work, but I think it's slower than looping.

 table2array(rowfun(@(A)(find(A > 2,1,'last')),table(A)))

Answer 1

[Matt J]

The following finds the final value in each row, but you can easily adapt it to other dimensions/directions.

 function k = rowwiseLast(A)
 %Finds locations k(i) of final non-zero value in each row A(i,:), with k(i)=NaN if 
 %the entire row is zero.

      m=size(A,2);

      [val,loc] = max(  fliplr(logical(A)),  [],2);

      k=m+1-loc;
       k(val==0)=nan;

 end

If there is no value satisfying the criterion, than k(i)=nan for that row.

[Oliver Johnson]

Hi Matt, the FEX submission that you gave the link for is substantially more complicated than your solution posted here. I prefer your apparently more concise and elegant solution, but I'm wondering what the difference is between yours and the FEX submission (i.e. is the FEX submission more robust)?

[Matt J]

My solution will only find at most one occurrence in each row. The FEX submission truly is enabled with all the functionality of the FIND command, and can find n>1 occurrences if desired. Also, to implement direction='last', I need to apply the flip() command, which creates a theoretically unnecessary extra copy of the input array. This can be avoided by using a MEX file, as the FEX version appears to do.

Answer 2

[Matt J]

See this FEX submission,

http://www.mathworks.com/matlabcentral/fileexchange/24641-vectorized-find-with--first--option/content/findfirst/findfirst.m

Answer 3

[Image Analyst]

The for loop is fine. I mean, it's not like you have tens of millions of columns or anything do you?

[Oliver Johnson]

Actually, I do have millions of rows, so my for-loop has millions of iterations.

[Image Analyst]

Sorry, I meant rows since you're iterating over rows. Still might not be a problem even for millions of rows. How long is it taking? I just did 10 million iterations and it took 0.03 seconds, though I didn't have a find() in there.

Did you time jgg's suggestion?