Deleting multiple points: UniqueRows

UniqueRows can be a faster alternative to the Matlab unique(X,'rows');

Description

The description has been moved to:

http://www.advancedmcode.org/deleting-multiple-points-uniquerows.html

• MATLAB release: MATLAB 7.6 (R2008a)
• Other requirements: should work on all platform
• Other Files: license.txt, ; UniqueRows.m

[John D'Errico]

I like that this author has been improving his work from his initial submissions. This function has pretty good help. It has error checks. Lots of informative internal comments. (Well done there.)

I'd suggest a better H1 line, designed to be useful when the lookfor function is used. This should be a descriptive sentence that describes the function and includes intelligently chosen keywords that can be searched for by lookfor. This enables the function to be found next year when you forget the name UniqueRows.

I did a quick test of the code to compare speed. First look at an integer array.

A = round(rand(10000,4)*10);

See that both unique and UniqueRows got the same size results.

B = unique(A,'rows');
B2 = UniqueRows(A);

size(B),size(B2)
ans =
        6891 4
ans =
        6891 4

See that in this case, unique was faster. UniqueRows will probably pick up some time on floating point arrays.

timeit(@() unique(A,'rows'))
ans =
     0.056572

timeit(@() UniqueRows(A))
ans =
     0.071071

What bothers me is that the two codes do not produce the same results. The order is not important. But UniqueRows seems to miss some rows.

[sortrows(B),sortrows(B2)]
ans =
     0 0 1 1 0 0 1 5
     0 0 1 5 0 0 1 5
     0 0 2 5 0 0 2 5
     0 0 2 6 0 0 5 3
     0 0 5 3 0 0 5 4
     0 0 5 4 0 0 5 8
     0 0 5 8 0 0 5 9
     0 0 5 9 0 0 6 0
     0 0 6 0 0 0 6 3
     0 0 6 3 0 0 6 6
     0 0 6 6 0 0 7 1
     0 0 7 1 0 0 8 1
     0 0 7 2 0 0 8 4
     0 0 7 8 0 0 8 8
     0 0 8 1 0 0 10 4
     0 0 8 3 0 1 1 2
     0 0 8 4 0 1 1 8
     0 0 8 8 0 1 2 4
     0 0 9 8 0 1 2 8
     0 0 10 1 0 1 2 8
     0 0 10 4 0 1 3 2
...

For example, is there a row [0 0 2 6] in the original array? In fact, that row was A(260,:).

find(all(bsxfun(@eq,A,[0 0 2 6]),2))
ans =
   260
 
A(260,:)
ans =
     0 0 2 6

In fact, it appears that UniqueRows does not produce truly unique rows, even though it generated the correct number of rows. So there must be an indexing bug in the code.

B3 = unique(B2,'rows');
size(B3)
ans =
        5305 4

I was expecting to give this a good review. I'll hold off for now and look for repairs. With an improved H1 line and this bug fixed, if it is truly more efficient, I'd have given it a high rating. As it is with the bug it needs improvement.

[Luigi Giaccari]

I apologize for that it was a last minute modification which I thought not to be effetive but it was.
 I am sending an update version which I hope works fine. If you ever found something wrong please aknoledge me.

[Óscar J. Rubio Martín]

A fast and smart alternative.
Just one comment:
As you want to replace unique(A,'rows'), include the possibility of getting two outputs:
- B: the matrix of unrepeated rows
- ind: the indices of the rows of B in the original matrix A