Slow while loop generating coordinates in 3d matrix

Hi Matlab community,

I'm pulling cross-validation data from a 3-d matrix (a satellite image time series). The way I'm doing this (matlab novice here) is generating x,y and z random integer coordinates and testing to make sure I have a value at this point, in addition to testing for duplicates. However, using 'unique' to test for row duplicates inside my while loop really slows it down - i.e. it takes 1 hour to generate ~8600 points from a 650x450x120 matrix (though to be fair, about 85% this matrix is missing data).

What this loop looks like:

[m,n,o]=size(before); 
  %before is the 650x450x120 matrix
coordinate=[];
k=1;
  %crossnum is generally between 18,000 and 51,000, ~4% of available points
while k<=crossnum 
      %x
  coordinate(k,1)=randi(m,1);
      %y
  coordinate(k,2)=randi(n,1);
      %timeslice
  coordinate(k,3)=randi(o,1);
      %Make sure unique values and no NaNs 
  coordinatetest=unique(coordinate,'rows');
  [i,j]=size(coordinatetest);clear coordinatetest j;
      if (~isnan(before(coordinate(k,1),coordinate(k,2),coordinate(k,3))) && (i==k))
          k=k+1;
      else
          k=k;
      end
  %disp(k)
end

All the modifications I've tried so far (i.e. preallocating 'coordinate', using a boolean test combined with 'sum' instead of unique) made it much slower. If anyone has suggestions on a more efficient way of doing this I would be really appreciative. Thanks in advance!

Answer 1

[dpb]

idx=find(isfinite(before));   % all valid locations in 3D array with data
[i,j,k]=ind2sub(size(before),idx(randperm(numel(idx),crossnum)));  % sample _p_ unique from _N_
coordinate=[i j k];           % pack into output array

Eliminate the need for testing by

1. Only use valid data locations to start, and
2. Select desired number of unique locations from total number valid
3. Retrieve index in original array associated with selected subset

Then, having chosen the desired number from a linear ordering of the valid locations, rearrange those linear indices back into the desired 3D coordinates. Done. :)

See

doc randperm  % for details on how the selection works...

[schckngs]

THANK YOU!! 0.002 seconds... Beautiful!