apply function along dimension

Hi,

I am trying to find a way to determine if all elements in a hyperplane of an n-dimensional array are NaNs. To exemplify, with a 4-D array of size (3,5,2,7), I would like to determine along dimension 2 which of the 5 sub-arrays (hyperplanes) of size (3,1,2,7) are filled with just NaNs. The result should be a 5 elems vector (the size of the original array along dimension 2) containing true or false depending on whether there are just NaNs or not in the corresponding sub-array. The for loop answer would be

% A = <3,5,2,7>
  z = false(size(A,2),1);
  for i = 1:size(A,2)
     % turning each sub-array in a column vector and perform the check
     z(i) = all(isnan(reshape(A(:,i,:,:),prod(size(A(:,i,:,:))),1)));
  end

Since the arrays I work with are quite large this seems impractical, plus the dimension along which to perform the check is not given at the outset, but itself an argument of the function (i.e. A(:,i,:,:) would not work, because it should be A(:,:,i,:) if the check was performed along dimension 3). I think bsxfun and/or arrayfun could help in this regard, but their complex implementation doesn't help. Any suggestions on how this might be achieved?

Thanks in advance

Dan

Answer 1

[James Tursa]

Does this do what you want?

% x = nD array, k = dimension to check for all NaN's per plane
function result = nanplane(x,k)
z = 1:numel(size(x));
X = permute(x,[z(k) z(1:k-1) z(k+1:end)]);
X = reshape(X,size(X,1),[]);
result = all(isnan(X),2);
end

[Daniele Rocchetta]

Hi James,

after having thoroughly put your function to work in several input array scenarios I may encounter, I am happy to confirm it works perfectly. I got very close to this solution at one point, but what I missed is the second argument of the permutation, that I couldn't figure out.

Thanks again.

Best regards

Dan

[James Tursa]

If you are working with very large arrays, it might be a bit faster to do the isnan part first. That way the permute would be working with 1-byte data instead of 8-byte data and the memory copies involved might be a bit quicker to accomplish. E.g. something like this (CAUTION: untested):

% x = nD array, k = dimension to check for all NaN's per plane
function result = nanplane(x,k)
N = isnan(x); % <-- do the isnan first
z = 1:numel(size(x));
N = permute(N,[z(k) z(1:k-1) z(k+1:end)]); % <-- permute the 1-byte data
N = reshape(N,size(N,1),[]);
result = all(N,2);
end

[Daniele Rocchetta]

Hi and thanks for the suggestion; I had actually thought the same and implemented it already, so I can confirm it's tested and working, down to the degenerate cases of vectors and scalars for input array x. Brilliant.

All the best

Daniele

Answer 2

[Daniele Rocchetta]

Hi James,

at a first glance and trial run I would say yes! It is quite late here now, so I'll put it under thorough testing in the morning. But before that I wanted to thank you for your timely and kind reply. I'll keep you posted

Best regards

Dan