Remove NaNs in matrices within cell array

I have a matrix that has some data in its columns like this:

A = [1 NaN NaN 1 1; NaN NaN NaN 1 NaN; 1 1 NaN 1 NaN]

I want to extract the non NaN elements so that they are returned in a cell array where a cell holds the non NaN elements of the columns like this:

C = {[1;1],[1],[],[1;1;1],[1]}

This far I have tried to use the cellfun:

 C = num2cell(A,1);
 C(cellfun(@isnan,C, 'UniformOutput', false)) = [];

But i receive an error: Function 'subsindex' is not defined for values of class 'cell'.

I Managed to use for loop:

C = num2cell(A,1);

for i = 1:length(ct)
   inds = isnan(ct{i});
   ct{i}(inds) = [];
end

This does the trick, but I'm worried about the performance as A can be huge in my actual code.

Any help is more than welcome!

Edit: typo

Answer 1

[Andrei Bobrov]

out = cellfun(@(x)x(~isnan(x)),num2cell(A,1),'un',0);

with for-end loop

n = size(A,2);
out = cell(n,1);
for jj = 1:n
    out{jj} = A(~isnan(A(:,jj)),jj);
end

with accumarray

s = size(A);
ii = ones(s(1),1)*(1:s(2));
out = accumarray(ii(:),A(:),[],@(x){x(~isnan(x))});

[Hannu K]

Thank you! As it turns out, the for-loop approach is quicker.

Answer 2

[Matt J]

This does the trick, but I'm worried about the performance as A can be huge in my actual code.

Cellfun is usually less efficient than a for-loop. It basically runs a for-loop internally, but in a slightly less efficient way than a direct for-loop.

Also, even though A is huge, length(ct) will not be, correct? If length(ct) is really huge, that's a bigger problem than the question of whether to use cellfun or for-loops. It means you'll have lots of data scattered discontiguously in RAM and access to the cell contents will be very slow.

[Hannu K]

Thank you for the insight. I always thought that cellfun does some magic that outperforms the for loop.

Answer 3

[Matt J]

If all elements of A are either NaN or non-zero, then it would be better to convert to a sparse matrix, rather than use cell arrays

 [i,j,s]=find(~isnan(A));
 [m,n]=size(A);

 A=sparse(i,j,s,m,n);

Answer 4

[Matt J]

Yet another alternative,

   map=~isnan(A); 

   [i,j]=find(map); 
   k=histc(j,1:size(A,2)); 

   C=mat2cell(A(map),k,1).';

Answer 5

[Azzi Abdelmalek]

out1=arrayfun(@(x) A(~isnan(A(:,x)),x),1:size(A,2),'un',0)

Answer 6

[David Sanchez]

for single line solution, A. Abdelmalek's is the solution. You can perform the same by using the following:

A = [3 NaN NaN 2 4; NaN NaN NaN 5 NaN; 6 7 NaN 8 NaN];
[rows cols] = size(A); % number of rows and columns in your matrix
D = mat2cell(A,rows,ones(1,cols)); % cell to hold your data
for k=1:numel(D)
    C = ~isnan(D{k});
    D{k}(C==0) = [];
end

It works with values different to 1 and the code is quite readable.

Answer 7

[Bob Freeman]

do find(NaN == 0)