passing matrix (dtype('float64')) containing NaNs to MATLAB

[lijing-lin]

However, what I got in MATLAB is a cell array, which flatten the matrix. Assume the original matrix is 5 by 4, I end up with a 20 by1 cell array. Moreover, all NaNs are treated as strings 'NaN' in the cell. A bug or it is designed to be like this?

[arokem]

Thanks for reporting - that does seem like a bug to me. Would be
non-trivial to fix, but let me take a look.

On Mon, Dec 1, 2014 at 6:30 PM, leizing notifications@github.com wrote:

However, what I got in MATLAB is a cell array, which flatten the matrix.
Assume the original matrix is 5 by 4, I end up with a 20 by1 cell array.
Moreover, all NaNs are treated as strings 'NaN' in the cell. A bug or it is
designed to be like this?

—
Reply to this email directly or view it on GitHub
#100.

[lijing-lin]

In the following is a quick fix for myself. Say I pass float matrix S to matlab. I add the following lines before using the matrix:

%%matlab -i S
for i=1:length(S)
    if isnumeric(S{i}), continue, end
    S{i}(cellfun(@(x) ~isnumeric(x),S{i})) = {nan};
    S{i} = cell2mat(S{i});
end
S = cell2mat(S);

[blink1073]

This behaviour is fixed in the latest master:

In [25]: S = np.ones((5,4), dtype=np.float64)

In [26]: S[0] = np.nan

In [27]: %%matlab -i S -o T
T = S;
   ....: 

In [28]: T
Out[28]: 
array([[ nan,  nan,  nan,  nan],
       [  1.,   1.,   1.,   1.],
       [  1.,   1.,   1.,   1.],
       [  1.,   1.,   1.,   1.],
       [  1.,   1.,   1.,   1.]])