How to vectorize operations on cells?

Question

Given B = {"7" "8" "9"}, I'd like to use its values as respective replacements for the ultimate elements of each cell in A, i.e. transform

A = {{1 2} {3 4} {5 6}}

into

A = {{1 "7"} {3 "8"} {5 "9"}}

In unvectorized form, this can be written as:

nn = cellfun(@numel, A)
for i = 1:numel(a)
  A{i}{nn(i)} = B{i};
end

All sub-cells may have the same length, but a solution for sub-cells of arbitrary length would be ideal.

The only solution I found involves making a reshaped copy of A:

nn = cellfun(@numel, A)      #=> 2 2 2
A2 = horzcat(A{:})           #=> {1 2 3 4 5 6}
jj = cumsum(nn)              #=> 2 4 6
[A2{jj}] = B{:}              #=> {1 "7" 3 "8" 5 "9"}
ii = [1 jj(1:end-1)+1]       #=> 1 3 5
A = cellslices(A2,ii,jj,2)   #=> {{1 "7"} {3 "8"} {5 "9"}

I'd have hoped something like the below would work - but it doesn't:

octave:1> [[A{:}]{ii}] = B{:}
error: invalid lvalue function called in expression

Is there a way to vectorize this operation?

If all sub-cells may have the same length you can use a matrix instead of nested cells. — rahnema1, May 21 '18 at 06:18
You can always expect to have unvectorised code when dealing with cells. Btw `cellfun` is a one-liner wrapper for a loop. It is still a loop. — Sardar Usama, May 21 '18 at 06:33
@SardarUsama It is a loop written in C++, and that makes it way faster than any octave-code loop. See [libinterp/corefcn/cellfun.cc](https://github.com/KubaO/Octave/blob/master/libinterp/corefcn/cellfun.cc) — Kuba hasn't forgotten Monica, May 21 '18 at 06:39
@rahnema1 Some of the cells may have non-numeric values, and thus a matrix won't work. — Kuba hasn't forgotten Monica, May 21 '18 at 06:39
You've explicitly asked for vectorisation, not avoiding a `for`/`while` loop. — Sardar Usama, May 21 '18 at 06:43
@KubaOber The fact that with `cellfun` the loop is in C++ or Java rather than in Matlab code doesn't necessarily make it fast. In fact, `cellfun` is often slower than `for`; see for example [here](https://es.mathworks.com/matlabcentral/answers/42335-array-cellfun-vs-for-loop) or [here](https://stackoverflow.com/questions/18284027/cellfun-versus-simple-matlab-loop-performance?utm_medium=organic&utm_source=google_rich_qa&utm_campaign=google_rich_qa). Anyway, performance results are dependent on the Matlab version — Luis Mendo, May 21 '18 at 10:32

Sardar Usama · Answer 1 · 2018-05-21T09:22:21.863

1

As learned from the comments that you actually just want to avoid an explicit for/while loop and are okay with cellfun (the use of which is not vectorisation), the process can be replicated with cellfun as follows:

A = cellfun(@(k) {A{k}{1:end-1} B{k}}, num2cell(1:numel(A)), 'un', 0);

or with arrayfun as:

 A = arrayfun(@(k) {A{k}{1:end-1} B{k}}, 1:numel(A), 'un', 0);

edited May 21 '18 at 09:22

answered May 21 '18 at 07:23

Sardar Usama

19,536
9
36
58

How to vectorize operations on cells?

1 Answers1