Alternative to using structs as reduction variables in parfor loop

My code initializes a struct (say init_s) with about 9 fields before going into a parfor loop of iterations (as s). Over a single iteration, the empty arrays in s are filled with data. I'd like to add these arrays together, without having to create 9 different variables for each field. How could I go about doing that? Here is a minimum working example:

init_s = struct('f1',ones(10,1),'f2',zeros(10,1),'f3',50*ones(15,1))
parfor iter = 1:10
    s = init_s;
    s.f1 = s.f1 * 2;
    s.f2 = s.f2 + 5;
    s.f3 = s.f3 - 1;
end

I'd like to add the generated arrays individually (s.f1 over all iter, s.f2 over all iter, etc.) so that I can average them later. How can I do that without having to create a lot of variables? Or if I have to, how do I access them in a loop using fieldnames?

Answer 1

[Walter Roberson]

s(1:2) = struct('f1',ones(10,1),'f2',zeros(10,1),'f3',50*ones(15,1))

s = 1x2 struct array with fields:
    f1
    f2
    f3

sc = struct2cell(s(:))

sc = 3x2 cell array
    {10×1 double}    {10×1 double}
    {10×1 double}    {10×1 double}
    {15×1 double}    {15×1 double}

totals = arrayfun(@(R) sum(cat(3,sc{R,:}),3), (1:size(sc,1)).', 'uniform', 0)

totals = 3x1 cell array
    {10×1 double}
    {10×1 double}
    {15×1 double}

struct_total = cell2struct(totals, fieldnames(s), 1)

struct_total = struct with fields:
    f1: [10×1 double]
    f2: [10×1 double]
    f3: [15×1 double]

[Tejas]

How would I carry the struct_total over the next iteration though? The struct that I obtain at the end of each iteration is different, so I'd be adding each to a sum of the previous structs. How can I access the sum of the previous structs in the next loop? Using your idea, I believe it'd be something like

init_s = struct('f1',ones(10,1),'f2',zeros(10,1),'f3',50*ones(15,1));
results = struct2cell(structfun(@(x) x*0,init_s,'un',0));
parfor iter = 1:10
    s = init_s;
    s.f1 = s.f1 * 2;
    s.f2 = s.f2 + 5;
    s.f3 = s.f3 - 1;
    temp_res = [results,struct2cell(s)];
    results = arrayfun(@(R) sum(cat(3,temp_res{R,:}),3), (1:size(temp_res,1)).', 'uniform', 0);
end

Except I cannot build temp_res because results is not a temporary variable.

[Walter Roberson]

If you need to know the value of the total up to this point, then you cannot do that with parfor, as it implies that order is important, but parfor cannot promise any particular order.

If you do not need to know the total until after the parfor, then store the values and do the total afterwards.

init_s = struct('f1',ones(10,1),'f2',zeros(10,1),'f3',50*ones(15,1));
results = struct2cell(structfun(@(x) x*0,init_s,'un',0));
parfor iter = 1:10
    s = init_s;
    s.f1 = s.f1 * 2;
    s.f2 = s.f2 + 5;
    s.f3 = s.f3 - 1;
    all_s(iter,1) = s;
end

sc = struct2cell(all_s(:));
totals = arrayfun(@(R) sum(cat(3,sc{R,:}),3), (1:size(sc,1)).', 'uniform', 0);
struct_total = cell2struct(totals, fieldnames(all_s), 1);

The code was designed to not care about whether all_s is Nx1 or 1xN or NxM or even higher dimension, as long as the individual arrays in the fields are vectors or 2D. It would, however, need a small change if the fields could be higher dimensional. Though if you care about the shape of all_s then afterwards

struct_total = reshape(struct_total, size(all_s));

[Tejas]

This works for me. I only care about the total after the iterations, since I want to average it. I think the point I missed was that variables collecting data every iteration are available outside the loop, unlike the temporary variables. I was confused why all_s can be accessed outside the loop. Anyway, thank you!