Automatically create a several layer structure from a cell array

Hello,

I have a signal that vary along three parameters a,b and c wich can take several values.

a can take n different values a1,a2...an

b can take m different values b1,b2,...,bm

c can take p different values c1,c2...cp.

So far I have all the variations stocked in a cell array C of size n*m*p such that C(x,y,z) correspond to my signal with parameter ax, by and cz.

This works but is not very easy to use. Let's say I want to get the signal corresponding to certain parameter I first need to go find the indexes of the wanted parameters (because I know the value of a that I want but not the index to wich it correspond).

My idea was to create a structure S such that

S.ax.by.cz=C(x,y,z)

that way I wont need a correpondance table

So far the only way to do it that I found is

for i=1:n
    for j=1:m
        for k=1:p
            S.(a_value(i)).(b_value(j)).(c_value(j))=C(i,j,k) 
        end
    end
end

I was wondering if a nicer/more efficient solution exist?

PS: I can't put the value directly in the structure when I generate it because the generation is inside a parfor loop. If you have a solution to that issue that would also solve my problem

thanks in advance

[Dyuman Joshi]

" Let's say I want to get the signal corresponding to certain parameter I first need to go find the indexes of the wanted parameters..."

Can you give an example? Say the data is

C{1}=[61 43 50 53 95]; %a
C{2}=[63 96 51 45 57 27]; %b
C{3}=[8 67 76 76 83 50 7]; %c

Now what do you want to access?

Also, it's not clear as to how you are storing the data as C(x,y,z)?

[Stephen23]

"I was wondering if a nicer/more efficient solution exist?"

ND arrays perhaps.

You indicate that the "signal" varies with some parameters which correspond very neatly to indices 1:m, 1:n, and 1:p. It is unclear why you think the replacing this simple index-correspondence with fieldnames would be an improvement (which most likely would be slower and more complex to access).

[Matteo Bonhomme]

Let me give a concrete exemple

Let's say I have a salary depending on three parameters age, job and country

What I have is

salary=arrayfun(@(x) x,randi(100000,[2,2,2]),'UniformOutput',false);

job_array=["doctor","waiter"];
country_array=["USA","France"];
age_array=["less_than_40","more_than_40"];
%How it is now (salary is a 2*2*2 cell array)
salary(1,1,1); %the salary of a doctor in USA under 40 years old
salary(1,1,2) %the salary of a doctor in USA over 40 years old

ans = 1×1 cell array
    {[65633]}

% Now let's say I want the salary of a french waiter less than 40 this is
% what I have to do
idx_waiter=find(strcmp(job_array,"waiter"));
idx_country=find(strcmp(country_array,"France"));
idx_age=find(strcmp(age_array,"less_than_40"));
salary_waiter_french_less_40=salary(idx_waiter,idx_country,idx_age)

salary_waiter_french_less_40 = 1×1 cell array
    {[96705]}

What I want is to create a structure so that

%could be created like this but is very slow
for i=1:length(job_array)
    for j=1:length(country_array)
        for k=1:length(age_array)
            salary_struct.(job_array(i)).(country_array(j)).(age_array(k))=salary(i,j,k) ;
        end
    end
end

salary_struct.doctor.USA.more_than_40 %the salary of a doctor in USA over 40 years old

ans = 1×1 cell array
    {[65633]}

salary_struct.waiter.France.less_than_40 %salary of a french waiter less than 40

ans = 1×1 cell array
    {[96705]}

What I have is technicaly working but the code would not be lisible.

I'm also supposed to share this data with other people so having a clear code is really usefull.

[Voss]

You can keep the indexing approach, but make a function that does the necessary indexing so as to keep the code more legible. If you make it a nested function then you don't have to pass job_array, country_array, etc., to it.

Example:

main()

salary_doctor_USA_less_40 = 1×1 cell array
    {[90751]}

salary_doctor_FRE_less_40 = 1×1 cell array
    {[40773]}

salary_waiter_USA_less_40 = 1×1 cell array
    {[96740]}

salary_waiter_FRE_less_40 = 1×1 cell array
    {[1566]}

salary_doctor_USA_more_40 = 1×1 cell array
    {[67015]}

salary_doctor_FRE_more_40 = 1×1 cell array
    {[93363]}

salary_waiter_USA_more_40 = 1×1 cell array
    {[22547]}

salary_waiter_FRE_more_40 = 1×1 cell array
    {[57732]}

function main()

salary=arrayfun(@(x) x,randi(100000,[2,2,2]),'UniformOutput',false);
job_array=["doctor","waiter"];
country_array=["USA","France"];
age_array=["less_than_40","more_than_40"];

salary_doctor_USA_less_40 = get_salary("doctor","USA","less_than_40")
salary_doctor_FRE_less_40 = get_salary("doctor","France","less_than_40")
salary_waiter_USA_less_40 = get_salary("waiter","USA","less_than_40")
salary_waiter_FRE_less_40 = get_salary("waiter","France","less_than_40")
salary_doctor_USA_more_40 = get_salary("doctor","USA","more_than_40")
salary_doctor_FRE_more_40 = get_salary("doctor","France","more_than_40")
salary_waiter_USA_more_40 = get_salary("waiter","USA","more_than_40")
salary_waiter_FRE_more_40 = get_salary("waiter","France","more_than_40")


    function s = get_salary(job,country,age_group)
        idx_job = find(strcmp(job_array,job));
        idx_country = find(strcmp(country_array,country));
        idx_age = find(strcmp(age_array,age_group));
        s = salary(idx_job,idx_country,idx_age);
    end

end

[Matteo Bonhomme]

Yeah that would work

I ended up biting the bullet and just do this

for i=1:length(job_array)
    for j=1:length(country_array)
        for k=1:length(age_array)
            salary_struct.(job_array(i)).(country_array(j)).(age_array(k))=salary(i,j,k) ;
        end
    end
end

It's not taking that much time (compare to the rest of the script).

[Stephen23]

Aaaah, so you actually have text data.... in which case, a few dynamic fieldnames as you show is probably reasonably efficient. Dynamic fieldnames in some loops is likely the most efficient approach (you could use SETFIELD, but it is slower).

@Voss: forcing meta-data into variable names should be avoided: https://www.mathworks.com/matlabcentral/answers/225435-save-variable-as-string-from-user-input#answer_184104

The approach you show is not easily generalizable nor easily expandable.

[Voss]

"forcing meta-data into variable names should be avoided"

Yes, you are right. I was following @Matteo Bonhomme's naming pattern, so that he could easily see how the nested function would work.

"The approach you show is not easily generalizable nor easily expandable."

I disagree.

Answer 1

[Eric Sofen]

Another approach is to string this out into a table, converting what were labels along each dimension into grouping variables. Starting with the same salary data example, the trick is to use ndgrid.

salary=arrayfun(@(x) x,randi(100000,[2,2,2]),'UniformOutput',false);
job_array=["doctor","waiter"];
country_array=["USA","France"];
age_array=["less_than_40","more_than_40"];

salary = [salary{:}]'

salary = 8×1

13085 40527 9337 23364 27223 34191 94748 16327

[Job, Country, Age] = ndgrid(job_array,country_array,age_array)

Job = 2×2×2 string array
Job(:,:,1) = 

    "doctor"    "doctor"
    "waiter"    "waiter"


Job(:,:,2) = 

    "doctor"    "doctor"
    "waiter"    "waiter"

Country = 2×2×2 string array
Country(:,:,1) = 

    "USA"    "France"
    "USA"    "France"


Country(:,:,2) = 

    "USA"    "France"
    "USA"    "France"

Age = 2×2×2 string array
Age(:,:,1) = 

    "less_than_40"    "less_than_40"
    "less_than_40"    "less_than_40"


Age(:,:,2) = 

    "more_than_40"    "more_than_40"
    "more_than_40"    "more_than_40"

t = table(Job(:), Country(:), Age(:), salary, VariableNames=["Job", "Country","Age","Salary"])

t = 8×4 table
      Job       Country          Age          Salary
    ________    ________    ______________    ______

    "doctor"    "USA"       "less_than_40"    13085 
    "waiter"    "USA"       "less_than_40"    40527 
    "doctor"    "France"    "less_than_40"     9337 
    "waiter"    "France"    "less_than_40"    23364 
    "doctor"    "USA"       "more_than_40"    27223 
    "waiter"    "USA"       "more_than_40"    34191 
    "doctor"    "France"    "more_than_40"    94748 
    "waiter"    "France"    "more_than_40"    16327 

% Now you can either use logical indexing to select data or grouping to do
% calculations that would be equivalent to slicing in your original array:

t.Salary(t.Job=="waiter" & t.Country == "France" & t.Age == "less_than_40")

ans = 23364

groupsummary(t,"Country","mean","Salary")

ans = 2×3 table
    Country     GroupCount    mean_Salary
    ________    __________    ___________

    "France"        4            35944   
    "USA"           4            28756