problem in using parralel computing toolbox

I want to speed up my computations using parfor. I am using the code written below,

 l=31:1:1000; % length-scale
 sig2=6.5:0.1:10; % noise-variance
 sigf2=1.3:0.1:10; % signal-variance

 L=zeros(length(l),length(sig2),length(sigf2));
 par=zeros(length(l),length(sig2),length(sigf2));
 lsig2=length(sig2);
 lsigf2=length(sigf2);

 % cross-validation
 parfor i=1:length(l)
    for j=1:lsig2
        for k=1:lsigf2
            par(i,j,k)=[l(i),sig2(j),sigf2(k)];
            L(i,j,k)=kcrossval(Xtrain,ytrain,M,l(i),sig2(j),sigf2(k));
        end
    end
 end

 but it is not working. I get error:

 Assignment has more non-singleton rhs dimensions than non-singleton subscripts.

When I remove the code

 par(i,j,k)=[l(i),sig2(j),sigf2(k)];

it works but gives the following warning:

 the entire array or struct 'sig2' is a broadcast variable. This might result in unnecessary communication overhead.

I have read the documentation but not able to find a solution. Please suggest the solutions.

Answer 1

[OCDER]

You are assigning 3 values, l(i),sig2(j),sigf2(k), into one value at par(i, j, k), which is not possible.

par(i,j,k)=[l(i),sig2(j),sigf2(k)];

But you don't really need this, since you know what values were used. If you want this, change par into a cell array instead, and do:

 par{i,j,k} = ...;

The broadcast variable warning is simply warning you that a constant vector is being passed to every worker every time, hence could be slow (especially if they're large matrices). In your case, sig2 and sigf2 are pretty small, so don't worry about this too much. This warning appears when you are accessing a small segment of a vector per loop. Notice that Xtrain and ytrain are not give you warning, because parfor knows these are broadcast variables (you aren't accessing " Xtrain(j)", but rather the full Xtrain vector).

To suppress these warnings, you could explicitly tell Matlab these are broadcast variables via another temporary variable like this:

 parfor i=1:length(l)
     sig2Temp  = sig2;  
     sigf2Temp = sigf2; 
     for j=1:lsig2
         for k=1:lsigf2
             L(i,j,k)=kcrossval(Xtrain,ytrain,M,l(i),sig2Temp(j),sigf2Temp(k));
         end
     end
 end

[Abhinav]

Thanks a lot for a nice explanation!

[OCDER]

You're welcome!