Speed up Some Code

Hi all. I'm trying to speed up the code below. It's from the SMOTE function that was wrtitten for MATLAB and works really well. The only thing is that the loops are not fast. I've looked at Parfor but that wouldn't work and I can't see how I can vectorise it either. It could be I just have to suck it up and wait but just in case, the code is :-

% This is where the magic happens ;-)
%   X : Observational matrix (rows are observations, columns are variables)
%   J : Synthesization vector. It has the same length as the number of
%       observations (rows) in X. J determines how many times each 
%       observation is used as a base for synthesization.
%   k : Number of nearest neighbors to consider when synthesizing.
function Xn = simpleSMOTE(X,J,k)
    tic
    HNSMdl = hnswSearcher(X);   % To remove this, comment out this line and replace the HNSMdl below with X
    [idx, ~] = knnsearch(HNSMdl,X,'k',k+1); % Find nearest neighbors (add one to the number of neighbors to find, as observations are their own nearest neighbor)
    toc
    Xn = nan(sum(J),size(X,2)); % Pre-allocate memory for synthesized observations
    % Iterate through observations to create to synthesize new observations
    for ii=1:numel(J)
        P = randperm(k,J(ii))+1; % Randomize nearest neighbor pick (never pick first nearest neighbor as this is the observation itself)
        for jj=1:J(ii)
            x = X(idx(ii,1),:); % Observation
            xk = X(idx(ii,P(jj)),:); % Nearest neighbor
            Xn(sum(J(1:ii-1))+jj,:) = (xk-x)*rand+x; % Synthesize observation
        end
    end
end

It's from the 'for ii' bit that is slow and there are around 750000 items of 13 variables.

Steve

Answer 1

[Voss]

% generate and collect all the P's first
nJ = numel(J);
P = cell(1,nJ);
for ii = 1:nJ
    P{ii} = randperm(k,J(ii))+1;
end
P = [P{:}];

% then do the rest in one fell swoop
ii = repelem(1:nJ,J);
xk = X(idx(sub2ind(size(idx),ii,P)),:);
x = X(idx(ii,1),:);
Xn = (xk-x).*rand(numel(ii),1)+x;

[Stephen Gray]

Much quicker thanks! Also a good example of how to vectorise.

Steve

[Voss]

You're welcome!