Nearest-Neighbour-Search-Neural-Networks-Product-Quantization/TEXMEX/WNN Library/QuantHammingCanonicalIndices_ceil.m

function [minDh, iCminDist]=QuantHammingCanonicalIndices_ceil(CWidx,nNets)

N=size(CWidx,1);
iCminDist=zeros(1,N);
idxs=1:nNets;
%rPermidx=randperm(N,nNets);
pperc=[];
minDh=0;

[~, ~, icw]=unique(CWidx,'rows');
uicw=unique(icw);
luicw=length(uicw);
%uniNets=nNets/luicw;
rPermidx=[];
for jj=1:luicw
     ficw=find(icw==uicw(jj));
     adaptidx=randperm(length(ficw),ceil(nNets*length(ficw)/N));
     %size(adaptidx)
     %size(ficw(adaptidx))
     rPermidx=[rPermidx ficw(adaptidx)'];
end

moreidx=length(rPermidx)-nNets;

if moreidx>0
    remrand=randperm(length(rPermidx),moreidx);
    rPermidx(remrand)=[];
end

%size(rPermidx)

for ii=1:N

     if any(rPermidx==ii)
         iCminDist(ii)=find(rPermidx==ii,1,'first');
     else
         % Approximation of half of the hamming distance
         dH=sum(bsxfun(@ne,CWidx(rPermidx,:),repmat(CWidx(ii,:),nNets,1))');
         minDh=minDh+min(dH);
         allOfTheMins=idxs(dH==min(dH));
         iCminDist(ii)=allOfTheMins(randi(length(allOfTheMins)));
         %[~,iCminDist(ii)]=min(dH);
     end
         perc=sprintf('%2.2f%%',ii/N*100);
         fprintf(1,'%s',sprintf('\b')*ones(1,length(pperc)));
         fprintf(1,'%s',perc);pperc=perc;
end
minDh=minDh/N;
fprintf(1,'%s',sprintf('\b')*ones(1,length(pperc)));
end