novelty-detection-ncm/Agnotron-model/Clustering Network/dataGen.m

%%% data generator
clear 
close all

%% Configs

numDims = 100;
numClusters = 10;
stdClust = 0.5; %std within cluster - not so bad at 1
clusterSep = 25; % std of distribution from which clusters are drawn
ptsPerCluster = 250;

%% 

stds = stdClust .* ones(numClusters, 1);
clustMeans = clusterSep*rand(numClusters, numDims); % original data is strictly pos.

allPts_unshuff = [];
clustIDs_unshuff = [];
for clustIdx = 1:numClusters
    newBlob = [];
    for dim = 1:numDims
        newBlobDim = clustMeans(clustIdx,dim)+stds(clustIdx).*randn(ptsPerCluster, 1);
        newBlob = [newBlob newBlobDim]; %100 pts in 100D space
    end
    
    clustIDs_unshuff(end+1:end+ptsPerCluster) = clustIdx;
    allPts_unshuff = [allPts_unshuff; newBlob]; 
end

allPts_dm_unshuff = allPts_unshuff;

for dim = 1:numDims
    allPts_dm_unshuff(:,dim) = allPts_dm_unshuff(:,dim) - mean(allPts_dm_unshuff(:,dim));
end


% calc clust means for use in dot product calculation
clustMeans = zeros(size(clustMeans));
for clustIdx = 1:numClusters
    startPt = ((clustIdx-1)*ptsPerCluster) + 1; %first pt in cluster
    endPt = startPt + ptsPerCluster - 1; %last pt in cluster
    thisBlob = allPts_dm_unshuff(startPt:endPt, :);
    thisMean = mean(thisBlob, 1); 
    clustMeans(clustIdx, :) = thisMean ./ norm(thisMean); %normalized
end

shuffIdx = randperm(ptsPerCluster*numClusters);

allPts = allPts_unshuff(shuffIdx, :);
allPts_dm = allPts_dm_unshuff(shuffIdx, :);
clustIDs = clustIDs_unshuff(shuffIdx);



ccDot = zeros(numClusters);
ccDist = zeros(numClusters);
for clust_i = 1:numClusters
    for clust_j = 1:numClusters
        ccDot(clust_i, clust_j) = abs(clustMeans(clust_i,:) * clustMeans(clust_j,:)');
        ccDist(clust_i, clust_j) = norm(clustMeans(clust_i,:) - clustMeans(clust_j,:));
    end
end

ptDist = zeros(numClusters*ptsPerCluster);
for ptIdx = 1:size(allPts, 1)
    for ptJdx = 1:size(allPts, 1)
        ptDist(ptIdx, ptJdx) =...
            norm(allPts_unshuff(ptIdx,:) - allPts_unshuff(ptJdx,:));
    end
end


if numDims == 2
    figure(10);
    plot(allPts_dm(:,1)', allPts_dm(:,2)', '.')
    grid on;
    hold on;
    plot(0, 0, 'r*')
    hold off;
end


figure(1); % useful bc weight vectors should be as separate as possible
imagesc(ccDot); colorbar
title('Dot Products of Cluster Means')
xlabel('Cluster')
ylabel('Cluster')

figure(2)
imagesc(ccDist); colorbar
title('Average Distance between Clusters')
xlabel('Cluster')
ylabel('Cluster')

figure(3)
imagesc(ptDist); colorbar
title('Pointwise Distances')



savename = ['genData' num2str(numDims) 'D_' num2str(numClusters)...
    'c_' num2str(ptsPerCluster) 'ppc.mat'];

clearvars -except allPts allPts_dm allPts_unshuff...
    allPts_dm_unshuff clustIDs clustIDs_unshuff savename

save(savename)