buildingHypernetwork.m

%function [pop, i_error, t_error] = buildingHypernetwork()
%UNTITLED4 Summary of this function goes here
%   Detailed explanation goes here
run 'configure';
episode_num = 4;
hn_path = [pop_path 'new' sprintf('%d', episode_num), '/'];
mkdir(hn_path);
load([region_centroid_path 'tot_reg_cent.mat']);
IK = size(C, 2);

%% load Dic
load([dic_path './dic.mat']);
%load([hn_path 'pop1127.mat']);
%% load preprocessing
total_pair = [];
total_bundle = [];
for pp=episode_num:episode_num
    disp(sprintf('The %d-th dataset', pp));
    disp('Loading data');
    pair_name = sprintf('./pair%d.mat', pp);
    load([pair_path pair_name]);
    total_pair = [total_pair; pair];
    bundle_name = sprintf('./bundle%d.mat', pp);
    load([bundle_path bundle_name]);
    total_bundle = [total_bundle; bundle];
end
pair = total_pair;
clear total_pair;
bundle = total_bundle; 
clear total_bundle;
%여기서 지금까지 갖고있는 에피소드 모두 total로해서 집어넣는다.

pop = [];
regionDB = cell(1,1);   % 하이퍼에지로 사용된 bundle region을 저장하는 저장소
regionIDXDB = zeros(1, 3); % 사용된 region의 인덱스 저장소 (0 0 0)
%regionDBcnt = 1;
regionIDXDBcnt = 1; %저장소의 현재 인덱스
regionClusterCnt = zeros(IK, 1);
dicCnt = zeros(size(dic, 1), 1);

i_error = zeros(1, 3);
t_error = zeros(1, 3);
error_cnt = 1;
for pp=1:size(pair, 1)
    st = sprintf('The %dth data instance===============', pp);
    disp(st);
    if size(fieldnames(bundle{pp, 1})) < 2
        pop_name = sprintf('pop%d.mat', pp);    
        save([hn_path pop_name], 'pop', '-v7.3');         
        continue;
    end   
    seed_region = [];
    seed_word = [];  
    org_img_clusters = [];
    for j=1:size(bundle{pp, 1}, 2)
        org_img_clusters(j, 1) = bundle{pp, 1}(1,j).region_cluster;
    end    
    for i=1:epoch     
        %% Seed 생성
        seed_region = bundle{pp,1}(1,randsample(size(bundle{pp, 1}, 2), 1, false)).region_cluster;
        seed_word = pair{pp, 2}{1, randsample(size(pair{pp, 2}, 2), 1, false)};
        %% Error estimation 
        [gen_data, selected_idx] = generate(pop, seed_region, seed_word, regionDB, size(bundle{pp, 1}, 2)*2, size(pair{pp, 2}, 2)*2);               
        gen_img_clusters = [];
        for j=1:size(gen_data{1,1}, 2)
            gen_img_clusters(1, j) = regionDB{gen_data{1,1}(1,j)}.region_cluster;
        end
        i_error(error_cnt, 1) = pp;i_error(error_cnt, 2) = i;
        i_error(error_cnt, 3) = 1-sum(ismember(org_img_clusters, gen_img_clusters)) / size(org_img_clusters, 1);
        t_error(error_cnt, 1) = pp;t_error(error_cnt, 2) = i;        
        if size(gen_data{1,2}, 2) > 0
            t_error(error_cnt, 3) = 1-sum(ismember(pair{pp, 2}, gen_data{1,2})) / size(pair{pp, 2}, 2);
        else
            t_error(error_cnt, 3) = 1;
        end
        error_cnt = error_cnt + 1;
        %% Correction and Update weights        
        HE = makeHyperedge(pop, pair{pp, 2}, bundle(pp, 1), pp, sampling_rate, bundle{pp, 1});
        for j=1:size(HE, 1)
            he = HE(j);
            for k=1:size(he.i_idx, 2)
                if sum(ismember(he.i_absIdx(k), regionIDXDB(:, 1))) < 1
                    regionIDXDB(regionIDXDBcnt, 1) = he.i_absIdx(k);
                    regionIDXDB(regionIDXDBcnt, 2) = he.file_idx;
                    regionIDXDB(regionIDXDBcnt, 3) = he.i_idx(k);
                    regionDB{regionIDXDBcnt, 1} = bundle{he.file_idx, 1}(1, he.i_idx(k));
                    HE(j).i_DBidx(1, k) = regionIDXDBcnt;
                    regionIDXDBcnt = regionIDXDBcnt + 1;                    
                else                    
                    HE(j).i_DBidx(1, k) = (find(ismember(regionIDXDB(:, 1), he.i_absIdx(k))));
                end
            end            
        end
        pop = [pop; HE];        
        pop = updateWeights(pop, selected_idx, pair{pp, 2}, org_img_clusters, regionClusterCnt, dicCnt, regionDB, dic);   

    end
    if mod(pp, 20) == 0
        pop_name = sprintf('pop%d.mat', pp);    
        save([hn_path pop_name], 'pop', '-v7.3');  
    end    
end
pop_name = sprintf('pop%d.mat', pp);    
save([hn_path pop_name], 'pop', '-v7.3');