wolffd@0: function [dPsi, M, SO_time] = cuttingPlaneFull(k, X, W, Ypos, Yneg, batchSize, SAMPLES, ClassScores)
wolffd@0: %
wolffd@0: % [dPsi, M, SO_time] = cuttingPlaneFull(k, X, W, Yp, Yn, batchSize, SAMPLES, ClassScores)
wolffd@0: %
wolffd@0: %   k           = k parameter for the SO
wolffd@0: %   X           = d*n data matrix
wolffd@0: %   W           = d*d PSD metric
wolffd@0: %   Yp          = cell-array of relevant results for each point
wolffd@0: %   Yn          = cell-array of irrelevant results for each point
wolffd@0: %   batchSize   = number of points to use in the constraint batch
wolffd@0: %   SAMPLES     = indices of valid points to include in the batch
wolffd@0: %   ClassScores = structure for synthetic constraints
wolffd@0: %
wolffd@0: %   dPsi        = dPsi vector for this batch
wolffd@0: %   M           = mean loss on this batch
wolffd@0: %   SO_time     = time spent in separation oracle
wolffd@0: 
wolffd@0:     global SO PSI DISTANCE CPGRADIENT;
wolffd@0: 
wolffd@0:     [d,n,m] = size(X);
wolffd@0:     D       = DISTANCE(W, X);
wolffd@0: 
wolffd@0:     M       = 0;
wolffd@0:     S       = zeros(n);
wolffd@0:     dIndex  = sub2ind([n n], 1:n, 1:n);
wolffd@0: 
wolffd@0:     SO_time = 0;
wolffd@0: 
wolffd@0:     if isempty(ClassScores)
wolffd@0:         TS  = zeros(batchSize, n);
wolffd@0:         parfor i = 1:batchSize
wolffd@0:             if i <= length(SAMPLES)
wolffd@0:                 j = SAMPLES(i);
wolffd@0: 
wolffd@0:                 if isempty(Ypos{j})
wolffd@0:                     continue;
wolffd@0:                 end
wolffd@0:                 if isempty(Yneg)
wolffd@0:                     % Construct a negative set 
wolffd@0:                     Ynegative = setdiff((1:n)', [j ; Ypos{j}]);
wolffd@0:                 else
wolffd@0:                     Ynegative = Yneg{j};
wolffd@0:                 end
wolffd@0:                 SO_start        = tic();
wolffd@0:                     [yi, li]    =   SO(j, D, Ypos{j}, Ynegative, k);
wolffd@0:                 SO_time         = SO_time + toc(SO_start);
wolffd@0: 
wolffd@0:                 M               = M + li /batchSize;
wolffd@0:                 TS(i,:)         = PSI(j, yi', n, Ypos{j}, Ynegative);
wolffd@0:             end
wolffd@0:         end
wolffd@0: 
wolffd@0:         % Reconstruct the S matrix from TS
wolffd@0:         S(SAMPLES,:)    = TS;
wolffd@0:         S(:,SAMPLES)    = S(:,SAMPLES) + TS';
wolffd@0:         S(dIndex)       = S(dIndex) - sum(TS, 1);
wolffd@0:     else
wolffd@0: 
wolffd@0:         % Do it class-wise for efficiency
wolffd@0:         batchSize = 0;
wolffd@0:         for j = 1:length(ClassScores.classes)
wolffd@0:             c       = ClassScores.classes(j);
wolffd@0:             points  = find(ClassScores.Y == c);
wolffd@0: 
wolffd@0:             Yneg    = find(ClassScores.Yneg{j});
wolffd@0:             yp      = ClassScores.Ypos{j};
wolffd@0:             
wolffd@0:             if length(points) <= 1
wolffd@0:                 continue;
wolffd@0:             end
wolffd@0: 
wolffd@0:             batchSize = batchSize + length(points);
wolffd@0:             TS      = zeros(length(points), n);
wolffd@0:             parfor x = 1:length(points)
wolffd@0:                 i           = points(x);
wolffd@0:                 yl          = yp;
wolffd@0:                 yl(i)       = 0;
wolffd@0:                 Ypos        = find(yl);
wolffd@0:                 SO_start    = tic();
wolffd@0:                     [yi, li]    = SO(i, D, Ypos, Yneg, k);
wolffd@0:                 SO_time     = SO_time + toc(SO_start);
wolffd@0: 
wolffd@0:                 M           = M + li;
wolffd@0:                 TS(x,:)     = PSI(i, yi', n, Ypos, Yneg);
wolffd@0:             end
wolffd@0: 
wolffd@0:             S(points,:) = S(points,:) + TS;
wolffd@0:             S(:,points) = S(:,points) + TS';
wolffd@0:             S(dIndex)   = S(dIndex) - sum(TS, 1);
wolffd@0:         end
wolffd@0:         M   = M / batchSize;
wolffd@0:     end
wolffd@0: 
wolffd@0:     dPsi    = CPGRADIENT(X, S, batchSize);
wolffd@0: 
wolffd@0: end