Daniel@0: function [X, C, idx] = get_itml_deltas(r, in)
Daniel@0: % [X, C, idx] = get_itml_deltas(r, in)
Daniel@0: 
Daniel@0: %ITML Specs
Daniel@0: % C: 4 column matrix
Daniel@0: %      column 1, 2: index of constrained points.  Indexes between 1 and n
Daniel@0: %      column 3: 1 if points are similar, -1 if dissimilar
Daniel@0: %      column 4: right-hand side (lower or upper bound, depending on 
Daniel@0: %                   whether points are similar or dissimilar)
Daniel@0: %
Daniel@0: % X: (n x m) data matrix - each row corresponds to a single instance
Daniel@0: % ---
Daniel@0: % NOTE: X is thus input in transposed shape for the ITML algorithm
Daniel@0: % ---
Daniel@0: 
Daniel@0: % ---
Daniel@0: % NOTE: this preallocation is not complete
Daniel@0: % ---
Daniel@0: X = zeros(size(in,1), 0);
Daniel@0: C = zeros(0,4);
Daniel@0: idx = zeros(0,2);
Daniel@0: 
Daniel@0: for i = 1:size(r,1)
Daniel@0:     
Daniel@0:     % feature indexing
Daniel@0:     a = i;
Daniel@0:     
Daniel@0:     % check if ranking is valid
Daniel@0:     if ~isempty(r{i,1}) && ~isempty(r{i,2})&& ...
Daniel@0:         isempty(intersect(r{i,1}, r{i,2}));
Daniel@0:     
Daniel@0:         % ---
Daniel@0:         % NOTE / TODO: the follwing is intended for compability
Daniel@0:         %  both sides of the ranking may have more than one entry.
Daniel@0:         %  for the MTT database, the ranking may be correct, but the 
Daniel@0:         %  inequalities build from non-singular rankings are not
Daniel@0:         %  based on the actual data
Daniel@0:         % ---
Daniel@0:         for j = 1:numel(r{i,1})
Daniel@0:             b = r{i,1}(j);
Daniel@0:             
Daniel@0:             for k = 1:numel(r{i,2})
Daniel@0:                 c = r{i,2}(k);
Daniel@0: 
Daniel@0:                 % ---
Daniel@0:                 % get vector deltas
Daniel@0:                 % ---
Daniel@0:                 [dab] = in(:,a) - in(:,b);
Daniel@0:                 [dac] = in(:,a) - in(:,c);
Daniel@0:                 
Daniel@0:                 % ---
Daniel@0:                 % save deltas in new feature matrix
Daniel@0:                 % TODO: this method has duplicate entries
Daniel@0:                 %  if the pairs appear more than once
Daniel@0:                 %  index the data set and use more efficiently!!!
Daniel@0:                 % ---
Daniel@0:                 X = [X dab];
Daniel@0:                 idx(end+1,:) = [a b];
Daniel@0:                 iab = size(idx, 1);
Daniel@0:                 
Daniel@0:                 X = [X dac];
Daniel@0:                 idx(end+1,:) = [a c];
Daniel@0:                 iac = size(idx, 1);
Daniel@0:                 
Daniel@0:                 % ---
Daniel@0:                 % NOTE:
Daniel@0:                 % in terms of the constraint,
Daniel@0:                 %  this should mean: dac - dab >= 1
Daniel@0:                 %
Daniel@0:                 % 4th position cannot be 0, converges to Inf if > 1
Daniel@0:                 % -1,-1 learns the opposite of what constraitns say
Daniel@0:                 % ---
Daniel@0:                 C(end+1, :) = [iab iac -1 -1];
Daniel@0:             end
Daniel@0:         end
Daniel@0:     end
Daniel@0: end
Daniel@0: 
Daniel@0: % % ---
Daniel@0: % % NOTE: here, we transpose the X for usage i nthe training
Daniel@0: % % ---
Daniel@0: % X = X';
Daniel@0: 
Daniel@0: end
Daniel@0: