|
wolffd@0
|
1 function [X, C, idx] = get_itml_deltas(r, in)
|
|
wolffd@0
|
2 % [X, C, idx] = get_itml_deltas(r, in)
|
|
wolffd@0
|
3
|
|
wolffd@0
|
4 %ITML Specs
|
|
wolffd@0
|
5 % C: 4 column matrix
|
|
wolffd@0
|
6 % column 1, 2: index of constrained points. Indexes between 1 and n
|
|
wolffd@0
|
7 % column 3: 1 if points are similar, -1 if dissimilar
|
|
wolffd@0
|
8 % column 4: right-hand side (lower or upper bound, depending on
|
|
wolffd@0
|
9 % whether points are similar or dissimilar)
|
|
wolffd@0
|
10 %
|
|
wolffd@0
|
11 % X: (n x m) data matrix - each row corresponds to a single instance
|
|
wolffd@0
|
12 % ---
|
|
wolffd@0
|
13 % NOTE: X is thus input in transposed shape for the ITML algorithm
|
|
wolffd@0
|
14 % ---
|
|
wolffd@0
|
15
|
|
wolffd@0
|
16 % ---
|
|
wolffd@0
|
17 % NOTE: this preallocation is not complete
|
|
wolffd@0
|
18 % ---
|
|
wolffd@0
|
19 X = zeros(size(in,1), 0);
|
|
wolffd@0
|
20 C = zeros(0,4);
|
|
wolffd@0
|
21 idx = zeros(0,2);
|
|
wolffd@0
|
22
|
|
wolffd@0
|
23 for i = 1:size(r,1)
|
|
wolffd@0
|
24
|
|
wolffd@0
|
25 % feature indexing
|
|
wolffd@0
|
26 a = i;
|
|
wolffd@0
|
27
|
|
wolffd@0
|
28 % check if ranking is valid
|
|
wolffd@0
|
29 if ~isempty(r{i,1}) && ~isempty(r{i,2})&& ...
|
|
wolffd@0
|
30 isempty(intersect(r{i,1}, r{i,2}));
|
|
wolffd@0
|
31
|
|
wolffd@0
|
32 % ---
|
|
wolffd@0
|
33 % NOTE / TODO: the follwing is intended for compability
|
|
wolffd@0
|
34 % both sides of the ranking may have more than one entry.
|
|
wolffd@0
|
35 % for the MTT database, the ranking may be correct, but the
|
|
wolffd@0
|
36 % inequalities build from non-singular rankings are not
|
|
wolffd@0
|
37 % based on the actual data
|
|
wolffd@0
|
38 % ---
|
|
wolffd@0
|
39 for j = 1:numel(r{i,1})
|
|
wolffd@0
|
40 b = r{i,1}(j);
|
|
wolffd@0
|
41
|
|
wolffd@0
|
42 for k = 1:numel(r{i,2})
|
|
wolffd@0
|
43 c = r{i,2}(k);
|
|
wolffd@0
|
44
|
|
wolffd@0
|
45 % ---
|
|
wolffd@0
|
46 % get vector deltas
|
|
wolffd@0
|
47 % ---
|
|
wolffd@0
|
48 [dab] = in(:,a) - in(:,b);
|
|
wolffd@0
|
49 [dac] = in(:,a) - in(:,c);
|
|
wolffd@0
|
50
|
|
wolffd@0
|
51 % ---
|
|
wolffd@0
|
52 % save deltas in new feature matrix
|
|
wolffd@0
|
53 % TODO: this method has duplicate entries
|
|
wolffd@0
|
54 % if the pairs appear more than once
|
|
wolffd@0
|
55 % index the data set and use more efficiently!!!
|
|
wolffd@0
|
56 % ---
|
|
wolffd@0
|
57 X = [X dab];
|
|
wolffd@0
|
58 idx(end+1,:) = [a b];
|
|
wolffd@0
|
59 iab = size(idx, 1);
|
|
wolffd@0
|
60
|
|
wolffd@0
|
61 X = [X dac];
|
|
wolffd@0
|
62 idx(end+1,:) = [a c];
|
|
wolffd@0
|
63 iac = size(idx, 1);
|
|
wolffd@0
|
64
|
|
wolffd@0
|
65 % ---
|
|
wolffd@0
|
66 % NOTE:
|
|
wolffd@0
|
67 % in terms of the constraint,
|
|
wolffd@0
|
68 % this should mean: dac - dab >= 1
|
|
wolffd@0
|
69 %
|
|
wolffd@0
|
70 % 4th position cannot be 0, converges to Inf if > 1
|
|
wolffd@0
|
71 % -1,-1 learns the opposite of what constraitns say
|
|
wolffd@0
|
72 % ---
|
|
wolffd@0
|
73 C(end+1, :) = [iab iac -1 -1];
|
|
wolffd@0
|
74 end
|
|
wolffd@0
|
75 end
|
|
wolffd@0
|
76 end
|
|
wolffd@0
|
77 end
|
|
wolffd@0
|
78
|
|
wolffd@0
|
79 % % ---
|
|
wolffd@0
|
80 % % NOTE: here, we transpose the X for usage i nthe training
|
|
wolffd@0
|
81 % % ---
|
|
wolffd@0
|
82 % X = X';
|
|
wolffd@0
|
83
|
|
wolffd@0
|
84 end
|
|
wolffd@0
|
85
|
