|
wolffd@0
|
1 function [partBinTrn, partBinTst, partBinNoTrn] = sim_from_comparison_UNfair_components(comparison, comparison_ids, k, trainpart, filename)
|
|
wolffd@0
|
2 %
|
|
wolffd@0
|
3 % FOR ISMIR 2012
|
|
wolffd@0
|
4 %
|
|
wolffd@0
|
5 % creates a cross-validation partitioning of the
|
|
wolffd@0
|
6 % similarity data in "multiG", NOT PRESERVING the
|
|
wolffd@0
|
7 % connected components in it during partitioning
|
|
wolffd@0
|
8
|
|
wolffd@0
|
9 % ---
|
|
wolffd@0
|
10 % get the similarity multigraph and remove cycles
|
|
wolffd@0
|
11 % ---
|
|
wolffd@0
|
12 cprint(2, 'creating graph')
|
|
wolffd@0
|
13 % Gm = ClipSimGraphMulti(comparison, comparison_ids);
|
|
wolffd@0
|
14 % Gm.remove_cycles_length2;
|
|
wolffd@0
|
15 cprint(2, 'loading Multigraph for Similarity Constraints')
|
|
wolffd@0
|
16 load('comp_SimGraphMulti.mat', 'G');
|
|
wolffd@0
|
17
|
|
wolffd@0
|
18
|
|
wolffd@0
|
19 % get similarity data
|
|
wolffd@0
|
20 [weights, a, b, c] = G.similarities();
|
|
wolffd@0
|
21
|
|
wolffd@0
|
22 % ---
|
|
wolffd@0
|
23 % We randomise the constraint succession
|
|
wolffd@0
|
24 % ---
|
|
wolffd@0
|
25 datPermu = randperm(numel(a));
|
|
wolffd@0
|
26 a = a(datPermu);
|
|
wolffd@0
|
27 b = b(datPermu);
|
|
wolffd@0
|
28 c = c(datPermu);
|
|
wolffd@0
|
29 weights = weights(datPermu);
|
|
wolffd@0
|
30
|
|
wolffd@0
|
31 % ---
|
|
wolffd@0
|
32 % NOTE: we try the easy route: partition the graphs
|
|
wolffd@0
|
33 % and look at which constraints balance we end up with
|
|
wolffd@0
|
34 % ---
|
|
wolffd@0
|
35 P = cvpartition(numel(a), 'k', k);
|
|
wolffd@0
|
36
|
|
wolffd@0
|
37 % ---
|
|
wolffd@0
|
38 % here we export similarity test sets
|
|
wolffd@0
|
39 % ---
|
|
wolffd@0
|
40 cprint(2, 'export test similarity')
|
|
wolffd@0
|
41 partBinTst = {};
|
|
wolffd@0
|
42 for i = 1:P.NumTestSets % test runs
|
|
wolffd@0
|
43
|
|
wolffd@0
|
44 partBinTst{i} = [a(P.test(i))' b(P.test(i))' c(P.test(i))' weights(P.test(i))];
|
|
wolffd@0
|
45 end
|
|
wolffd@0
|
46
|
|
wolffd@0
|
47
|
|
wolffd@0
|
48 % ---
|
|
wolffd@0
|
49 % Note: This uses a "truly" increasing training set
|
|
wolffd@0
|
50 % to do the partial training partition
|
|
wolffd@0
|
51 % ---
|
|
wolffd@0
|
52 cprint(2, 'export train similarity')
|
|
wolffd@0
|
53 for m = 1:numel(trainpart)
|
|
wolffd@0
|
54
|
|
wolffd@0
|
55 % save test indices
|
|
wolffd@0
|
56 Ptrain(m) = cvpartition_trunctrain_incsubsets(P, trainpart(m));
|
|
wolffd@0
|
57 end
|
|
wolffd@0
|
58
|
|
wolffd@0
|
59 % ---
|
|
wolffd@0
|
60 % here we export similarity training sets
|
|
wolffd@0
|
61 % ---
|
|
wolffd@0
|
62 partBinTrn = {};
|
|
wolffd@0
|
63 for i = 1:P.NumTestSets % train runs
|
|
wolffd@0
|
64
|
|
wolffd@0
|
65 for m = 1:numel(trainpart) % increasing training sets
|
|
wolffd@0
|
66
|
|
wolffd@0
|
67 % get training indices
|
|
wolffd@0
|
68 idxB = Ptrain(m).training(i);
|
|
wolffd@0
|
69
|
|
wolffd@0
|
70 % save into cell
|
|
wolffd@0
|
71 partBinTrn{i,m} = [a(idxB)' b(idxB)' c(idxB)' weights(idxB)];
|
|
wolffd@0
|
72 end
|
|
wolffd@0
|
73 end
|
|
wolffd@0
|
74
|
|
wolffd@0
|
75 partBinNoTrn = {};
|
|
wolffd@0
|
76 for i = 1:P.NumTestSets % train runs
|
|
wolffd@0
|
77
|
|
wolffd@0
|
78 for m = 1:numel(trainpart) % increasing training sets
|
|
wolffd@0
|
79
|
|
wolffd@0
|
80 % get training indices
|
|
wolffd@0
|
81 idxB = ~Ptrain(m).training(i) & ~Ptrain(m).test(i);
|
|
wolffd@0
|
82
|
|
wolffd@0
|
83 % save into cell
|
|
wolffd@0
|
84 partBinNoTrn{i,m} = [a(idxB)' b(idxB)' c(idxB)' weights(idxB)];
|
|
wolffd@0
|
85 end
|
|
wolffd@0
|
86 end
|
|
wolffd@0
|
87
|
|
wolffd@0
|
88 if nargin == 5
|
|
wolffd@0
|
89 save(filename, 'partBinTrn', 'partBinTst', 'partBinNoTrn')
|
|
wolffd@0
|
90 end
|
|
wolffd@0
|
91 end |
