|
wolffd@0
|
1 function [mu, N, pick] = parzen_fit_select_unif(data, labels, max_proto, varargin)
|
|
wolffd@0
|
2 % PARZEN_FIT_SELECT_UNIF Fit a parzen density estimator by selecting prototypes uniformly from data
|
|
wolffd@0
|
3 % [mu, N, pick] = parzen_fit_select_unif(data, max_proto, labels, ...)
|
|
wolffd@0
|
4 %
|
|
wolffd@0
|
5 % We partition the data into different subsets based on the labels.
|
|
wolffd@0
|
6 % We then choose up to max_proto columns from each subset, chosen uniformly.
|
|
wolffd@0
|
7 %
|
|
wolffd@0
|
8 % INPUTS
|
|
wolffd@0
|
9 % data(:,t)
|
|
wolffd@0
|
10 % labels(t) - should be in {1,2,..,Q}
|
|
wolffd@0
|
11 % max_proto - max number of prototypes per partition
|
|
wolffd@0
|
12 %
|
|
wolffd@0
|
13 % Optional args
|
|
wolffd@0
|
14 % partition_names{m} - for debugging
|
|
wolffd@0
|
15 % boundary - do not choose prototypes which are within 'boundary' of the label transition
|
|
wolffd@0
|
16 %
|
|
wolffd@0
|
17 % OUTPUTS
|
|
wolffd@0
|
18 % mu(:, m, q) for label q, prototype m for 1 <= m <= N(q)
|
|
wolffd@0
|
19 % N(q) = number of prototypes for label q
|
|
wolffd@0
|
20 % pick{q} = identity of the prototypes
|
|
wolffd@0
|
21
|
|
wolffd@0
|
22 nclasses = max(labels);
|
|
wolffd@0
|
23 [boundary, partition_names] = process_options(...
|
|
wolffd@0
|
24 varargin, 'boundary', 0, 'partition_names', []);
|
|
wolffd@0
|
25
|
|
wolffd@0
|
26 [D T] = size(data);
|
|
wolffd@0
|
27 mu = zeros(D, 1, nclasses); % dynamically determine num prototypes (may be less than K)
|
|
wolffd@0
|
28 mean_feat = mean(data,2);
|
|
wolffd@0
|
29 pick = cell(1,nclasses);
|
|
wolffd@0
|
30 for c=1:nclasses
|
|
wolffd@0
|
31 ndx = find(labels==c);
|
|
wolffd@0
|
32 if isempty(ndx)
|
|
wolffd@0
|
33 %fprintf('no training images have label %d (%s)\n', c, partition_names{c})
|
|
wolffd@0
|
34 fprintf('no training images have label %d\n', c);
|
|
wolffd@0
|
35 nviews = 1;
|
|
wolffd@0
|
36 mu(:,1,c) = mean_feat;
|
|
wolffd@0
|
37 else
|
|
wolffd@0
|
38 foo = linspace(boundary+1, length(ndx-boundary), max_proto);
|
|
wolffd@0
|
39 pick{c} = ndx(unique(floor(foo)));
|
|
wolffd@0
|
40 nviews = length(pick{c});
|
|
wolffd@0
|
41 %fprintf('picking %d views for class %d=%s\n', nviews, c, class_names{c});
|
|
wolffd@0
|
42 mu(:,1:nviews,c) = data(:, pick{c});
|
|
wolffd@0
|
43 end
|
|
wolffd@0
|
44 N(c) = nviews;
|
|
wolffd@0
|
45 end
|
