|
wolffd@0
|
1 function [y, l] = knnfwd(net, x)
|
|
wolffd@0
|
2 %KNNFWD Forward propagation through a K-nearest-neighbour classifier.
|
|
wolffd@0
|
3 %
|
|
wolffd@0
|
4 % Description
|
|
wolffd@0
|
5 % [Y, L] = KNNFWD(NET, X) takes a matrix X of input vectors (one vector
|
|
wolffd@0
|
6 % per row) and uses the K-nearest-neighbour rule on the training data
|
|
wolffd@0
|
7 % contained in NET to produce a matrix Y of outputs and a matrix L of
|
|
wolffd@0
|
8 % classification labels. The nearest neighbours are determined using
|
|
wolffd@0
|
9 % Euclidean distance. The IJth entry of Y counts the number of
|
|
wolffd@0
|
10 % occurrences that an example from class J is among the K closest
|
|
wolffd@0
|
11 % training examples to example I from X. The matrix L contains the
|
|
wolffd@0
|
12 % predicted class labels as an index 1..N, not as 1-of-N coding.
|
|
wolffd@0
|
13 %
|
|
wolffd@0
|
14 % See also
|
|
wolffd@0
|
15 % KMEANS, KNN
|
|
wolffd@0
|
16 %
|
|
wolffd@0
|
17
|
|
wolffd@0
|
18 % Copyright (c) Ian T Nabney (1996-2001)
|
|
wolffd@0
|
19
|
|
wolffd@0
|
20
|
|
wolffd@0
|
21 errstring = consist(net, 'knn', x);
|
|
wolffd@0
|
22 if ~isempty(errstring)
|
|
wolffd@0
|
23 error(errstring);
|
|
wolffd@0
|
24 end
|
|
wolffd@0
|
25
|
|
wolffd@0
|
26 ntest = size(x, 1); % Number of input vectors.
|
|
wolffd@0
|
27 nclass = size(net.tr_targets, 2); % Number of classes.
|
|
wolffd@0
|
28
|
|
wolffd@0
|
29 % Compute matrix of squared distances between input vectors from the training
|
|
wolffd@0
|
30 % and test sets. The matrix distsq has dimensions (ntrain, ntest).
|
|
wolffd@0
|
31
|
|
wolffd@0
|
32 distsq = dist2(net.tr_in, x);
|
|
wolffd@0
|
33
|
|
wolffd@0
|
34 % Now sort the distances. This generates a matrix kind of the same
|
|
wolffd@0
|
35 % dimensions as distsq, in which each column gives the indices of the
|
|
wolffd@0
|
36 % elements in the corresponding column of distsq in ascending order.
|
|
wolffd@0
|
37
|
|
wolffd@0
|
38 [vals, kind] = sort(distsq);
|
|
wolffd@0
|
39 y = zeros(ntest, nclass);
|
|
wolffd@0
|
40
|
|
wolffd@0
|
41 for k=1:net.k
|
|
wolffd@0
|
42 % We now look at the predictions made by the Kth nearest neighbours alone,
|
|
wolffd@0
|
43 % and represent this as a 1-of-N coded matrix, and then accumulate the
|
|
wolffd@0
|
44 % predictions so far.
|
|
wolffd@0
|
45
|
|
wolffd@0
|
46 y = y + net.tr_targets(kind(k,:),:);
|
|
wolffd@0
|
47
|
|
wolffd@0
|
48 end
|
|
wolffd@0
|
49
|
|
wolffd@0
|
50 if nargout == 2
|
|
wolffd@0
|
51 % Convert this set of outputs to labels, randomly breaking ties
|
|
wolffd@0
|
52 [temp, l] = max((y + 0.1*rand(size(y))), [], 2);
|
|
wolffd@0
|
53 end |
