1 | function [B, ll] = loglc2(A, W) |
---|
2 | %LOGLC2 Logistic Linear Classifier |
---|
3 | % |
---|
4 | % W = LOGLC2(A, L) |
---|
5 | % |
---|
6 | % INPUT |
---|
7 | % A Dataset |
---|
8 | % L Regularization parameter (L2) |
---|
9 | % |
---|
10 | % OUTPUT |
---|
11 | % W Logistic linear classifier |
---|
12 | % |
---|
13 | % DESCRIPTION |
---|
14 | % Computation of the linear classifier for the dataset A by maximizing the |
---|
15 | % L2-regularized likelihood criterion using the logistic (sigmoid) function. |
---|
16 | % The default value for L is 0. |
---|
17 | % |
---|
18 | % |
---|
19 | % SEE ALSO |
---|
20 | % MAPPINGS, DATASETS, LDC, FISHERC |
---|
21 | |
---|
22 | |
---|
23 | name = 'Logistic regressor (implementation 2)'; |
---|
24 | addpath(genpath('minFunc')); |
---|
25 | addpath(genpath('~/prtools')); |
---|
26 | if exist('minFunc', 'file') |
---|
27 | prwarning(1, 'LOGLC2 requires the minFunc optimizer. Please download it from www.di.ens.fr/~mschmidt/Software/minFunc.html and add it to the Matlab path.'); |
---|
28 | end |
---|
29 | |
---|
30 | % Handle untrained calls like W = loglc2([]); |
---|
31 | if nargin == 0 || isempty(A) |
---|
32 | B = mapping(mfilename); |
---|
33 | B = setname(B, name); |
---|
34 | return; |
---|
35 | |
---|
36 | % Handle training on dataset A (use A * loglc2, A * loglc2([]), and loglc2(A)) |
---|
37 | elseif (nargin == 1 && isa(A, 'dataset')) || (isa(A, 'dataset') && isa(W, 'double')) |
---|
38 | if nargin < 2 |
---|
39 | W = 0; |
---|
40 | end |
---|
41 | islabtype(A, 'crisp'); |
---|
42 | isvaldfile(A, 1, 2); |
---|
43 | A = testdatasize(A, 'features'); |
---|
44 | A = setprior(A, getprior(A)); |
---|
45 | [~, k, c] = getsize(A); |
---|
46 | |
---|
47 | % Train the logistic regressor |
---|
48 | [data.E, data.E_bias] = train_logreg(+A', getnlab(A)', W); |
---|
49 | B = mapping(mfilename, 'trained', data, getlablist(A), k, c); |
---|
50 | B = setname(B, name); |
---|
51 | |
---|
52 | % Handle evaluation of a trained LOGLC2 W for a dataset A |
---|
53 | elseif isa(A, 'dataset') && isa(W, 'mapping') |
---|
54 | |
---|
55 | % Evaluate logistic classifier |
---|
56 | [~, test_post] = eval_logreg(+A', W.data.E, W.data.E_bias); |
---|
57 | A = dataset(A); |
---|
58 | B = setdata(A, test_post', getlabels(W)); |
---|
59 | ll = []; |
---|
60 | |
---|
61 | % This should not happen |
---|
62 | else |
---|
63 | error('Illegal call'); |
---|
64 | end |
---|
65 | end |
---|
66 | |
---|
67 | function [E, E_bias] = train_logreg(train_X, train_labels, lambda, E_init, E_bias_init) |
---|
68 | |
---|
69 | % Uses fancy optimizer |
---|
70 | addpath(genpath('minFunc')); |
---|
71 | |
---|
72 | % Initialize solution |
---|
73 | if ~iscell(train_X) |
---|
74 | D = size(train_X, 1); |
---|
75 | else |
---|
76 | D = 0; |
---|
77 | for i=1:length(train_X) |
---|
78 | D = max(D, max(train_X{i})); |
---|
79 | end |
---|
80 | end |
---|
81 | [lablist, foo, train_labels] = unique(train_labels); |
---|
82 | K = length(lablist); |
---|
83 | if ~exist('E_init', 'var') || isempty(E_init) |
---|
84 | E = randn(D, K) * .0001; |
---|
85 | else |
---|
86 | E = E_init; clear E_init |
---|
87 | end |
---|
88 | if ~exist('E_bias_init', 'var') || isempty(E_bias_init) |
---|
89 | E_bias = zeros(1, K); |
---|
90 | else |
---|
91 | E_bias = E_bias_init; clear E_bias_init; |
---|
92 | end |
---|
93 | |
---|
94 | % Compute positive part of gradient |
---|
95 | pos_E = zeros(D, K); |
---|
96 | pos_E_bias = zeros(1, K); |
---|
97 | if ~iscell(train_X) |
---|
98 | for k=1:K |
---|
99 | pos_E(:,k) = sum(train_X(:,train_labels == k), 2); |
---|
100 | end |
---|
101 | else |
---|
102 | for i=1:length(train_X) |
---|
103 | pos_E(train_X{i}, train_labels(i)) = pos_E(train_X{i}, train_labels(i)) + 1; |
---|
104 | end |
---|
105 | end |
---|
106 | for k=1:K |
---|
107 | pos_E_bias(k) = sum(train_labels == k); |
---|
108 | end |
---|
109 | |
---|
110 | % Perform learning using L-BFGS |
---|
111 | x = [E(:); E_bias(:)]; |
---|
112 | options.Method = 'lbfgs'; |
---|
113 | options.Display = 'on'; |
---|
114 | options.TolFun = 1e-4; |
---|
115 | options.TolX = 1e-4; |
---|
116 | options.MaxIter = 5000; |
---|
117 | if ~iscell(train_X) |
---|
118 | x = minFunc(@logreg_grad, x, options, train_X, train_labels, lambda, pos_E, pos_E_bias); |
---|
119 | else |
---|
120 | x = minFunc(@logreg_discrete_grad, x, options, train_X, train_labels, lambda, pos_E, pos_E_bias); |
---|
121 | end |
---|
122 | |
---|
123 | % Decode solution |
---|
124 | E = reshape(x(1:D * K), [D K]); |
---|
125 | E_bias = reshape(x(D * K + 1:end), [1 K]); |
---|
126 | end |
---|
127 | |
---|
128 | |
---|
129 | function [est_labels, posterior] = eval_logreg(test_X, E, E_bias) |
---|
130 | |
---|
131 | % Perform labeling |
---|
132 | if ~iscell(test_X) |
---|
133 | log_Pyx = bsxfun(@plus, E' * test_X, E_bias'); |
---|
134 | else |
---|
135 | log_Pyx = zeros(length(E_bias), length(test_X)); |
---|
136 | for i=1:length(test_X) |
---|
137 | for j=1:length(test_X{i}) |
---|
138 | log_Pyx(:,i) = log_Pyx(:,i) + sum(E(test_X{i}{j},:), 1)'; |
---|
139 | end |
---|
140 | end |
---|
141 | log_Pyx = bsxfun(@plus, log_Pyx, E_bias'); |
---|
142 | end |
---|
143 | [~, est_labels] = max(log_Pyx, [], 1); |
---|
144 | |
---|
145 | % Compute posterior |
---|
146 | if nargout > 1 |
---|
147 | posterior = exp(bsxfun(@minus, log_Pyx, max(log_Pyx, [], 1))); |
---|
148 | posterior = bsxfun(@rdivide, posterior, sum(posterior, 1)); |
---|
149 | end |
---|
150 | end |
---|
151 | |
---|
152 | |
---|
153 | function [C, dC] = logreg_grad(x, train_X, train_labels, lambda, pos_E, pos_E_bias) |
---|
154 | %LOGREG_GRAD Gradient of L2-regularized logistic regressor |
---|
155 | % |
---|
156 | % [C, dC] = logreg_grad(x, train_X, train_labels, lambda, pos_E, pos_E_bias) |
---|
157 | % |
---|
158 | % Gradient of L2-regularized logistic regressor. |
---|
159 | |
---|
160 | |
---|
161 | % Decode solution |
---|
162 | [D, N] = size(train_X); |
---|
163 | K = numel(x) / (D + 1); |
---|
164 | E = reshape(x(1:D * K), [D K]); |
---|
165 | E_bias = reshape(x(D * K + 1:end), [1 K]); |
---|
166 | |
---|
167 | % Compute p(y|x) |
---|
168 | gamma = bsxfun(@plus, E' * train_X, E_bias'); |
---|
169 | gamma = exp(bsxfun(@minus, gamma, max(gamma, [], 1))); |
---|
170 | gamma = bsxfun(@rdivide, gamma, max(sum(gamma, 1), realmin)); |
---|
171 | |
---|
172 | % Compute conditional log-likelihood |
---|
173 | C = 0; |
---|
174 | for n=1:N |
---|
175 | C = C - log(max(gamma(train_labels(n), n), realmin)); |
---|
176 | end |
---|
177 | C = C + lambda .* sum(x .^ 2); |
---|
178 | |
---|
179 | % Only compute gradient when required |
---|
180 | if nargout > 1 |
---|
181 | |
---|
182 | % Compute positive part of gradient |
---|
183 | if ~exist('pos_E', 'var') || isempty(pos_E) |
---|
184 | pos_E = zeros(D, K); |
---|
185 | for k=1:K |
---|
186 | pos_E(:,k) = sum(train_X(:,train_labels == k), 2); |
---|
187 | end |
---|
188 | end |
---|
189 | if ~exist('pos_E_bias', 'var') || isempty(pos_E_bias) |
---|
190 | pos_E_bias = zeros(1, K); |
---|
191 | for k=1:K |
---|
192 | pos_E_bias(k) = sum(train_labels == k); |
---|
193 | end |
---|
194 | end |
---|
195 | |
---|
196 | % Compute negative part of gradient |
---|
197 | neg_E = train_X * gamma'; |
---|
198 | neg_E_bias = sum(gamma, 2)'; |
---|
199 | |
---|
200 | % Compute gradient |
---|
201 | dC = -[pos_E(:) - neg_E(:); pos_E_bias(:) - neg_E_bias(:)] + 2 .* lambda .* x; |
---|
202 | end |
---|
203 | end |
---|