comparison toolboxes/distance_learning/mlr/util/rmlr_admm.m @ 0:e9a9cd732c1e tip

first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
comparison
equal deleted inserted replaced
-1:000000000000 0:e9a9cd732c1e
1 function [W, Xi, Diagnostics] = rmlr_admm(C, K, Delta, H, Q, lam)
2 % [W, Xi, D] = mlr_admm(C, Delta, W, X)
3 %
4 % C >= 0 Slack trade-off parameter
5 % K = data matrix (or kernel)
6 % Delta = array of mean margin values
7 % H = structural kernel matrix
8 % Q = kernel-structure interaction vector
9 %
10 % W (output) = the learned metric
11 % Xi = 1-slack
12 % D = diagnostics
13
14 global DEBUG REG FEASIBLE LOSS INIT STRUCTKERNEL DUALW THRESH;
15
16 %%%
17 % Initialize the gradient directions for each constraint
18 %
19 global PsiR;
20
21 global ADMM_Z ADMM_V ADMM_UW ADMM_UV;
22
23 global ADMM_STEPS;
24
25 global RHO;
26
27 numConstraints = length(PsiR);
28
29 Diagnostics = struct( 'f', [], ...
30 'num_steps', [], ...
31 'stop_criteria', []);
32
33
34 % Convergence settings
35 if ~isempty(ADMM_STEPS)
36 MAX_ITER = ADMM_STEPS;
37 else
38 MAX_ITER = 10;
39 end
40 ABSTOL = 1e-4 * sqrt(numel(ADMM_Z));
41 RELTOL = 1e-3;
42 SCALE_THRESH = 10;
43 RHO_RESCALE = 2;
44 stopcriteria= 'MAX STEPS';
45
46 % Objective function
47 F = zeros(1,MAX_ITER);
48
49 % how many constraints
50
51 alpha = zeros(numConstraints, 1);
52 Gamma = zeros(numConstraints, 1);
53
54 ln1 = 0;
55 ln2 = 0;
56
57 % figure(2)
58 % hold off
59 % plot(0)
60 % delete(abc)
61 % delete(abc2)
62 for step = 1:MAX_ITER
63 % do a w-update
64 % dubstep needs:
65 % C <-- static
66 % RHO <-- static
67 % H <-- static
68 % Q <-- static
69 % Delta <-- static
70 % Gamma <-- this one's dynamic
71
72 for i = 1:numConstraints
73 Gamma(i) = STRUCTKERNEL(ADMM_Z-ADMM_UW, PsiR{i});
74 end
75 % d = length(K);
76 alpha = mlr_dual(C, RHO, H, Q, Delta, Gamma, alpha);
77
78 %%%
79 % 3) convert back to W
80 %
81 W = DUALW(alpha, ADMM_Z, ADMM_UW, RHO, K);
82
83 % figure(1), imagesc(W), drawnow;
84
85 % Update V
86 ADMM_V = THRESH(ADMM_Z - ADMM_UV, lam/RHO);
87
88 % Update Z
89 Zold = ADMM_Z;
90 ADMM_Z = FEASIBLE(0.5* (W + ADMM_V + ADMM_UW + ADMM_UV));
91
92 % Update residuals
93 ADMM_UW = ADMM_UW + W - ADMM_Z;
94 ADMM_UV = ADMM_UV + ADMM_V - ADMM_Z;
95
96 % Compute primal objective
97 % slack term
98 Xi = 0;
99 for R = numConstraints:-1:1
100 Xi = max(Xi, LOSS(ADMM_Z, PsiR{R}, Delta(R), 0));
101 end
102 F(step) = C * Xi + REG(W, K, 0) + lam * sum(sqrt(sum(W.^2)));
103
104 % figure(2), loglog(1:step, F(1:step)), xlim([0, MAX_ITER]), drawnow;
105 % Test for convergence
106
107 %WIP
108 N1 = norm(ADMM_V(:) + W(:) - 2* ADMM_Z(:));
109 N2 = RHO * norm(2* (Zold(:) - ADMM_Z(:)));
110
111 eps_primal = ABSTOL + RELTOL * max(norm(W(:)), norm(ADMM_Z(:)));
112 eps_dual = ABSTOL + RELTOL * RHO * norm(ADMM_UW(:));
113 %end WIP
114
115
116 % figure(2), loglog(step + (-1:0), [ln1, N1/eps_primal], 'b'), xlim([0, MAX_ITER]), hold('on');
117 % figure(2), loglog(step + (-1:0), [ln2, N2/eps_dual], 'r-'), xlim([0, MAX_ITER]), hold('on'), drawnow;
118 % ln1 = N1/eps_primal;
119 % ln2 = N2/eps_dual;
120
121 if N1 < eps_primal && N2 < eps_dual
122 stopcriteria = 'CONVERGENCE';
123 break;
124 end
125
126 if N1 > SCALE_THRESH * N2
127 dbprint(3, sprintf('RHO: %.2e UP %.2e', RHO, RHO * RHO_RESCALE));
128 RHO = RHO * RHO_RESCALE;
129 ADMM_UW = ADMM_UW / RHO_RESCALE;
130 elseif N2 > SCALE_THRESH * N1
131 dbprint(3, sprintf('RHO: %.2e DN %.2e', RHO, RHO / RHO_RESCALE));
132 RHO = RHO / RHO_RESCALE;
133 ADMM_UW = ADMM_UW * RHO_RESCALE;
134 end
135 end
136 % figure(2), hold('off');
137
138 %%%
139 % Ensure feasibility
140 %
141 W = FEASIBLE(W);
142
143
144 %%%
145 % Compute the slack
146 %
147 Xi = 0;
148 for R = numConstraints:-1:1
149 Xi = max(Xi, LOSS(W, PsiR{R}, Delta(R), 0));
150 end
151
152 %%%
153 % Update diagnostics
154 %
155
156 Diagnostics.f = F(1:step)';
157 Diagnostics.stop_criteria = stopcriteria;
158 Diagnostics.num_steps = step;
159
160 dbprint(1, '\t%s after %d steps.\n', stopcriteria, step);
161 end
162
163 function alpha = mlr_dual(C, RHO, H, Q, Delta, Gamma, alpha)
164
165 global PsiClock;
166
167 m = length(Delta);
168
169 if nargin < 7
170 alpha = zeros(m,1);
171 end
172
173 %%%
174 % 1) construct the QP parameters
175 %
176 b = RHO * (Gamma - Delta) - Q;
177
178 %%%
179 % 2) solve the QP
180 %
181 alpha = qplcprog(H, b, ones(1, m), C, [], [], 0, []);
182
183 %%%
184 % 3) update the Psi clock
185 %
186 PsiClock(alpha > 0) = 0;
187
188 end