-
Notifications
You must be signed in to change notification settings - Fork 0
/
cpd_als.m
341 lines (320 loc) · 13.4 KB
/
cpd_als.m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
function [U,output] = cpd_als(T,U0,varargin)
%CPD_ALS CPD by alternating least squares.
% [U,OUTPUT] = CPD_ALS(T,U0) computes the factor matrices U{1}, ..., U{N}
% belonging to a canonical polyadic decomposition of the N-th order
% tensor T. The algorithm is initialized with the factor matrices U0{n}.
% The structure output returns additional information:
%
% output.alpha - The value of the line or plane search step
% length(s) in every iteration.
% output.fval - The value of the objective function
% 0.5*frob(T-cpdgen(U))^2 in every iteration.
% output.info - The circumstances under which the procedure
% terminated:
% 1: Objective function tolerance reached.
% 2: Step size tolerance reached.
% 3: Maximum number of iterations reached.
% output.iterations - The number of iterations.
% output.relfval - The difference in objective function value
% between every two successive iterates, relative
% to its initial value.
% output.relstep - The step size relative to the norm of the
% current iterate in every iteration.
%
% CPD_ALS(T,U0,OPTIONS) or CPD_ALS(T,U0,'key',value) may be used to set the
% following options:
%
% options.Display = 0 - Displays output information each
% options.Display iterations.
% options.LineSearch = - A function handle to the desired line
% [{false}|@cpd_aels|... search algorithm.
% @cpd_els|@cpd_lsb]
% options.LineSearchOptions - An options structure passed to the
% selected line search algorithm.
% options.PlaneSearch = - A function handle to the desired plane
% [{false}|@cpd_eps] search algorithm. Searches in the plane
% spanned by the last two ALS updates.
% options.PlaneSearchOptions - An options structure passed to the
% selected plane search algorithm.
% options.MaxIter = 500 - The maximum number of iterations.
% options.Order = 1:N - The order in which to update the factor
% matrices. Must be a permutation of 1:N.
% options.TolFun = 1e-8 - The tolerance for output.relfval. Note
% that because the objective function is
% a squared norm, TolFun can be as small
% as eps^2.
% options.TolX = 1e-6 - The tolerance for output.relstep.
% options.Delta = 1 - Step length restriction for update. If
% Delta = 1, no restriction is used.
% Authors: Laurent Sorber (Laurent.Sorber@cs.kuleuven.be)
% Nico Vervliet (Nico.Vervliet@esat.kuleuven.be)
% Marc Van Barel (Marc.VanBarel@cs.kuleuven.be)
% Lieven De Lathauwer (Lieven.DeLathauwer@kuleuven-kulak.be)
%
% Version History:
% - 2016/01/30 NV Support for structured tensors
% Format the tensor T.
type = getstructure(T);
isstructured = ~any(strcmp(type, {'full', 'incomplete', 'sparse'}));
if ~isstructured, T = fmt(T,true); end
% Check the initial factor matrices U0.
U = U0(:).'; U1 = {};
R = size(U{1},2);
N = length(U);
if any(cellfun('size',U,2) ~= R)
error('cpd_als:U0','size(U0{n},2) should be the same for all n.');
end
sz = cellfun('size', U, 1);
size_tens = getsize(T);
if length(sz) < length(size_tens)
error('cpd_als:U0','length(U0) should >= getorder(T).');
end
if any(sz(1:length(size_tens)) ~= size_tens)
error('cpd_als:U0','size(U0{n},1) should be size(T,n).');
end
if any(sz(length(size_tens)+1:end) ~= 1)
error('cpd_als:U0','size(U0{n},1) should be 1 for n > getorder(T).');
end
p = inputParser;
p.addOptional('Display', 0);
p.addOptional('FastUpdate', true);
p.addOptional('LineSearch', false);
p.addOptional('LineSearchOptions', struct);
p.addOptional('PlaneSearch', false);
p.addOptional('PlaneSearchOptions', struct);
p.addOptional('MaxIter', 500);
p.addOptional('Order', 1:N);
p.addOptional('TolFun', 1e-8);
p.addOptional('TolX', 1e-6);
p.addOptional('TolAbs', 0);
p.addOptional('Delta', nan);
p.KeepUnmatched = true;
p.parse(varargin{:});
options = p.Results;
% Check the options structure.
isfunc = @(f)isa(f,'function_handle');
xsfunc = @(f)isfunc(f)&&exist(func2str(f),'file');
if ~xsfunc(options.LineSearch) ...
&& (~isa(options.LineSearch,'function_handle') && options.LineSearch)
error('cpd_als:LineSearch','Not a valid line search algorithm.');
end
if ~xsfunc(options.PlaneSearch) ...
&& (~isa(options.PlaneSearch,'function_handle') && options.PlaneSearch)
error('cpd_als:PlaneSearch','Not a valid line search algorithm.');
end
if any(strcmpi(p.UsingDefaults, 'TolAbs')) && isstructured
options.TolAbs = 1e-15;
end
% Cache some intermediate variables.
try
T2 = frob(T,'squared');
catch e
if strcmpi(e.identifier, 'frob:notImplemented');
error('cpd_als:notImplemented', ...
['cpd_als does not support the structured tensor type %s, yet. Use ' ...
'ful(T) instead.'], type);
end
end
K = cell(1,N);
UHU = zeros(R,R,N);
for n = 1:N, UHU(:,:,n) = U{n}'*U{n}; end
% Alternating least squares.
first = options.Order(1);
last = options.Order(end);
K{first} = mtkrprod(T,U,first);
if isstructured,
D = 0.5*T2 - real(inprod(T,U)) + 0.5*frob(U,'squared');
else
D = cpdres(T,U);
D = 0.5*(D(:)'*D(:));
end
output.alpha = [];
output.fval = D;
output.info = false;
output.iterations = 0;
output.relgain = [];
output.relfval = [];
output.relstep = [];
while ~output.info
% Save current state.
U2 = U1;
U1 = U;
% Update factor matrices.
for n = options.Order
W = prod(UHU(:,:,[1:n-1 n+1:N]),3);
if n ~= first
K{n} = mtkrprod(T,U,n);
end
if isnan(options.Delta) || options.Delta == 1
U{n} = K{n}/conj(W);
else
U{n} = (1-options.Delta)*U{n} + options.Delta*K{n}/conj(W);
end
UHU(:,:,n) = U{n}'*U{n};
end
% Line/plane search.
[U,alpha,outputsrch] = search();
% Update the output structure.
output.alpha(:,end+1) = alpha;
if isfield(outputsrch,'relgain')
output.relgain(end+1) = outputsrch.relgain;
end
output.fval(end+1) = outputsrch.fval;
output.iterations = output.iterations+1;
output.relfval(end+1) = ...
abs(diff(output.fval(end:-1:end-1)))/abs(output.fval(1));
output.relstep(end+1) = sqrt(sum(cellfun(@(u,v)(u(:)-v(:))'* ...
(u(:)-v(:)),U,U1)))/sqrt(sum(cellfun(@(u)u(:)'*u(:),U)));
if output.relfval(end) <= options.TolFun, output.info = 1; end
if output.relstep(end) <= options.TolX, output.info = 2; end
if output.iterations >= options.MaxIter, output.info = 3; end
if 2*output.fval(end) <= options.TolAbs*T2, output.info = 4; end
% Display progress.
if options.Display > 0 && (output.iterations == 1 || output.info || ...
mod(output.iterations,options.Display) == 0)
if output.iterations == 1
bold = '%s';
[~,~,~,~,v] = regexp(version('-release'),'([0-9]+)([ab])');
if usejava('Desktop') && str2double(v{1}{1}) > 2011 || ...
(str2double(v{1}{1}) == 2011 && strcmpi(v{1}{2},'b'))
bold = '<strong>%s</strong>';
end
end
if output.iterations == 1 || ...
mod(output.iterations,15*options.Display) == 0
fprintf('\n%7s%s','',sprintf(bold,'fval'));
fprintf('%13s%s','',sprintf(bold,'relfval'));
fprintf('%10s%s','',sprintf(bold,'relstep'));
fprintf('%10s%s','',sprintf(bold,'alpha'));
fprintf('\n%30s = %4.e %6s = %4.e\n\n', ...
'TolFun',options.TolFun,'TolX',options.TolX);
end
if output.iterations == 1
fprintf('%4i: % 14.8e |\n',0,output.fval(1));
end
stralpha = [repmat('%10.4e ',1,size(output.alpha,1)) '\n'];
fprintf(['%4i: % 14.8e | %14.8e | %14.8e | ' stralpha], ...
output.iterations,output.fval(end), ...
output.relfval(end),output.relstep(end), ...
abs(output.alpha(:,end)));
end
end
% Display termination message.
if options.Display > 0
ahref = '\n%s\n\n';
x = round(linspace(0,output.iterations,min(500,output.iterations)));
if length(bold) > 2
ahref = sprintf(['\n<a href="matlab:semilogy(%s,%s);' ...
'xlabel(''iteration'');legend(''fval'',' ...
'''relfval'',''relstep'')">%%s</a>\n\n'],mat2str(x'), ...
mat2str([output.fval(x+1)' [nan output.relfval(x(2:end))]' ...
[nan output.relstep(x(2:end))]'],3));
end
switch output.info
case 1, fprintf(ahref,'Objective function tolerance reached.');
case 2, fprintf(ahref,'Step size tolerance reached.');
case 3, fprintf(ahref,'Maximum number of iterations reached.');
end
end
function [Ua,alpha,outputsrch,dU,dU1,dU2,D,n] = search()
% Line/plane search wrapper that guarantees the return values alpha and
% outputsrch.fval, and prepares the next ALS iteration.
% Attempt line/plane search, and fill in missing alpha.
outputsrch = struct;
state = output; state.T2 = T2; state.UHU = UHU;
if options.FastUpdate
state.K1 = first; state.KN = last; state.K = K;
end
if isfunc(options.LineSearch)
dU = cellfun(@(u,v)u-v,U,U1,'UniformOutput',false);
[alpha,outputsrch] = options.LineSearch(T,U1,dU, ...
state,options.LineSearchOptions);
if any(isnan(alpha)) || isempty(alpha), alpha = [1 1];
elseif length(alpha) == 1, alpha(2) = 1; end
elseif isfunc(options.PlaneSearch)
if output.iterations >= 1
dU1 = cellfun(@(u,v)u-v,U,U1,'UniformOutput',false);
dU2 = cellfun(@(u,v)u-v,U1,U2,'UniformOutput',false);
[alpha,outputsrch] = options.PlaneSearch(T,U1,dU1,dU2, ...
state,options.PlaneSearchOptions);
else
alpha = nan;
end
if any(isnan(alpha)) || isempty(alpha), alpha = [1 0 1];
elseif length(alpha) == 2, alpha(3) = 1; end
else
alpha = 1;
end
% Fill in missing fval.
if ~isfield(outputsrch,'fval')
if isfunc(options.LineSearch)
Ua = cellfun(@(u,v)alpha(2)*(u+alpha(1)*v),U1,dU, ...
'UniformOutput',false);
elseif isfunc(options.PlaneSearch) && output.iterations >= 1
Ua = cellfun(@(u,v,w)alpha(3)*(u+alpha(1)*v+alpha(2)*w), ...
U1,dU1,dU2,'UniformOutput',false);
else
Ua = U;
end
if options.FastUpdate
K1 = mtkrprod(T,Ua,first);
UHU1 = zeros(size(UHU));
for n = 1:N, UHU1(:,:,n) = Ua{n}'*Ua{n}; end
end
if options.FastUpdate && log10(output.fval(end)) > log10(T2)-16+2.5
outputsrch.fval = abs(.5*(T2+sum(sum(real(prod(UHU1,3)))))- ...
real(sum(dot(K1,Ua{first}))));
else
if isstructured
D = 0.5*T2 - real(inprod(T,Ua)) + 0.5*frob(Ua,'squared');
outputsrch.fval = D;
else
D = cpdres(T,Ua);
outputsrch.fval = 0.5*(D(:)'*D(:));
end
end
if outputsrch.fval > output.fval(end), K1 = []; end
end
% If line search did not improve the objective function, don't take it.
if outputsrch.fval > output.fval(end)
if isfunc(options.LineSearch)
alpha = [1 1];
elseif isfunc(options.PlaneSearch) && output.iterations >= 1
alpha = [1 0 1];
else
alpha = 1;
end
if options.FastUpdate && log10(output.fval(end)) > log10(T2)-16+2.5
outputsrch.fval = abs(0.5*(T2+sum(sum(real( ...
W.*UHU(:,:,last)))))-real(sum(dot(K{last},U{last}))));
else
if isstructured
D = 0.5*T2 - real(inprod(T,U)) + 0.5*frob(U,'squared');
outputsrch.fval = D;
else
D = cpdres(T,U);
outputsrch.fval = 0.5*(D(:)'*D(:));
end
end
end
% Prepare next ALS iteration, if necessary.
if options.FastUpdate
if exist('K1','var') && ~isempty(K1)
K{first} = K1;
UHU = UHU1;
else
if isfunc(options.LineSearch)
Ua = cellfun(@(u,v)alpha(2)*(u+alpha(1)*v),U1,dU, ...
'UniformOutput',false);
elseif isfunc(options.PlaneSearch) && output.iterations >= 1
Ua = cellfun(@(u,v,w)alpha(3)*(u+alpha(1)*v+alpha(2)*w),...
U1,dU1,dU2,'UniformOutput',false);
else
Ua = U;
end
K{first} = mtkrprod(T,Ua,first);
for n = 1:N, UHU(:,:,n) = Ua{n}'*Ua{n}; end
end
end
end
end