File size: 12,509 Bytes
d4035c1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
function [conf, jobId] = bk_trainAppModel(conf, varargin)
% BK_TRAINAPPMODEL
%
%  modelDirs:    one for each class
%  roiDir:       where to find GT rois
%
%  Author:: Andrea Vedaldi

% AUTORIGHTS
% Copyright (C) 2008-09 Andrea Vedaldi
%
% This file is part of the VGG MKL Class and VGG MKL Det code packages,
% available in the terms of the GNU General Public License version 2.

conf_.kerDir         = '' ;
conf_.kerDb          = struct ;
conf_.trainRoiDbPath = '' ;
conf_.trainHistDir   = '' ;
conf_.testRoiDbPath  = '' ;
conf_.testHistDir    = '' ;
conf_.modelPath      = '' ;
conf_.featWeights    = [] ;
conf_.learnWeights   = 'none' ;
conf_.posWeight      = 1 ;
conf_.noClobber      = false ;

if nargin == 0, conf = conf_ ; return ; end
conf = override(conf_, conf, 1) ;

[jobId, taskId] = parallelDriver('numNodes', 1, ...
                                 'freeMemMb', 4 * 1024, ...
                                 varargin{:}) ;
if ~isnan(jobId) & isnan(taskId) ; return ; end

% --------------------------------------------------------------------
%                                      Process all classes and aspects
% --------------------------------------------------------------------

if conf.noClobber & checkFile(conf.modelPath)
  fprintf('\tSkipping to avoid clobbering ''%s''.\n', conf.modelPath) ;
  return ;
end

rand('state', 0) ;
randn('state', 0) ;

ensuredir(fileparts(conf.modelPath)) ;

% ------------------------------------------------------------------
%                                               Load training ROI DB
% ------------------------------------------------------------------

fprintf('\tLoading training ROI DB ''%s''.\n', ...
        conf.trainRoiDbPath) ;
trdb = load(conf.trainRoiDbPath) ;
labels = trdb.labels(:)' ;

fprintf('\tWeight learning stragegy: %s\n', conf.learnWeights) ;
fprintf('\tInitial weights: %s\n', sprintf('%g ', conf.featWeights)) ;

% ------------------------------------------------------------------
%                                                      Train weights
% ------------------------------------------------------------------

if ~isempty(conf.featWeights)
  initWeights = conf.featWeights ;
else
  initWeights = ones(1, length(conf.kerDb)) ;
end

% if there is only one kernel, MKL is not needed
if length(conf.kerDb) == 1,
  conf.learnWeights = 'none' ;
end

switch conf.learnWeights
  case 'none'
    weights = initWeights ;

  case 'equalMean' % ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    [base, kernels, labels] = ...
        loadBase(conf.kerDir, conf.kerDb, trdb.labels(:)', ...
                 'weights', initWeights, ...
                 'maxSize', 5000, ...
                 'squeeze', false) ;
    for i=1:size(base,3)
      tmp = base(:,:,i) ;
      weights(i) = initWeights(i) ./ mean(tmp(:)) ;
      clear tmp ;
    end
    svm.d = weights ;
    fprintf('\tComputed weights: %s\n', sprintf('%g ', weights)) ;

    % KERNELS contains the full list of kernels, with gamma for the exp


  case 'manik' % ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    [base, kernels, labels] = ...
        loadBase(conf.kerDir, conf.kerDb, trdb.labels(:)', ...
                 'weights', initWeights, ...
                 'maxSize', 5000, ...
                 'squeeze', false) ;
    svm = learnGmklSvm(base, labels(:)) ;
    svm = svmflip(svm, labels) ;
    weights = svm.d' .* initWeights ;
    svm.d = weights ;
    fprintf('\tComputed weights: %s\n', sprintf('%g ', weights)) ;

    % KERNELS contains the full list of kernels, with gamma for the exp

  case 'manikPartition' % ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
    numPartitions = 0 ;
    kerTypes = uniqueStrings({conf.kerDb.type}) ;
    featNames = uniqueStrings({conf.kerDb.feat}) ;
    for kerType = kerTypes
      kerType = char(kerType) ;

      for featName = featNames
        featName = char(featName) ;
        fprintf('Partition learning %d: %s %s\n', ...
                numPartitions, kerType, featName) ;

        selKer = findKernels(conf.kerDb, ...
                             'type', kerType,  ...
                             'feat', featName) ;

        if isempty(selKer), continue ; end
        numPartitions = numPartitions + 1 ;

        [base, kernels{numPartitions}, labels] = ...
            loadBase(conf.kerDir, conf.kerDb(selKer), trdb.labels(:)', ...
                     'weights', initWeights(selKer), ...
                     'maxSize', 10000, ...
                     'squeeze', false) ;
        svm = learnGmklSvm(base, labels(:)) ;
        svm = svmflip(svm, labels) ;
        weights(selKer) = svm.d' .* initWeights(selKer) ;
        fprintf('\tRelative weights: %s\n', sprintf('%g ', svm.d)) ;
        clear base ;
      end
    end

    fprintf('Partition learning: intra weights\n') ;
    base = [] ;
    thisPartition = 0 ;
    for kerType = kerTypes
      kerType = char(kerType) ;

      for featName = featNames
        featName = char(featName) ;

        selKer = findKernels(conf.kerDb, ...
                             'type', kerType,  ...
                             'feat', featName) ;

        if isempty(selKer), continue ; end
        thisPartition = thisPartition + 1 ;

        [partition, drop, labels] = ...
            loadBase(conf.kerDir, conf.kerDb(selKer), ...
                     trdb.labels(:)', ...
                     'weights', weights(selKer), ...
                     'maxSize', 7000, ...
                     'squeeze', true) ;

        if isempty(base)
          base = zeros([size(partition) numPartitions]) ;
        end
        base(:,:,thisPartition) = partition ;
        clear partition ;
      end
    end

    svm = learnGmklSvm(base, labels(:)) ;
    svm = svmflip(svm, labels) ;
    fprintf('\tRelative weights: %s\n', sprintf('%g ', svm.d)) ;
    clear base ;

    thisPartition = 0 ;
    for kerType = kerTypes
      kerType = char(kerType) ;

      for featName = featNames
        featName = char(featName) ;

        selKer = findKernels(conf.kerDb, ...
                             'type', kerType,  ...
                             'feat', featName) ;

        if isempty(selKer), continue ; end
        thisPartition = thisPartition + 1 ;

        weights(selKer) = svm.d(thisPartition) .* weights(selKer) ;
      end
    end

    % kernels contains the full list of kernels, with gamma for the exp
    kernels = cat(2, kernels{:}) ;
    svm.d = weights ;

  otherwise
    error('Unknown weight learning method') ;
end

% ------------------------------------------------------------------
%                                                          Train SVM
% ------------------------------------------------------------------

if ~exist('svm', 'var') | length(labels) < length(trdb.labels) | ...
    conf.posWeight
  fprintf('\tFinal SVM learning (pos weight: %f)\n', conf.posWeight) ;
  [base,kernels,labels] = ...
      loadBase(conf.kerDir, conf.kerDb, trdb.labels(:)', ...
               'weights', weights, ...
               'squeeze', true) ;

  svm = svmkernellearn(base, labels(:)',   ...
                       'type', 'C',        ...
                       'C', 10,            ...
                       'verbosity', 1,     ...
                       'weights', [+1 conf.posWeight ; -1 1]') ;
  svm = svmflip(svm, trdb.labels) ;

  % test it on train
  scores = svm.alphay' * base(svm.svind, :) + svm.b ;
  errs = scores .* trdb.labels(:)' < 0 ;
  err  = mean(errs) ;
  selPos = find(trdb.labels(:)' > 0) ;
  selNeg = find(trdb.labels(:)' < 0) ;
  werr = sum(errs(selPos)) * conf.posWeight + sum(errs(selNeg)) ;
  werr = werr / (length(selPos) * conf.posWeight + length(selNeg)) ;
  fprintf('\tSVM training error: %.2f%% (weighed: %.2f%%).\n', ...
          err*100, werr*100) ;

  svm.d = weights ;
end

svm.kernels = kernels ;
clear base ;



% ------------------------------------------------------------------
%                                               Compress linear SVMs
% ------------------------------------------------------------------


isLinear = true ;

for ki=1:length(svm.kernels)
  kernel = svm.kernels(ki) ;
  if ~strcmp(kernel.type, 'kl2'), isLinear = false ; continue ; end
  histPath = fullfile(conf.trainHistDir, ['hist-' kernel.histName '.mat']) ;
  fprintf('\tLoading histograms''%s''.\n', histPath) ;
  tmp = load(histPath) ;
  svm.(kernel.histName) = ...
      svm.d(ki) * tmp.hists(:, svm.svind) * svm.alphay ;
  clear tmp ;
end

if isLinear
  svm.type = 'linear' ;
else
  svm.type = 'nonlinear' ;
end

fprintf('\tNumber of support vectors: %d\n', length(svm.svind)) ;
fprintf('\tSVM type: %s\n', svm.type) ;

fprintf('\tSaving model ''%s''.\n', conf.modelPath) ;
ssave(conf.modelPath, '-STRUCT', 'svm') ;

return

% ------------------------------------------------------------------
%                                                            Test it
% ------------------------------------------------------------------

if isempty(conf.testRoiDbPath), return ; end

fprintf('\tLoading test ROI DB ''%s''.\n', conf.testRoiDbPath) ;
tsdb = load(conf.testRoiDbPath) ;

scores = zeros(size(tsdb.labels)) + svm.b ;
for ki=1:K
  matPath = fullfile(conf.kerDir, ['test-' kernels(ki).name '.mat']) ;
  fprintf('\tLoading kernel matrix ''%s''.\n', matPath) ;
  kernels_ = load(matPath) ;

  scores = scores + svm.d(ki) * ...
           (svm.alphay' * kernels_.matrix(svm.svind, :)) ;
  clear kernels_ ;
end

figure(100) ; clf ; hold on ;
[recall,precision,info] = pr(tsdb.labels, scores) ;
plot(recall,precision,'b-','linewidth',2) ;
grid on ;
xlim([0 1]) ;
ylim([0 1]) ;
xlabel('recall') ;
ylabel('precision') ;
title(sprintf('PR on test ROI DB - AUC: %.2f %% (%.2f %%)', ...
              info.auc, info.auc_pa08)) ;
axis square ;
drawnow ;

[a,b,c] = fileparts(conf.modelPath) ;

figPath = fullfile(a, [b '-pr-test.eps']) ;
printsize(.97 + 0.001*randn) ; % ugly hack for matlab repeat
print('-depsc', figPath) ;

prPath = fullfile(a, [b '-pr-test.mat']) ;
ssave(prPath, 'recall', 'precision', 'info') ;


% ------------------------------------------------------------------
function [base, kernels, labels] =  ...
    loadBase(kerDir, kernels, labels, varargin)
% ------------------------------------------------------------------

opts.squeeze = false ;
opts.weights = NaN ;
opts.maxSize = +inf ;
opts = vl_argparse(opts,varargin{:}) ;

base = [] ;
sel  = NaN ;
K = length(kernels) ;

% calculate subset of kernel to load ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
if length(labels) > opts.maxSize
  selp = find(labels > 0) ;
  seln = find(labels < 0) ;
  np = length(selp) ;
  nm = length(seln) ;
  alpha = opts.maxSize / length(labels) ;
  mp = floor(alpha * np) ;
  mm = floor(alpha * nm) ;
  if (mp + mm < opts.maxSize), mp = mp + 1 ; end
  if (mp + mm < opts.maxSize), mm = mm + 1 ; end

  randn('state', 0) ;
  rand('state', 0) ;
  selp = randcol(selp, mp) ;
  seln = randcol(seln, mm) ;
  sel  = [selp seln] ;

  labels = labels(sel) ;

  fprintf('\tLoading kernels for only %.2f %% of the data (pos: %d neg: %d).\n', ...
          alpha * 100, length(selp), length(seln)) ;
end

% load matrices ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
clear kernels_ kernels__ ;
for ki = 1:K
  kerPath = fullfile(kerDir, [kernels(ki).name '.mat']) ;
  fprintf('\tLoading kernel matrix ''%s''.\n', kerPath) ;
  kernels_ = load(kerPath) ;
  if ~isfield(kernels_,'mu')
    [kernels_.mu] = deal([]) ;
  end
  if ~isnan(sel)
    fprintf('\t\tCutting down data.\n') ;
    kernels_.matrix = kernels_.matrix(sel, sel) ;
  end
  if ~isnan(opts.weights)
    fprintf('\t\tWeighing by %g.\n', opts.weights(ki)) ;
    kernels_.matrix = kernels_.matrix * opts.weights(ki) ;
  end
  if opts.squeeze
    if isempty(base)
      base = zeros(size(kernels_.matrix)) ;
    end
    base = base + kernels_.matrix ;
  else
    if isempty(base)
      base = zeros([size(kernels_.matrix) K]) ;
    end
    base(:,:,ki) = kernels_.matrix ;
  end
  kernels__(ki) = normalizeKernelStructure(kernels_);
  clear kernels_ ;
end

kernels = kernels__ ;
clear kernels__ ;

info = whos('base') ;
fprintf('\tKernel matrices size %.2f GB\n', info.bytes / 1024^3) ;
clear info ;

% ------------------------------------------------------------------
function kernel_ = normalizeKernelStructure(kernel)
% ------------------------------------------------------------------
kernel_.type     = kernel.type ;
kernel_.feat     = kernel.feat ;
kernel_.pyrLevel = kernel.pyrLevel ;
kernel_.histName = kernel.histName ;
kernel_.name     = kernel.name ;