rcnn_train.m 13.7 KB
Newer Older
Ross Girshick's avatar
Ross Girshick committed
1
2
function [rcnn_model, rcnn_k_fold_model] = ...
    rcnn_train(imdb, varargin)
3
4
5
6
% [rcnn_model, rcnn_k_fold_model] = rcnn_train(imdb, varargin)
%   Trains an R-CNN detector for all classes in the imdb.
%   
%   Keys that can be passed in:
Ross Girshick's avatar
Ross Girshick committed
7
%
8
9
10
11
12
13
14
15
16
17
%   svm_C             SVM regularization parameter
%   bias_mult         Bias feature value (for liblinear)
%   pos_loss_weight   Cost factor on hinge loss for positives
%   layer             Feature layer to use (either 5, 6 or 7)
%   k_folds           Train on folds of the imdb
%   checkpoint        Save the rcnn_model every checkpoint images
%   crop_mode         Crop mode (either 'warp' or 'square')
%   crop_padding      Amount of padding in crop
%   net_file          Path to the Caffe CNN to use
%   cache_name        Path to the precomputed feature cache
Ross Girshick's avatar
Ross Girshick committed
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

% AUTORIGHTS
% ---------------------------------------------------------
% Copyright (c) 2014, Ross Girshick
% 
% This file is part of the R-CNN code and is available 
% under the terms of the Simplified BSD License provided in 
% LICENSE. Please retain this notice and LICENSE if you use 
% this file (or any portion of it) in your project.
% ---------------------------------------------------------

% TODO:
%  - allow training just a subset of classes

ip = inputParser;
ip.addRequired('imdb', @isstruct);
ip.addParamValue('svm_C',           10^-3,  @isscalar);
ip.addParamValue('bias_mult',       10,     @isscalar);
ip.addParamValue('pos_loss_weight', 2,      @isscalar);
ip.addParamValue('layer',           7,      @isscalar);
ip.addParamValue('k_folds',         2,      @isscalar);
ip.addParamValue('checkpoint',      0,      @isscalar);
ip.addParamValue('crop_mode',       'warp', @isstr);
ip.addParamValue('crop_padding',    16,     @isscalar);
ip.addParamValue('net_file', ...
Ross Girshick's avatar
Ross Girshick committed
43
    './data/caffe_nets/finetune_voc_2007_trainval_iter_70k', ...
Ross Girshick's avatar
Ross Girshick committed
44
    @isstr);
Ross Girshick's avatar
Ross Girshick committed
45
46
47
ip.addParamValue('cache_name', ...
    'v1_finetune_voc_2007_trainval_iter_70000', @isstr);

Ross Girshick's avatar
Ross Girshick committed
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

ip.parse(imdb, varargin{:});
opts = ip.Results;

opts.net_def_file = './model-defs/rcnn_batch_256_output_fc7.prototxt';

conf = rcnn_config('sub_dir', imdb.name);

% Record a log of the training and test procedure
timestamp = datestr(datevec(now()), 'dd.mmm.yyyy:HH.MM.SS');
diary_file = [conf.cache_dir 'rcnn_train_' timestamp '.txt'];
diary(diary_file);
fprintf('Logging output in %s\n', diary_file);

fprintf('\n\n~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n');
fprintf('Training options:\n');
disp(opts);
fprintf('~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n\n');

% ------------------------------------------------------------------------
% Create a new rcnn model
rcnn_model = rcnn_create_model(opts.net_def_file, opts.net_file, opts.cache_name);
rcnn_model = rcnn_load_model(rcnn_model, conf.use_gpu);
rcnn_model.detectors.crop_mode = opts.crop_mode;
rcnn_model.detectors.crop_padding = opts.crop_padding;
rcnn_model.classes = imdb.classes;
% ------------------------------------------------------------------------

% ------------------------------------------------------------------------
% Get the average norm of the features
opts.feat_norm_mean = rcnn_feature_stats(imdb, opts.layer, rcnn_model);
fprintf('average norm = %.3f\n', opts.feat_norm_mean);
rcnn_model.training_opts = opts;
% ------------------------------------------------------------------------

% ------------------------------------------------------------------------
% Get all positive examples
% We cache only the pool5 features and convert them on-the-fly to
% fc6 or fc7 as required
save_file = sprintf('./feat_cache/%s/%s/gt_pos_layer_5_cache.mat', ...
    rcnn_model.cache_name, imdb.name);
try
  load(save_file);
  fprintf('Loaded saved positives from ground truth boxes\n');
catch
  [X_pos, keys_pos] = get_positive_pool5_features(imdb, opts);
  save(save_file, 'X_pos', 'keys_pos', '-v7.3');
end
% Init training caches
caches = {};
for i = imdb.class_ids
  fprintf('%14s has %6d positive instances\n', ...
      imdb.classes{i}, size(X_pos{i},1));
  X_pos{i} = rcnn_pool5_to_fcX(X_pos{i}, opts.layer, rcnn_model);
  X_pos{i} = rcnn_scale_features(X_pos{i}, opts.feat_norm_mean);
  caches{i} = init_cache(X_pos{i}, keys_pos{i});
end
% ------------------------------------------------------------------------

% ------------------------------------------------------------------------
% Train with hard negative mining
first_time = true;
% one pass over the data is enough
max_hard_epochs = 1;

for hard_epoch = 1:max_hard_epochs
  for i = 1:length(imdb.image_ids)
    fprintf('%s: hard neg epoch: %d/%d image: %d/%d\n', ...
            procid(), hard_epoch, max_hard_epochs, i, length(imdb.image_ids));

    % Get hard negatives for all classes at once (avoids loading feature cache
    % more than once)
    [X, keys] = sample_negative_features(first_time, rcnn_model, caches, ...
        imdb, i);

    % Add sampled negatives to each classes training cache, removing
    % duplicates
    for j = imdb.class_ids
      if ~isempty(keys{j})
        [~, ~, dups] = intersect(caches{j}.keys_neg, keys{j}, 'rows');
        assert(isempty(dups));
        caches{j}.X_neg = cat(1, caches{j}.X_neg, X{j});
        caches{j}.keys_neg = cat(1, caches{j}.keys_neg, keys{j});
        caches{j}.num_added = caches{j}.num_added + size(keys{j},1);
      end

      % Update model if
      %  - first time seeing negatives
      %  - more than retrain_limit negatives have been added
      %  - its the final image of the final epoch
      is_last_time = (hard_epoch == max_hard_epochs && i == length(imdb.image_ids));
      hit_retrain_limit = (caches{j}.num_added > caches{j}.retrain_limit);
      if (first_time || hit_retrain_limit || is_last_time) && ...
          ~isempty(caches{j}.X_neg)
        fprintf('>>> Updating %s detector <<<\n', imdb.classes{j});
        fprintf('Cache holds %d pos examples %d neg examples\n', ...
                size(caches{j}.X_pos,1), size(caches{j}.X_neg,1));
        [new_w, new_b] = update_model(caches{j}, opts);
        rcnn_model.detectors.W(:, j) = new_w;
        rcnn_model.detectors.B(j) = new_b;
        caches{j}.num_added = 0;

        z_pos = caches{j}.X_pos * new_w + new_b;
        z_neg = caches{j}.X_neg * new_w + new_b;

        caches{j}.pos_loss(end+1) = opts.svm_C * opts.pos_loss_weight * ...
                                    sum(max(0, 1 - z_pos));
        caches{j}.neg_loss(end+1) = opts.svm_C * sum(max(0, 1 + z_neg));
        caches{j}.reg_loss(end+1) = 0.5 * new_w' * new_w + ...
                                    0.5 * (new_b / opts.bias_mult)^2;
        caches{j}.tot_loss(end+1) = caches{j}.pos_loss(end) + ...
                                    caches{j}.neg_loss(end) + ...
                                    caches{j}.reg_loss(end);

        for t = 1:length(caches{j}.tot_loss)
          fprintf('    %2d: obj val: %.3f = %.3f (pos) + %.3f (neg) + %.3f (reg)\n', ...
                  t, caches{j}.tot_loss(t), caches{j}.pos_loss(t), ...
                  caches{j}.neg_loss(t), caches{j}.reg_loss(t));
        end

        % store negative support vectors for visualizing later
        SVs_neg = find(z_neg > -1 - eps);
        rcnn_model.SVs.keys_neg{j} = caches{j}.keys_neg(SVs_neg, :);
        rcnn_model.SVs.scores_neg{j} = z_neg(SVs_neg);

        % evict easy examples
        easy = find(z_neg < caches{j}.evict_thresh);
        caches{j}.X_neg(easy,:) = [];
        caches{j}.keys_neg(easy,:) = [];
        fprintf('  Pruning easy negatives\n');
        fprintf('  Cache holds %d pos examples %d neg examples\n', ...
                size(caches{j}.X_pos,1), size(caches{j}.X_neg,1));
        fprintf('  %d pos support vectors\n', numel(find(z_pos <  1 + eps)));
        fprintf('  %d neg support vectors\n', numel(find(z_neg > -1 - eps)));
      end
    end
    first_time = false;

    if opts.checkpoint > 0 && mod(i, opts.checkpoint) == 0
      save([conf.cache_dir 'rcnn_model'], 'rcnn_model');
    end
  end
end
% save the final rcnn_model
save([conf.cache_dir 'rcnn_model'], 'rcnn_model');
% ------------------------------------------------------------------------

% ------------------------------------------------------------------------
if opts.k_folds > 0
  rcnn_k_fold_model = rcnn_model;
  [W, B, folds] = update_model_k_fold(rcnn_model, caches, imdb);
  rcnn_k_fold_model.folds = folds;
  for f = 1:length(folds)
    rcnn_k_fold_model.detectors(f).W = W{f};
    rcnn_k_fold_model.detectors(f).B = B{f};
  end
  save([conf.cache_dir 'rcnn_k_fold_model'], 'rcnn_k_fold_model');
else
  rcnn_k_fold_model = [];
end
% ------------------------------------------------------------------------


% ------------------------------------------------------------------------
function [X_neg, keys] = sample_negative_features(first_time, rcnn_model, ...
                                                  caches, imdb, ind)
% ------------------------------------------------------------------------
opts = rcnn_model.training_opts;

d = rcnn_load_cached_pool5_features(opts.cache_name, ...
    imdb.name, imdb.image_ids{ind});

class_ids = imdb.class_ids;

if isempty(d.feat)
  X_neg = cell(max(class_ids), 1);
  keys = cell(max(class_ids), 1);
  return;
end

d.feat = rcnn_pool5_to_fcX(d.feat, opts.layer, rcnn_model);
d.feat = rcnn_scale_features(d.feat, opts.feat_norm_mean);

neg_ovr_thresh = 0.3;

if first_time
  for cls_id = class_ids
    I = find(d.overlap(:, cls_id) < neg_ovr_thresh);
    X_neg{cls_id} = d.feat(I,:);
    keys{cls_id} = [ind*ones(length(I),1) I];
  end
else
  zs = bsxfun(@plus, d.feat*rcnn_model.detectors.W, rcnn_model.detectors.B);
  for cls_id = class_ids
    z = zs(:, cls_id);
    I = find((z > caches{cls_id}.hard_thresh) & ...
             (d.overlap(:, cls_id) < neg_ovr_thresh));

    % Avoid adding duplicate features
    keys_ = [ind*ones(length(I),1) I];
    [~, ~, dups] = intersect(caches{cls_id}.keys_neg, keys_, 'rows');
    keep = setdiff(1:size(keys_,1), dups);
    I = I(keep);

    % Unique hard negatives
    X_neg{cls_id} = d.feat(I,:);
    keys{cls_id} = [ind*ones(length(I),1) I];
  end
end


% ------------------------------------------------------------------------
function [w, b] = update_model(cache, opts, pos_inds, neg_inds)
% ------------------------------------------------------------------------
solver = 'liblinear';
liblinear_type = 3;  % l2 regularized l1 hinge loss
%liblinear_type = 5; % l1 regularized l2 hinge loss

if ~exist('pos_inds', 'var') || isempty(pos_inds)
  num_pos = size(cache.X_pos, 1);
  pos_inds = 1:num_pos;
else
  num_pos = length(pos_inds);
  fprintf('[subset mode] using %d out of %d total positives\n', ...
      num_pos, size(cache.X_pos,1));
end
if ~exist('neg_inds', 'var') || isempty(neg_inds)
  num_neg = size(cache.X_neg, 1);
  neg_inds = 1:num_neg;
else
  num_neg = length(neg_inds);
  fprintf('[subset mode] using %d out of %d total negatives\n', ...
      num_neg, size(cache.X_neg,1));
end

switch solver
  case 'liblinear'
    ll_opts = sprintf('-w1 %.5f -c %.5f -s %d -B %.5f', ...
                      opts.pos_loss_weight, opts.svm_C, ...
                      liblinear_type, opts.bias_mult);
    fprintf('liblinear opts: %s\n', ll_opts);
    X = sparse(size(cache.X_pos,2), num_pos+num_neg);
    X(:,1:num_pos) = cache.X_pos(pos_inds,:)';
    X(:,num_pos+1:end) = cache.X_neg(neg_inds,:)';
    y = cat(1, ones(num_pos,1), -ones(num_neg,1));
    llm = liblinear_train(y, X, ll_opts, 'col');
    w = single(llm.w(1:end-1)');
    b = single(llm.w(end)*opts.bias_mult);

  otherwise
    error('unknown solver: %s', solver);
end


% ------------------------------------------------------------------------
function [W, B, folds] = update_model_k_fold(rcnn_model, caches, imdb)
% ------------------------------------------------------------------------
opts = rcnn_model.training_opts;
num_images = length(imdb.image_ids);
folds = create_folds(num_images, opts.k_folds);
W = cell(opts.k_folds, 1);
B = cell(opts.k_folds, 1);

fprintf('Training k-fold models\n');
for i = imdb.class_ids
  fprintf('\n\n~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n');
  fprintf('Training folds for class %s\n', imdb.classes{i});
  fprintf('~~~~~~~~~~~~~~~~~~~~~~~~~~~~\n\n');
  for f = 1:length(folds)
    fprintf('Holding out fold %d\n', f);
    [pos_inds, neg_inds] = get_cache_inds_from_fold(caches{i}, folds{f});
    [new_w, new_b] = update_model(caches{i}, opts, ...
        pos_inds, neg_inds);
    W{f}(:,i) = new_w;
    B{f}(i) = new_b;
  end
end


% ------------------------------------------------------------------------
function [pos_inds, neg_inds] = get_cache_inds_from_fold(cache, fold)
% ------------------------------------------------------------------------
pos_inds = find(ismember(cache.keys_pos(:,1), fold) == false);
neg_inds = find(ismember(cache.keys_neg(:,1), fold) == false);


% ------------------------------------------------------------------------
function [X_pos, keys] = get_positive_pool5_features(imdb, opts)
% ------------------------------------------------------------------------
X_pos = cell(max(imdb.class_ids), 1);
keys = cell(max(imdb.class_ids), 1);

for i = 1:length(imdb.image_ids)
  tic_toc_print('%s: pos features %d/%d\n', ...
                procid(), i, length(imdb.image_ids));

  d = rcnn_load_cached_pool5_features(opts.cache_name, ...
      imdb.name, imdb.image_ids{i});

  for j = imdb.class_ids
    if isempty(X_pos{j})
      X_pos{j} = single([]);
      keys{j} = [];
    end
    sel = find(d.class == j);
    if ~isempty(sel)
      X_pos{j} = cat(1, X_pos{j}, d.feat(sel,:));
      keys{j} = cat(1, keys{j}, [i*ones(length(sel),1) sel]);
    end
  end
end


% ------------------------------------------------------------------------
function cache = init_cache(X_pos, keys_pos)
% ------------------------------------------------------------------------
cache.X_pos = X_pos;
cache.X_neg = single([]);
cache.keys_neg = [];
cache.keys_pos = keys_pos;
cache.num_added = 0;
cache.retrain_limit = 2000;
cache.evict_thresh = -1.2;
cache.hard_thresh = -1.0001;
cache.pos_loss = [];
cache.neg_loss = [];
cache.reg_loss = [];
cache.tot_loss = [];