-
Notifications
You must be signed in to change notification settings - Fork 6
/
training.m
275 lines (224 loc) · 8.9 KB
/
training.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
function training(experiment, varargin)
%TRAINING Performs training for any solver, model and dataset.
% Name-value pairs are supported, listed below. See readme for examples.
% options (override by calling script with name-value pairs).
% (-D-) := if left empty, the default value for the chosen model will be used.
opts.dataset = 'cifar'; % mnist, cifar or imagenet
[opts, varargin] = vl_argparse(opts, varargin, 'nonrecursive');
switch opts.dataset
case 'mnist'
opts.dataDir = [vl_rootnn() '/data/mnist']; % data location
opts.model = models.LeNet(); % choose model (type 'help models' for a list)
numClasses = 10;
case 'cifar'
opts.dataDir = [vl_rootnn() '/data/cifar'];
opts.model = models.BasicCifarNet('batchNorm', true);
numClasses = 10;
case {'imagenet', 'imagenet-100'}
opts.dataDir = [vl_rootnn() '/data/ilsvrc12'];
if strcmp(opts.dataset, 'imagenet-100') % smaller imagenet subset
numClasses = 100;
else
numClasses = 1000;
end
opts.model = models.VGG8('batchNorm', true, 'numClasses', numClasses);
otherwise
error('Unknown dataset.');
end
opts.resultsDir = [vl_rootnn() '/data/curveball/' opts.dataset '/' experiment]; % results location
opts.conserveMemory = true; % whether to conserve memory
opts.numEpochs = []; % epochs (-D-)
opts.batchSize = []; % batch size (-D-)
opts.learningRate = []; % learning rate (-D-)
opts.weightDecay = 0; % weight decay (-D-)
opts.solver = solvers.SGD(); % solver instance to use (type 'help solvers' for a list)
opts.gpu = 1; % GPU index, empty for CPU mode
opts.numThreads = 12; % number of threads for image reading
opts.augmentation = []; % data augmentation (see datasets.ImageFolder) (-D-)
opts.savePlot = true; % whether to save the plot as a PDF file
opts.continue = false; % continue from last checkpoint if available
opts.cifarShift = 2 ; % number of pixels to shift in CIFAR data augmentation
opts = vl_argparse(opts, varargin, 'nonrecursive'); % let user override options
setup_curveball(); % add functions to path
mkdir(opts.resultsDir);
%
% set up network
%
% use chosen model's output as the predictions
assert(isa(opts.model, 'Layer'), 'Model must be a CNN (e.g. models.AlexNet()).')
predictions = opts.model;
% change the model's input name
images = predictions.find('Input', 1);
images.name = 'images';
images.gpu = true;
% validate the prediction size (must predict 1000 classes)
defaults = predictions.meta; % get model's meta information (default learning rate, etc)
outputSize = predictions.evalOutputSize('images', [defaults.imageSize 5]);
assert(isequal(outputSize, [1 1 numClasses 5]), 'Model output does not have the correct shape.');
% mark conv layer that comes before the predictions, if possible
if isequal(predictions.func, @vl_nnconv)
prelayer = predictions.inputs{1};
prelayer.name = 'prelayer';
end
% replace empty options with model-specific default values
for name_ = {'numEpochs', 'batchSize', 'learningRate', 'weightDecay', 'augmentation'}
name = name_{1}; % dereference cell array
if isempty(opts.(name)) && isfield(defaults, name)
opts.(name) = defaults.(name);
end
end
% create losses
labels = Input();
obj = vl_nnloss(predictions, labels, 'loss', 'softmaxlog');
% assign layer names automatically, and compile network
Layer.workspaceNames();
% compile net
baseline = ~isa(opts.solver, 'CurveBall');
if baseline % standard net with loss
output = obj;
else % net outputs the prediction only, loss is computed by hand
output = predictions;
if isequal(output.func, @vl_nnrelu)
output = 1 * output; % hack to keep ReLUs from being short-circuited in some models
end
end
net = Net(output, 'conserveMemory', opts.conserveMemory);
%
% set up solver and dataset
%
% set solver learning rate
solver = opts.solver;
solver.learningRate = opts.learningRate(1) ;
solver.weightDecay = opts.weightDecay ;
if ~baseline
solver.net = net;
end
% initialize dataset
augmenter = @deal;
switch opts.dataset
case 'mnist'
dataset = datasets.MNIST(opts.dataDir, 'batchSize', opts.batchSize);
case 'cifar'
dataset = datasets.CIFAR10(opts.dataDir, 'batchSize', opts.batchSize);
% data augmentation: horizontal flips, and shifts of a few pixels
shift = opts.cifarShift;
augmenter = @(images) small_image_augmentation(images, shift);
case {'imagenet', 'imagenet-100'}
dataset = datasets.ImageNet('dataDir', opts.dataDir, ...
'imageSize', defaults.imageSize, 'useGpu', ~isempty(opts.gpu));
dataset.augmentation = opts.augmentation;
dataset.numThreads = opts.numThreads;
% smaller imagenet subset
if strcmp(opts.dataset, 'imagenet-100')
subset = dlmread('imagenet-subset.txt');
imagenet_subset(dataset, subset);
end
end
dataset.batchSize = opts.batchSize;
stats = Stats();
stats.registerVars({'obj', 'err', 'time'}, false);
%
% training
%
% continue from last checkpoint if there is one
startEpoch = 1;
if opts.continue
[filename, startEpoch] = models.checkpoint([opts.resultsDir '/epoch-*.mat']);
end
if startEpoch > 1
load(filename, 'net', 'stats', 'solver');
end
% enable GPU mode
net.useGpu(opts.gpu);
non_grad_params = [net.params([net.params.trainMethod] ~= 1).var];
for epoch = startEpoch : opts.numEpochs
% get the learning rate for this epoch, if there is a schedule
if epoch <= numel(opts.learningRate)
solver.learningRate = opts.learningRate(epoch);
end
% clear batch-norm moments derivatives for FMAD
net.setDer(non_grad_params, repmat({0}, size(non_grad_params)));
% training phase
train_clock = tic();
numBatches = floor(numel(dataset.trainSet) / opts.batchSize);
for batch = dataset.train()
% draw samples
[images, labels] = dataset.get(batch);
images = augmenter(images); % apply data augmentation
tic();
% take one solver step
if baseline
% simple backprop
net.eval({'images', images, 'labels', labels});
pred_value = net.getValue('predictions');
solver.step(net);
obj_value = net.getValue('obj');
else
% use CurveBall method. first, evaluate network on inputs
net.eval({'images', images}, 'forward');
pred_value = net.getValue('predictions');
% do a step
solver.labels = labels;
solver.step(net);
obj_value = solver.last_loss;
end
% compute classification error and update statistics
assert(all(isfinite(pred_value(:))), 'Optimization diverged.');
err_value = vl_nnloss(pred_value, labels, 'loss', 'classerror');
stats.update('obj', obj_value, 'err', err_value);
iter = stats.counts(1); % number of iterations
if iter > 3, stats.update('time', toc() * 1000); end % skip first 3 batches
if isa(solver, 'CurveBall') && solver.autolambda
stats.update('lambda', solver.lambda, 'ratio', solver.ratio);
end
% report statistics and iteration number
fprintf('ep%d %d/%d eta %s ', epoch, iter, numBatches, ...
eta(stats.average('time'), iter, numBatches, epoch, opts.numEpochs));
stats.print();
end
% push average objective and error (after one epoch) into the plot
stats.update('total_time', toc(train_clock));
stats.push('train');
% validation phase
numBatches = floor(numel(dataset.valSet) / opts.batchSize);
for batch = dataset.val()
[images, labels] = dataset.get(batch);
% evaluate network
tic();
net.eval({'images', images}, 'test');
% compute objective and error, and update statistics
pred_value = net.getValue('predictions');
obj_value = vl_nnloss(pred_value, labels, 'loss', 'softmaxlog');
err_value = vl_nnloss(pred_value, labels, 'loss', 'classerror');
stats.update('obj', obj_value, 'err', err_value, 'time', toc() * 1000);
% report statistics and iteration number
fprintf('val ep%d %d/%d ', epoch, stats.counts(1), numBatches);
stats.print();
end
stats.push('val');
% plot statistics
stats.plot('figure', 1, 'names', {'obj', 'err'});
if ~isempty(opts.resultsDir)
% save the plot
if opts.savePlot
print(1, [opts.resultsDir '/plot.pdf'], '-dpdf');
end
% save checkpoint every few epochs
if mod(epoch, 5) == 0
save(sprintf('%s/epoch-%d.mat', opts.resultsDir, epoch), ...
'net', 'stats', 'solver');
end
end
end
% save results
if ~isempty(opts.resultsDir)
save([opts.resultsDir '/results.mat'], 'net', 'stats', 'solver');
end
end
function str = eta(batchTime, batch, batchesPerEpoch, epoch, numEpochs)
% generate string estimating the time of completion
completed = batch + batchesPerEpoch * (epoch - 1);
total = batchesPerEpoch * numEpochs;
secs = (total - completed) * batchTime / 1000 ;
str = duration(0, 0, secs) ;
end