annotate deep/convolutional_dae/sgd_opt.py @ 303:ef28cbb5f464

Use sigmoids with cross-entropy cost in the ConvAutoencoders.
author Arnaud Bergeron <abergeron@gmail.com>
date Wed, 31 Mar 2010 15:54:47 -0400
parents 8babd43235dd
children
rev   line source
276
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
1 import time
294
8babd43235dd Save best valid score and test score in the db.
Arnaud Bergeron <abergeron@gmail.com>
parents: 288
diff changeset
2 import sys, os
276
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
3
288
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
4 from ift6266.utils.seriestables import *
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
5
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
6 default_series = {
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
7 'train_error' : DummySeries(),
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
8 'valid_error' : DummySeries(),
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
9 'test_error' : DummySeries()
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
10 }
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
11
276
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
12 def sgd_opt(train, valid, test, training_epochs=10000, patience=10000,
288
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
13 patience_increase=2., improvement_threshold=0.995, net=None,
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
14 validation_frequency=None, series=default_series):
276
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
15
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
16 if validation_frequency is None:
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
17 validation_frequency = patience/2
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
18
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
19 start_time = time.clock()
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
20
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
21 best_params = None
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
22 best_validation_loss = float('inf')
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
23 test_score = 0.
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
24
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
25 start_time = time.clock()
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
26
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
27 for epoch in xrange(1, training_epochs+1):
288
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
28 series['train_error'].append((epoch,), train())
276
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
29
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
30 if epoch % validation_frequency == 0:
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
31 this_validation_loss = valid()
288
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
32 series['valid_error'].append((epoch,), this_validation_loss*100.)
276
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
33 print('epoch %i, validation error %f %%' % \
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
34 (epoch, this_validation_loss*100.))
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
35
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
36 # if we got the best validation score until now
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
37 if this_validation_loss < best_validation_loss:
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
38
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
39 #improve patience if loss improvement is good enough
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
40 if this_validation_loss < best_validation_loss * \
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
41 improvement_threshold :
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
42 patience = max(patience, epoch * patience_increase)
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
43
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
44 # save best validation score and epoch number
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
45 best_validation_loss = this_validation_loss
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
46 best_epoch = epoch
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
47
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
48 # test it on the test set
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
49 test_score = test()
288
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
50 series['test_error'].append((epoch,), test_score*100.)
276
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
51 print((' epoch %i, test error of best model %f %%') %
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
52 (epoch, test_score*100.))
288
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
53 if net is not None:
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
54 net.save('best.net.new')
80ee63c3e749 Add net saving (only the best model) and error saving using SeriesTable
Arnaud Bergeron <abergeron@gmail.com>
parents: 276
diff changeset
55 os.rename('best.net.new', 'best.net')
276
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
56
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
57 if patience <= epoch:
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
58 break
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
59
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
60 end_time = time.clock()
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
61 print(('Optimization complete with best validation score of %f %%,'
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
62 'with test performance %f %%') %
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
63 (best_validation_loss * 100., test_score*100.))
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
64 print ('The code ran for %f minutes' % ((end_time-start_time)/60.))
727ed56fad12 Add reworked code for convolutional auto-encoder.
Arnaud Bergeron <abergeron@gmail.com>
parents:
diff changeset
65
294
8babd43235dd Save best valid score and test score in the db.
Arnaud Bergeron <abergeron@gmail.com>
parents: 288
diff changeset
66 return best_validation_loss, test_score