np_rng=rng)
#%%============================================================================
# Training the RBM
#==============================================================================
start_time = timeit.default_timer()
accuracies = []
argmax_acc = 0
for epoch in xrange(training_epochs):
epoch_time = timeit.default_timer()
mean_cost = []
for batch_index, batch in enumerate(batches):
rbm.update(batch, persistent=True, k=k)
sys.stdout.write("\rEpoch advancement: %d%%" % (100*float(batch_index)/len(batches)))
sys.stdout.flush()
# Training Logistic regression
sys.stdout.write("\rTraining softmax...")
sm_time = timeit.default_timer()
softmax_classifier = LogisticRegression(penalty='l1',
C=1.0,
solver='lbfgs',
multi_class='multinomial')
softmax_classifier.fit(rbm.propup(train_set[:,n_labels:], np.ones((len(train_set),n_hidden))),
train_labels)
sys.stdout.write('\rSoftmax trained in %f minutes.\n' % ((timeit.default_timer()-sm_time) / 60.))
sys.stdout.write("Evaluating accuracy...")
cv_time = timeit.default_timer()
acc = softmax_classifier.score(rbm.propup(test_set[:,n_labels:], np.ones((len(test_set),n_hidden))),
# Training the RBM
#==========================================================================
max_score = -np.inf
argmax_score = RBM(n_visible=n_visible,
n_labels=n_labels,
n_hidden=n_hidden,
dropout_rate=dropout_rate,
batch_size=batch_size,
np_rng=rng)
start_time = timeit.default_timer()
for epoch in xrange(training_epochs):
epoch_time = timeit.default_timer()
mean_cost = []
for batch_index in xrange(n_train_batches):
rbm.update(np_train_set[batch_index*batch_size:(batch_index+1)*batch_size,:], persistent=True, k=k)
sys.stdout.write("\rEpoch advancement: %d%%" % (100*float(batch_index)/n_train_batches))
sys.stdout.flush()
sys.stdout.write("\rEvaluating accuracy...")
sys.stdout.flush()
cv_time = timeit.default_timer()
acc = rbm.cv_accuracy(np_test_set)
sys.stdout.write('''\rEpoch %i took %f minutes,
accuracy (computed in %f minutes) is %f.\n'''
% (epoch,
((cv_time-epoch_time) / 60.),
((timeit.default_timer()-cv_time) / 60.),
acc))
if scoring=='cost':
score = np.mean(mean_cost)
