def train(pretrain_set, train_set, test_set,
pretrain_config, finetune_config, n_layers=10, n_hidden=140, coxphfit=False,
dropout_rate=0.5, non_lin=None):
finetune_lr = theano.shared(numpy.asarray(finetune_config['ft_lr'], dtype=theano.config.floatX))
learning_rate_decay = .989
# changed to theano shared variable in order to do minibatch
#train_set = theano.shared(value=train_set, name='train_set')
# numpy random generator
numpy_rng = numpy.random.RandomState(121212)
print '... building the model'
# construct the stacked denoising autoencoder and the corresponding regression network
model = Model(
numpy_rng = numpy_rng,
n_ins = train_set['X'].shape[1],
hidden_layers_sizes = [n_hidden] * n_layers,
n_outs = 1,
dropout_rate=dropout_rate,
non_lin=non_lin)
#########################
# PRETRAINING THE MODEL #
#########################
if pretrain_config is not None:
n_train_batches = len(train_set) / pretrain_config['pt_batchsize'] if pretrain_config['pt_batchsize'] else 1
print '... getting the pretraining functions'
pretraining_fns = model.pretraining_functions(pretrain_set,
pretrain_config['pt_batchsize'])
print '... pre-training the model'
start_time = timeit.default_timer()
# de-noising level
corruption_levels = [pretrain_config['corruption_level']] * n_layers
for i in xrange(model.n_layers): #Layerwise pre-training
# go through pretraining epochs
for epoch in xrange(pretrain_config['pt_epochs']):
# go through the training set
c = []
for batch_index in xrange(n_train_batches):
c.append(pretraining_fns[i](index=batch_index,
corruption=corruption_levels[i],
lr=pretrain_config['pt_lr']))
print "Pre-training layer %i, epoch %d, cost" % (i, epoch),
print numpy.mean(c)
end_time = timeit.default_timer()
print >> sys.stderr, ('The pretraining code for file ' +
os.path.split(__file__)[1] +
' ran for %.2fm' % ((end_time - start_time) / 60.))
########################
# FINETUNING THE MODEL #
########################
print '... getting the finetuning functions'
forward, backward = model.build_finetune_functions(
learning_rate=finetune_lr
)
print '... finetunning the model'
# early-stopping parameters
cindex_train = []
cindex_test = []
train_cost_list = []
test_cost_list = []
epoch = 0
while epoch < finetune_config['ft_epochs']:
epoch += 1
#print epoch
train_cost, train_risk, train_features = forward(train_set['X'], train_set['O'], train_set['A'], 1)
backward(train_set['X'], train_set['O'], train_set['A'], 1)
train_c_index = _naive_concordance_index(train_set['T'], -train_risk, train_set['O'])
test_cost, test_risk, test_features = forward(test_set['X'], test_set['O'], test_set['A'], 0)
test_c_index = _naive_concordance_index(test_set['T'], -test_risk, test_set['O'])
cindex_train.append(train_c_index)
cindex_test.append(test_c_index)
train_cost_list.append(train_cost)
test_cost_list.append(test_cost)
print 'epoch = %d, trn_cost = %f, trn_ci = %f, tst_cost = %f, tst_ci = %f' % (epoch, train_cost, train_c_index, test_cost, test_c_index)
decay_learning_rate = theano.function(inputs=[], outputs=finetune_lr, \
updates={finetune_lr: finetune_lr * learning_rate_decay})
decay_learning_rate()
print 'best score is: %f' % max(cindex_test)
return train_cost_list, cindex_train, test_cost_list, cindex_test
