def sample_rbm(rbm, dataset = dataset, n_chains = 20, n_samples = 10, output_folder = location + 'rbm_plots'):
if not os.path.isdir(output_folder):
os.makedirs(output_folder)
os.chdir(output_folder)
datasets = load_data(dataset)
test_set_x, test_set_y = datasets[2]
test_set_x = shared_data(test_set_x)
###############
#Sampling RBM #
###############
print '...start sampling'
rng = numpy.random.RandomState(123)
n_test_samples = test_set_x.get_value(borrow = True).shape[0]
test_idx = rng.randint(n_test_samples - n_chains)
persistent_vis_chain = theano.shared(
numpy.asarray(test_set_x.get_value(borrow = True)[test_idx : test_idx + n_chains],
dtype = theano.config.floatX)
)
plot_every = 1000
([pre_h, h_m, h_samples, pre_v, v_m, v_samples], updates) = \
theano.scan(rbm.gibbs_vhv, outputs_info = [None, None, None, None, None, persistent_vis_chain], n_steps = plot_every)
updates.update({persistent_vis_chain : v_samples[-1]})
sample_fn = theano.function([], [v_m[-1], v_samples[-1]], \
updates = updates, name = 'sample_fn')
img_data = numpy.zeros((29 * n_samples + 1, 29 * n_chains - 1), dtype = 'uint8')
for idx in range(n_samples):
v_m, v_sample = sample_fn()
print "plot sample %d" %idx
img_data[29 * idx : 29 * idx + 28, :] = \
tile_raster_images(X = v_m,
img_shape = (28, 28),
tile_shape = (1, n_chains),
tile_spacing = (1, 1))
image = Image.fromarray(img_data)
image.save('sample.png')
os.chdir('../')
def train_DBN(dataset = dataset):
datasets = load_data(dataset, shared = True)
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
index = T.lscalar()
x = T.matrix('x')
y = T.ivector('y')
rng = numpy.random.RandomState(123)
theano_rng = RandomStreams(rng.randint(2 ** 30))
rbm1 = RBM(input = x, numpy_rng = rng, theano_rng = theano_rng,\
n_visible = 784, n_hidden = 529)
rbm2 = RBM(input = x, numpy_rng = rng, theano_rng = theano_rng,\
n_visible = 529, n_hidden = 529)
lgs = LogisticRegression(input = x, n_in = 529, n_out = 10)
print '...train rbm1'
time_train_rbm1 = train_rbm(rbm1, index, x, train_set_x, 'rbm1')
#rbm1 = cPickle.load( open(location + 'rbm1.pkl', 'r'))
print '...train rbm2'
pre_sigm_h1, train_set_h1 = rbm1.prob_up(train_set_x)
rbm2_train_x = shared_data(train_set_h1.eval())
time_train_rbm2 = train_rbm(rbm2, index, x, rbm2_train_x, 'rbm2')
#rbm2 = cPickle.load( open(location + 'rbm2.pkl', 'r'))
print '...dbn fintuning'
dbn = DBN(x, [rbm1.W, rbm2.W], [rbm1.hbias, rbm2.hbias], lgs.W, lgs.b)
time_train_dbn = train_classifier(dbn, index, x, y,
datasets = [[train_set_x, train_set_y],
[valid_set_x, valid_set_y],
[test_set_x, test_set_y]])
print 'Train DBN total cost %.2f' %((time_train_rbm1 + time_train_rbm2 + time_train_dbn) / 60.)
def StackedDA_sgd(dataset= dataset):
print '...load data'
datasets = load_data(dataset, shared = True)
train_set_x, train_set_y = datasets[0]
valid_set_x, valid_set_y = datasets[1]
test_set_x, test_set_y = datasets[2]
print '...build model'
index = T.lscalar() # index to a [mini]batch
x = T.matrix('x') # data, presented as rasterized images
y = T.ivector('y') # labels, presented as 1D vector of [int] labels
rng = numpy.random.RandomState(123)
theano_rng = RandomStreams(rng.randint(2 ** 30))
da1 = DA(numpy_rng = rng, theano_rng = theano_rng, input = x, n_visible = 784, n_hide = 529)
da2 = DA(numpy_rng = rng, theano_rng = theano_rng, input = x, n_visible = 529, n_hide = 484)
lgs = LogisticRegression(input = x, n_in = 484, n_out = 10)
print '...train 1st Da'
time_train_da1 = train_da(da1, index, x, train_set_x, 'da1_30.png', corruption_level = 0.3,
learning_rate = 0.1, training_epochs = 15, batch_size=20)
print '...train 2ed Da'
da2_train_x = shared_data( da1.get_hidden_value(train_set_x).eval())
time_train_da2 = train_da(da2, index, x, da2_train_x, 'da2_30.png', corruption_level = 0.15,
learning_rate = 0.1, training_epochs = 15, batch_size=20)
print '...fine-tuning'
stacked_da = StackedDA(x, [da1.W, da2.W], [da1.b, da2.b], lgs.W, lgs.b)
time_train_sda = train_classifier(stacked_da, index, x, y,
[[train_set_x, train_set_y],
[valid_set_x, valid_set_y],
[test_set_x, test_set_y]])
print 'Tiraning Total cost time %.2fm'%((time_train_da1 + time_train_da2 + time_train_sda) / 60.)
