def test_make_symbolic_state():
# Tests whether the returned p_sample and h_sample have the right
# dimensions
num_examples = 40
theano_rng = MRG_RandomStreams(2012+11+1)
visible_layer = BinaryVector(nvis=100)
rval = visible_layer.make_symbolic_state(num_examples=num_examples,
theano_rng=theano_rng)
hidden_layer = BinaryVectorMaxPool(detector_layer_dim=500,
pool_size=1,
layer_name='h',
irange=0.05,
init_bias=-2.0)
p_sample, h_sample = hidden_layer.make_symbolic_state(num_examples=num_examples,
theano_rng=theano_rng)
softmax_layer = Softmax(n_classes=10, layer_name='s', irange=0.05)
h_sample_s = softmax_layer.make_symbolic_state(num_examples=num_examples,
theano_rng=theano_rng)
required_shapes = [(40, 100), (40, 500), (40, 500), (40, 10)]
f = function(inputs=[],
outputs=[rval, p_sample, h_sample, h_sample_s])
for s, r in zip(f(), required_shapes):
assert s.shape == r
def test_make_symbolic_state():
# Tests whether the returned p_sample and h_sample have the right
# dimensions
num_examples = 40
theano_rng = MRG_RandomStreams(2012+11+1)
visible_layer = BinaryVector(nvis=100)
rval = visible_layer.make_symbolic_state(num_examples=num_examples,
theano_rng=theano_rng)
hidden_layer = BinaryVectorMaxPool(detector_layer_dim=500,
pool_size=1,
layer_name='h',
irange=0.05,
init_bias=-2.0)
p_sample, h_sample = hidden_layer.make_symbolic_state(num_examples=num_examples,
theano_rng=theano_rng)
softmax_layer = Softmax(n_classes=10, layer_name='s', irange=0.05)
h_sample_s = softmax_layer.make_symbolic_state(num_examples=num_examples,
theano_rng=theano_rng)
required_shapes = [(40, 100), (40, 500), (40, 500), (40, 10)]
f = function(inputs=[],
outputs=[rval, p_sample, h_sample, h_sample_s])
for s, r in zip(f(), required_shapes):
assert s.shape == r
def setup():
"""
Create pickle file with a simple model.
"""
# tearDown is guaranteed to run pop_load_data.
control.push_load_data(False)
with open('dbm.pkl', 'wb') as f:
dataset = MNIST(which_set='train', start=0, stop=100, binarize=True)
vis_layer = BinaryVector(nvis=784, bias_from_marginals=dataset)
hid_layer1 = BinaryVectorMaxPool(layer_name='h1',
pool_size=1,
irange=.05,
init_bias=-2.,
detector_layer_dim=50)
hid_layer2 = BinaryVectorMaxPool(layer_name='h2',
pool_size=1,
irange=.05,
init_bias=-2.,
detector_layer_dim=10)
model = DBM(batch_size=20,
niter=2,
visible_layer=vis_layer,
hidden_layers=[hid_layer1, hid_layer2])
model.dataset_yaml_src = """
!obj:pylearn2.datasets.binarizer.Binarizer {
raw: !obj:pylearn2.datasets.mnist.MNIST {
which_set: "train",
start: 0,
stop: 100
}
}
"""
model.layer_to_chains = model.make_layer_to_state(1)
cPickle.dump(model, f, protocol=cPickle.HIGHEST_PROTOCOL)
def make_random_basic_binary_dbm(
rng,
pool_size_1,
num_vis = None,
num_pool_1 = None,
num_pool_2 = None,
pool_size_2 = None,
center = False
):
"""
Makes a DBM with BinaryVector for the visible layer,
and two hidden layers of type BinaryVectorMaxPool.
The weights and biases are initialized randomly with
somewhat large values (i.e., not what you'd want to
use for learning)
rng: A numpy RandomState.
pool_size_1: The size of the pools to use in the first
layer.
"""
if num_vis is None:
num_vis = rng.randint(1,11)
if num_pool_1 is None:
num_pool_1 = rng.randint(1,11)
if num_pool_2 is None:
num_pool_2 = rng.randint(1,11)
if pool_size_2 is None:
pool_size_2 = rng.randint(1,6)
num_h1 = num_pool_1 * pool_size_1
num_h2 = num_pool_2 * pool_size_2
v = BinaryVector(num_vis, center=center)
v.set_biases(rng.uniform(-1., 1., (num_vis,)).astype(config.floatX), recenter=center)
h1 = BinaryVectorMaxPool(
detector_layer_dim = num_h1,
pool_size = pool_size_1,
layer_name = 'h1',
center = center,
irange = 1.)
h1.set_biases(rng.uniform(-1., 1., (num_h1,)).astype(config.floatX), recenter=center)
h2 = BinaryVectorMaxPool(
center = center,
detector_layer_dim = num_h2,
pool_size = pool_size_2,
layer_name = 'h2',
irange = 1.)
h2.set_biases(rng.uniform(-1., 1., (num_h2,)).astype(config.floatX), recenter=center)
dbm = DBM(visible_layer = v,
hidden_layers = [h1, h2],
batch_size = 1,
niter = 50)
return dbm
def test_variational_cd():
# Verifies that VariationalCD works well with make_layer_to_symbolic_state
visible_layer = BinaryVector(nvis=100)
hidden_layer = BinaryVectorMaxPool(detector_layer_dim=500,
pool_size=1,
layer_name='h',
irange=0.05,
init_bias=-2.0)
model = DBM(visible_layer=visible_layer,
hidden_layers=[hidden_layer],
batch_size=100,
niter=1)
cost = VariationalCD(num_chains=100, num_gibbs_steps=2)
data_specs = cost.get_data_specs(model)
mapping = DataSpecsMapping(data_specs)
space_tuple = mapping.flatten(data_specs[0], return_tuple=True)
source_tuple = mapping.flatten(data_specs[1], return_tuple=True)
theano_args = []
for space, source in safe_zip(space_tuple, source_tuple):
name = '%s' % (source)
arg = space.make_theano_batch(name=name)
theano_args.append(arg)
theano_args = tuple(theano_args)
nested_args = mapping.nest(theano_args)
grads, updates = cost.get_gradients(model, nested_args)
def test_bvmp_make_state():
# Verifies that BinaryVector.make_state creates
# a shared variable whose value passes check_binary_samples
num_pools = 3
num_samples = 1000
tol = .04
rng = np.random.RandomState([2012,11,1,9])
# pool_size=1 is an important corner case
for pool_size in [1, 2, 5]:
n = num_pools * pool_size
layer = BinaryVectorMaxPool(
detector_layer_dim=n,
layer_name='h',
irange=1.,
pool_size=pool_size)
# This is just to placate mf_update below
input_space = VectorSpace(1)
class DummyDBM(object):
def __init__(self):
self.rng = rng
layer.set_dbm(DummyDBM())
layer.set_input_space(input_space)
layer.set_biases(rng.uniform(-pool_size, 1., (n,)).astype(config.floatX))
# To find the mean of the samples, we use mean field with an input of 0
mean = layer.mf_update(
state_below=T.alloc(0., 1, 1),
state_above=None,
layer_above=None)
mean = function([], mean)()
mean = [ mn[0,:] for mn in mean ]
state = layer.make_state(num_examples=num_samples,
numpy_rng=rng)
value = [elem.get_value() for elem in state]
check_bvmp_samples(value, num_samples, n, pool_size, mean, tol)
def test_bvmp_make_state():
# Verifies that BinaryVector.make_state creates
# a shared variable whose value passes check_binary_samples
num_pools = 3
num_samples = 1000
tol = .04
rng = np.random.RandomState([2012,11,1,9])
# pool_size=1 is an important corner case
for pool_size in [1, 2, 5]:
n = num_pools * pool_size
layer = BinaryVectorMaxPool(
detector_layer_dim=n,
layer_name='h',
irange=1.,
pool_size=pool_size)
# This is just to placate mf_update below
input_space = VectorSpace(1)
class DummyDBM(object):
def __init__(self):
self.rng = rng
layer.set_dbm(DummyDBM())
layer.set_input_space(input_space)
layer.set_biases(rng.uniform(-pool_size, 1., (n,)).astype(config.floatX))
# To find the mean of the samples, we use mean field with an input of 0
mean = layer.mf_update(
state_below=T.alloc(0., 1, 1),
state_above=None,
layer_above=None)
mean = function([], mean)()
mean = [ mn[0,:] for mn in mean ]
state = layer.make_state(num_examples=num_samples,
numpy_rng=rng)
value = [elem.get_value() for elem in state]
check_bvmp_samples(value, num_samples, n, pool_size, mean, tol)
def make_random_basic_binary_dbm(
rng,
pool_size_1,
num_vis = None,
num_pool_1 = None,
num_pool_2 = None,
pool_size_2 = None,
center = False
):
"""
Makes a DBM with BinaryVector for the visible layer,
and two hidden layers of type BinaryVectorMaxPool.
The weights and biases are initialized randomly with
somewhat large values (i.e., not what you'd want to
use for learning)
rng: A numpy RandomState.
pool_size_1: The size of the pools to use in the first
layer.
"""
if num_vis is None:
num_vis = rng.randint(1,11)
if num_pool_1 is None:
num_pool_1 = rng.randint(1,11)
if num_pool_2 is None:
num_pool_2 = rng.randint(1,11)
if pool_size_2 is None:
pool_size_2 = rng.randint(1,6)
num_h1 = num_pool_1 * pool_size_1
num_h2 = num_pool_2 * pool_size_2
v = BinaryVector(num_vis, center=center)
v.set_biases(rng.uniform(-1., 1., (num_vis,)).astype(config.floatX), recenter=center)
h1 = BinaryVectorMaxPool(
detector_layer_dim = num_h1,
pool_size = pool_size_1,
layer_name = 'h1',
center = center,
irange = 1.)
h1.set_biases(rng.uniform(-1., 1., (num_h1,)).astype(config.floatX), recenter=center)
h2 = BinaryVectorMaxPool(
center = center,
detector_layer_dim = num_h2,
pool_size = pool_size_2,
layer_name = 'h2',
irange = 1.)
h2.set_biases(rng.uniform(-1., 1., (num_h2,)).astype(config.floatX), recenter=center)
dbm = DBM(visible_layer = v,
hidden_layers = [h1, h2],
batch_size = 1,
niter = 50)
return dbm
def test_ais():
"""
Test ais computation by comparing the output of estimate_likelihood to
Russ's code's output for the same parameters.
"""
try:
# TODO: the one_hot=True is only necessary because one_hot=False is
# broken, remove it after one_hot=False is fixed.
trainset = MNIST(which_set='train', one_hot=True)
testset = MNIST(which_set='test', one_hot=True)
except NoDataPathError:
raise SkipTest("PYLEARN2_DATA_PATH environment variable not defined")
nvis = 784
nhid = 20
# Random initialization of RBM parameters
numpy.random.seed(98734)
w_hid = 10 * numpy.cast[theano.config.floatX](numpy.random.randn(
nvis, nhid))
b_vis = 10 * numpy.cast[theano.config.floatX](numpy.random.randn(nvis))
b_hid = 10 * numpy.cast[theano.config.floatX](numpy.random.randn(nhid))
# Initialization of RBM
visible_layer = BinaryVector(nvis)
hidden_layer = BinaryVectorMaxPool(detector_layer_dim=nhid,
pool_size=1,
layer_name='h',
irange=0.1)
rbm = DBM(100, visible_layer, [hidden_layer], 1)
rbm.visible_layer.set_biases(b_vis)
rbm.hidden_layers[0].set_weights(w_hid)
rbm.hidden_layers[0].set_biases(b_hid)
rbm.nvis = nvis
rbm.nhid = nhid
# Compute real logz and associated train_ll and test_ll using rbm_tools
v_sample = T.matrix('v_sample')
h_sample = T.matrix('h_sample')
W = theano.shared(rbm.hidden_layers[0].get_weights())
hbias = theano.shared(rbm.hidden_layers[0].get_biases())
vbias = theano.shared(rbm.visible_layer.get_biases())
wx_b = T.dot(v_sample, W) + hbias
vbias_term = T.dot(v_sample, vbias)
hidden_term = T.sum(T.log(1 + T.exp(wx_b)), axis=1)
free_energy_v = -hidden_term - vbias_term
free_energy_v_fn = theano.function(inputs=[v_sample],
outputs=free_energy_v)
wh_c = T.dot(h_sample, W.T) + vbias
hbias_term = T.dot(h_sample, hbias)
visible_term = T.sum(T.log(1 + T.exp(wh_c)), axis=1)
free_energy_h = -visible_term - hbias_term
free_energy_h_fn = theano.function(inputs=[h_sample],
outputs=free_energy_h)
real_logz = rbm_tools.compute_log_z(rbm, free_energy_h_fn)
real_ais_train_ll = -rbm_tools.compute_nll(
rbm, trainset.get_design_matrix(), real_logz, free_energy_v_fn)
real_ais_test_ll = -rbm_tools.compute_nll(rbm, testset.get_design_matrix(),
real_logz, free_energy_v_fn)
# Compute train_ll, test_ll and logz using dbm_metrics
train_ll, test_ll, logz = dbm_metrics.estimate_likelihood([W],
[vbias, hbias],
trainset,
testset,
pos_mf_steps=100)
assert (real_logz - logz) < 2.0
assert (real_ais_train_ll - train_ll) < 2.0
assert (real_ais_test_ll - test_ll) < 2.0
