def test_get_weights():
#Tests that the RBM, when constructed
#with nvis and nhid arguments, supports the
#weights interface
model = RBM(nvis = 2, nhid = 3)
W = model.get_weights()
def test_get_weights():
#Tests that the RBM, when constructed
#with nvis and nhid arguments, supports the
#weights interface
model = RBM(nvis=2, nhid=3)
W = model.get_weights()
def load_matlab_dbm(path, num_chains = 1):
"""
.. todo::
WRITEME properly
Loads a two layer DBM stored in the format used by Ruslan Salakhutdinov's
matlab demo
"""
d = io.loadmat(path)
for key in d:
try:
d[key] = np.cast[config.floatX](d[key])
except:
pass
visbiases = d['visbiases']
assert len(visbiases.shape) == 2
assert visbiases.shape[0] == 1
visbiases = visbiases[0,:]
hidbiases = d['hidbiases']
assert len(hidbiases.shape) == 2
assert hidbiases.shape[0] == 1
hidbiases = hidbiases[0,:]
penbiases = d['penbiases']
assert len(penbiases.shape) == 2
assert penbiases.shape[0] == 1
penbiases = penbiases[0,:]
vishid = d['vishid']
hidpen = d['hidpen']
D ,= visbiases.shape
N1 ,= hidbiases.shape
N2 ,= penbiases.shape
assert vishid.shape == (D,N1)
assert hidpen.shape == (N1,N2)
rbms = [ RBM( nvis = D, nhid = N1),
RBM( nvis = N1, nhid = N2) ]
dbm = DBM(rbms, negative_chains = num_chains)
dbm.bias_vis.set_value(visbiases)
dbm.bias_hid[0].set_value(hidbiases)
dbm.bias_hid[1].set_value(penbiases)
dbm.W[0].set_value(vishid)
dbm.W[1].set_value(hidpen)
return dbm
def test_gibbs_step_for_v():
# Just tests that gibbs_step_for_v can be called
# without crashing
model = RBM(nvis=2, nhid=3)
theano_rng = make_theano_rng(17, which_method='binomial')
X = T.matrix()
Y = model.gibbs_step_for_v(X, theano_rng)
def test_gibbs_step_for_v():
#Just tests that gibbs_step_for_v can be called
#without crashing (protection against refactoring
#damage, aren't interpreted languages great?)
model = RBM(nvis=2, nhid=3)
theano_rng = RandomStreams(17)
X = T.matrix()
Y = model.gibbs_step_for_v(X, theano_rng)
def test_gibbs_step_for_v():
#Just tests that gibbs_step_for_v can be called
#without crashing (protection against refactoring
#damage, aren't interpreted languages great?)
model = RBM(nvis = 2, nhid = 3)
theano_rng = make_theano_rng(17, which_method='binomial')
X = T.matrix()
Y = model.gibbs_step_for_v(X, theano_rng)
def get_rbm(structure):
n_input, n_output = structure
config = {
'nvis': n_input,
'nhid': n_output,
"irange": 0.05,
"init_bias_hid": 0.0,
"init_bias_vis": 0.0,
}
return RBM(**config)
def test_train_batch():
# Just tests that train_batch can be called without crashing
m = 1
dim = 2
rng = np.random.RandomState([2014, 03, 17])
X = rng.randn(m, dim)
train = DenseDesignMatrix(X=X)
rbm = RBM(nvis=dim, nhid=3)
trainer = DefaultTrainingAlgorithm(batch_size=1, batches_per_iter=10)
trainer.setup(rbm, train)
trainer.train(train)
def __init__(self):
""" gets a small batch of data
sets up a PD-DBM model
"""
self.tol = 1e-5
X = np.random.RandomState([1,2,3]).randn(1000,5)
X -= X.mean()
X /= X.std()
m, D = X.shape
N = 6
N2 = 7
s3c = S3C(nvis = D,
nhid = N,
irange = .1,
init_bias_hid = -1.5,
init_B = 3.,
min_B = 1e-8,
max_B = 1000.,
init_alpha = 1., min_alpha = 1e-8, max_alpha = 1000.,
init_mu = 1., e_step = None,
m_step = Grad_M_Step(),
min_bias_hid = -1e30, max_bias_hid = 1e30,
)
rbm = RBM(nvis = N, nhid = N2, irange = .1, init_bias_vis = -1.5, init_bias_hid = 1.5)
#don't give the model an inference procedure or learning rate so it won't spend years compiling a learn_func
self.model = PDDBM(
dbm = DBM( use_cd = 1,
rbms = [ rbm ]),
s3c = s3c
)
self.model.make_pseudoparams()
self.inference_procedure = InferenceProcedure(
clip_reflections = True,
rho = .5 )
self.inference_procedure.register_model(self.model)
self.X = X
self.N = N
self.N2 = N2
self.m = m
def test_unspecified_batch_size():
# Test that failing to specify the batch size results in a
# NoBatchSizeError
m = 1
dim = 2
rng = np.random.RandomState([2014, 03, 17])
X = rng.randn(m, dim)
train = DenseDesignMatrix(X=X)
rbm = RBM(nvis=dim, nhid=3)
trainer = DefaultTrainingAlgorithm()
try:
trainer.setup(rbm, train)
except NoBatchSizeError:
return
raise AssertionError("Missed the lack of a batch size")
def __init__(self, model = None, X = None, tol = 1e-5,
init_H = None, init_S = None, init_G = None):
""" gets a small batch of data
sets up a PD-DBM model
"""
self.tol = tol
if X is None:
X = np.random.RandomState([1,2,3]).randn(1000,5)
X -= X.mean()
X /= X.std()
m, D = X.shape
if model is None:
N = 6
N2 = 7
s3c = S3C(nvis = D,
nhid = N,
irange = .1,
init_bias_hid = -1.5,
init_B = 3.,
min_B = 1e-8,
max_B = 1000.,
init_alpha = 1., min_alpha = 1e-8, max_alpha = 1000.,
init_mu = 1., e_step = None,
m_step = Grad_M_Step(),
min_bias_hid = -1e30, max_bias_hid = 1e30,
)
rbm = RBM(nvis = N, nhid = N2, irange = .5, init_bias_vis = -1.5, init_bias_hid = 1.5)
#don't give the model an inference procedure or learning rate so it won't spend years compiling a learn_func
self.model = PDDBM(
dbm = DBM( use_cd = 1,
rbms = [ rbm ]),
s3c = s3c
)
self.model.make_pseudoparams()
self.inference_procedure = InferenceProcedure(
clip_reflections = True,
rho = .5 )
self.inference_procedure.register_model(self.model)
else:
self.model = model
self.inference_procedure = model.inference_procedure
N = model.s3c.nhid
N2 = model.dbm.rbms[0].nhid
self.X = X
self.N = N
self.N2 = N2
self.m = m
if init_H is None:
self.init_H = np.cast[config.floatX](self.model.rng.uniform(0.,1.,(self.m, self.N)))
self.init_S = np.cast[config.floatX](self.model.rng.uniform(-5.,5.,(self.m, self.N)))
self.init_G = np.cast[config.floatX](self.model.rng.uniform(0.,1.,(self.m,self.N2)))
else:
assert init_S is not None
assert init_G is not None
self.init_H = init_H
self.init_S = init_S
self.init_G = init_G
for i in xrange(2 ** D):
all_states.append(int_to_bits(i))
all_states = np.asarray(all_states, dtype = 'float32')
good_states = []
for i in xrange(D):
good_states.append(int_to_bits(2 ** i) )
good_states = np.asarray(good_states, dtype = 'float32')
all_states_var = T.matrix()
rbm = RBM( nvis = D, nhid = D)
Z = T.exp( - rbm.free_energy_given_v(all_states_var) ).sum()
good_states_var = T.matrix()
good_prob = T.exp( - rbm.free_energy_given_v(good_states_var) ).sum() / Z
good_prob_func = function([all_states_var, good_states_var], good_prob)
def run_rbm( pos_weight = 1., neg_weight = 1., bias_hid = -1.,
bias_vis = inverse_sigmoid_numpy( 1. / float(D) ) ):
rbm.bias_vis.set_value( np.ones( (D,), dtype='float32') * bias_vis)
rbm.bias_hid.set_value( np.ones( (D,), dtype='float32') * bias_hid)
rbm.transformer._W.set_value( np.identity(D, dtype='float32') * \
(pos_weight + neg_weight) - np.ones( (D,D), dtype = 'float32') \
def test_get_input_space():
model = RBM(nvis = 2, nhid = 3)
space = model.get_input_space()
def test_get_weights():
model = RBM(nvis = 2, nhid = 3)
W = model.get_weights()
def test_get_input_space():
#Tests that the RBM supports
#the Space interface
model = RBM(nvis=2, nhid=3)
space = model.get_input_space()
for i in xrange(2**D):
all_states.append(int_to_bits(i))
all_states = np.asarray(all_states, dtype='float32')
good_states = []
for i in xrange(D):
good_states.append(int_to_bits(2**i))
good_states = np.asarray(good_states, dtype='float32')
all_states_var = T.matrix()
rbm = RBM(nvis=D, nhid=D)
Z = T.exp(-rbm.free_energy_given_v(all_states_var)).sum()
good_states_var = T.matrix()
good_prob = T.exp(-rbm.free_energy_given_v(good_states_var)).sum() / Z
good_prob_func = function([all_states_var, good_states_var], good_prob)
def run_rbm(pos_weight=1.,
neg_weight=1.,
bias_hid=-1.,
bias_vis=inverse_sigmoid_numpy(1. / float(D))):
rbm.bias_vis.set_value(np.ones((D, ), dtype='float32') * bias_vis)
def __init__(self):
""" gets a small batch of data
sets up a PD-DBM model with its DBM weights set to 0
so that it represents the same distribution as an S3C
model
Makes the S3C model it matches
(Note that their learning rules don't match since the
complete partition function of the S3C model is tractable
but the PD-DBM has to approximate the h partition function
via sampling)
"""
self.tol = 1e-5
X = np.random.RandomState([1,2,3]).randn(1000,5)
X -= X.mean()
X /= X.std()
m, D = X.shape
N = 6
N2 = 7
self.X = X
self.N = N
self.N2 = N2
self.m = m
self.D = D
s3c_for_pddbm = self.make_s3c()
self.s3c = self.make_s3c()
self.s3c.W.set_value(s3c_for_pddbm.W.get_value())
rbm = RBM(nvis = N, nhid = N2, irange = .0, init_bias_vis = -1.5, init_bias_hid = 6.)
#don't give the model an inference procedure or learning rate so it won't spend years compiling a learn_func
self.model = PDDBM(
dbm = DBM( use_cd = 1,
rbms = [ rbm ]),
s3c = s3c_for_pddbm
)
self.model.make_pseudoparams()
self.inference_procedure = InferenceProcedure(
schedule = [ ['s',.1], ['h',.1], ['g',0, 0.2], ['s', 0.2], ['h',0.2],
['s',0.3], ['g',0,.3], ['h',0.3], ['s',0.4], ['h',0.4],
['g',0,.4], ['s',0.4], ['h',0.4],
['g',0,.5], ['s',0.5], ['h', 0.5], ['s',0.1],
['h',0.5] ],
clip_reflections = True,
rho = .5 )
self.inference_procedure.register_model(self.model)
self.e_step = make_e_step_from_inference_procedure(self.inference_procedure)
self.e_step.register_model(self.s3c)
self.s3c.make_pseudoparams()
#check that all the parameters match
assert np.abs(self.s3c.W.get_value() - self.model.s3c.W.get_value()).max() == 0.0
assert np.abs(self.s3c.bias_hid.get_value() - self.model.s3c.bias_hid.get_value()).max() == 0.0
assert np.abs(self.s3c.alpha.get_value() - self.model.s3c.alpha.get_value()).max() == 0.0
assert np.abs(self.s3c.mu.get_value() - self.model.s3c.mu.get_value()).max() == 0.0
assert np.abs(self.s3c.B_driver.get_value() - self.model.s3c.B_driver.get_value()).max() == 0.0
#check that the assumptions making these tests valid are met
assert np.abs(self.model.dbm.W[0].get_value()).max() == 0.0
def get_model(self):
self.model = RBM(nvis=self.vis, nhid=self.hid, irange=.05)
return self.model
def test_get_input_space():
#Tests that the RBM supports
#the Space interface
model = RBM(nvis = 2, nhid = 3)
space = model.get_input_space()
