def __init__(self, activation_fcn, rng=None, shape=None):
"""
:param activation_fcn: A string identifying the type of activation function.
{'bernoulli', 'gaussian', 'adaptive_gaussian', 'rect-lin'}
:param rng: Numpy random number generator for the stochastic component
:param shape: Optionally, reshape the output to this shape.
"""
rng = RandomStreams(rng.randint(1e9) if rng is not None else None)
self.activation_fcn = activation_fcn
self._smooth_activation_fcn, self._stochastic_activation_fcn, self._free_energy_fcn, self._params = \
self._stochastic_layer_name_to_functions(activation_fcn, rng)
self._shape = shape
def __init__(self, activation_fcn, rng = None, shape = None):
"""
:param activation_fcn: A string identifying the type of activation function.
{'bernoulli', 'gaussian', 'adaptive_gaussian', 'rect-lin'}
:param rng: Numpy random number generator for the stochastic component
:param shape: Optionally, reshape the output to this shape.
"""
rng = RandomStreams(rng.randint(1e9) if rng is not None else None)
self.activation_fcn = activation_fcn
self._smooth_activation_fcn, self._stochastic_activation_fcn, self._free_energy_fcn, self._params = \
self._stochastic_layer_name_to_functions(activation_fcn, rng)
self._shape = shape
def test_indeterministic_reconstruct_scan_vs_theano(self):
self.setUpRBM()
self.assertTrue(self.rbm.h_n == 10)
rbm = self.rbm
W = rbm.W.get_value(borrow=True)
U = rbm.U.get_value(borrow=True)
vb1 = rbm.v_bias.eval()
vb2 = rbm.v_bias2.eval()
hb = rbm.h_bias.eval()
k = 100
# Initial values
rand = np.random.RandomState(123)
rand = RandomStreams(rand.randint(2 ** 30))
x1 = self.rbmx1
x2 = rand.binomial(size=self.rbmx2.shape, n=1, p=0.5, dtype=t_float_x).eval()
def gibbs(ux, u2):
h, hp = rbm.prop_up(ux, u2)
hs = rbm.rand.binomial(size=hp.shape, n=1, p=hp, dtype=t_float_x)
v, vp = rbm.prop_down(hs)
vs = rbm.rand.binomial(size=vp.shape, n=1, p=vp, dtype=t_float_x)
v2, v2p = rbm.prop_down_assoc(hs)
v2s = rbm.rand.binomial(size=v2p.shape, n=1, p=v2p, dtype=t_float_x)
return [h, hp, hs, v, v2p, ux, v2, v2p, v2s]
# THEANO
x = T.dmatrix("x")
y = T.dmatrix("y")
x_start = x
y_start = y
(
res,
updates
) = theano.scan(
gibbs,
outputs_info=[None, None, None, None, None,
x_start, None, None, y_start],
n_steps=k
)
f = theano.function([x, y], res, updates=updates)
rand = np.random.RandomState(1234)
rand = RandomStreams(rand.randint(2 ** 30))
rbm.rand = rand
[h, hp, hs, v, vp, vs, v2, v2p, v2s] = f(self.rbmx1, x2)
# print h
# print hp
# print "h: \n{}".format(h)
# print "hp: \n{}".format(hp)
# print "hs: \n{}".format(hs)
# print "v: \n{}".format(v)
# print "vp: \n{}".format(vp)
# print "vs: \n{}".format(vs)
# print "v2: \n{}".format(v)
# print "v2p: \n{}".format(v2p)
# print "v2s: \n{}".format(v2s)
# =============== NUMPY ================
rand = np.random.RandomState(1234)
rand = RandomStreams(rand.randint(2 ** 30))
for i in xrange(0, k):
# Sample h
h, ph = np_prop_up(x1, W, hb, x2, U)
# sample using same seed
hs = rand.binomial(size=ph.shape, n=1, p=ph, dtype=t_float_x).eval()
# print h
# Sample x, x2
u, pu = np_prop_down(hs, W, vb1)
# dummy call, just to adjust seed
us = rand.binomial(size=pu.shape, n=1, p=pu, dtype=t_float_x).eval()
u2, pu2 = np_prop_down(hs, U, vb2)
x2 = pu2
x2 = rand.binomial(size=pu2.shape, n=1, p=pu2, dtype=t_float_x).eval()
def test_indeterministic_reconstruct_scan_vs_theano(self):
self.setUpRBM()
self.assertTrue(self.rbm.h_n == 10)
rbm = self.rbm
W = rbm.W.get_value(borrow=True)
U = rbm.U.get_value(borrow=True)
vb1 = rbm.v_bias.eval()
vb2 = rbm.v_bias2.eval()
hb = rbm.h_bias.eval()
k = 100
# Initial values
rand = np.random.RandomState(123)
rand = RandomStreams(rand.randint(2**30))
x1 = self.rbmx1
x2 = rand.binomial(size=self.rbmx2.shape, n=1, p=0.5,
dtype=t_float_x).eval()
def gibbs(ux, u2):
h, hp = rbm.prop_up(ux, u2)
hs = rbm.rand.binomial(size=hp.shape, n=1, p=hp, dtype=t_float_x)
v, vp = rbm.prop_down(hs)
vs = rbm.rand.binomial(size=vp.shape, n=1, p=vp, dtype=t_float_x)
v2, v2p = rbm.prop_down_assoc(hs)
v2s = rbm.rand.binomial(size=v2p.shape,
n=1,
p=v2p,
dtype=t_float_x)
return [h, hp, hs, v, v2p, ux, v2, v2p, v2s]
# THEANO
x = T.dmatrix("x")
y = T.dmatrix("y")
x_start = x
y_start = y
(res, updates) = theano.scan(gibbs,
outputs_info=[
None, None, None, None, None, x_start,
None, None, y_start
],
n_steps=k)
f = theano.function([x, y], res, updates=updates)
rand = np.random.RandomState(1234)
rand = RandomStreams(rand.randint(2**30))
rbm.rand = rand
[h, hp, hs, v, vp, vs, v2, v2p, v2s] = f(self.rbmx1, x2)
# print h
# print hp
# print "h: \n{}".format(h)
# print "hp: \n{}".format(hp)
# print "hs: \n{}".format(hs)
# print "v: \n{}".format(v)
# print "vp: \n{}".format(vp)
# print "vs: \n{}".format(vs)
# print "v2: \n{}".format(v)
# print "v2p: \n{}".format(v2p)
# print "v2s: \n{}".format(v2s)
# =============== NUMPY ================
rand = np.random.RandomState(1234)
rand = RandomStreams(rand.randint(2**30))
for i in xrange(0, k):
# Sample h
h, ph = np_prop_up(x1, W, hb, x2, U)
# sample using same seed
hs = rand.binomial(size=ph.shape, n=1, p=ph,
dtype=t_float_x).eval()
# print h
# Sample x, x2
u, pu = np_prop_down(hs, W, vb1)
# dummy call, just to adjust seed
us = rand.binomial(size=pu.shape, n=1, p=pu,
dtype=t_float_x).eval()
u2, pu2 = np_prop_down(hs, U, vb2)
x2 = pu2
x2 = rand.binomial(size=pu2.shape, n=1, p=pu2,
dtype=t_float_x).eval()
