def get_numeric_gradient(stochastic, pvalue):
e = 1e-9
initial_value = pvalue.value
i_shape = shape(initial_value)
i_size = size(initial_value)
initial_logp = utils.logp_of_set(stochastic)
numeric_gradient = zeros(i_shape)
if not (pvalue.dtype in float_dtypes):
return numeric_gradient
for i in range(i_size):
delta = zeros(i_shape)
try:
delta[unravel_index(i, i_shape or (1, ))] += e
except IndexError:
delta += e
pvalue.value = initial_value + delta
logp = utils.logp_of_set(stochastic)
try:
numeric_gradient[unravel_index(i, i_shape or
(1, ))] = (logp - initial_logp) / e
except IndexError:
numeric_gradient = (logp - initial_logp) / e
pvalue.value = initial_value
return numeric_gradient
def get_numeric_gradient(stochastic, pvalue):
e = 1e-9
initial_value = pvalue.value
i_shape = shape(initial_value)
i_size = size(initial_value)
initial_logp = utils.logp_of_set(stochastic)
numeric_gradient = zeros(i_shape)
if not (pvalue.dtype in float_dtypes):
return numeric_gradient
for i in range(i_size):
delta = zeros(i_shape)
try:
delta[unravel_index(i, i_shape or (1,))] += e
except IndexError:
delta += e
pvalue.value = initial_value + delta
logp = utils.logp_of_set(stochastic)
try:
numeric_gradient[unravel_index(i, i_shape or (1,))] = (logp - initial_logp) / e
except IndexError:
numeric_gradient = (logp - initial_logp) / e
pvalue.value = initial_value
return numeric_gradient
def test_other_err(self):
self.B.rand()
for i in xrange(1000):
try:
utils.logp_of_set(set([self.A,self.B,self.D,self.E]))
except:
cls, inst, tb = sys.exc_info()
assert(cls is RuntimeError)
def test_ZeroProb(self):
self.B.value = -1
for i in xrange(1000):
try:
utils.logp_of_set(set([self.A,self.B,self.D, self.E]))
except:
cls, inst, tb = sys.exc_info()
assert(cls is ZeroProbability)
def test_other_err(self):
self.B.rand()
for i in xrange(1000):
try:
utils.logp_of_set(set([self.A, self.B, self.D, self.E]))
except:
cls, inst, tb = sys.exc_info()
assert (cls is RuntimeError)
def test_ZeroProb(self):
self.B.value = -1
for i in xrange(1000):
try:
utils.logp_of_set(set([self.A, self.B, self.D, self.E]))
except:
cls, inst, tb = sys.exc_info()
assert (cls is ZeroProbability)
def test_logp(self):
self.B.rand()
lp1 = utils.logp_of_set(set([self.A,self.B,self.D]))
assert_equal(lp1, self.A.logp+self.B.logp+self.D.logp)
def test_logp(self):
self.B.rand()
lp1 = utils.logp_of_set(set([self.A, self.B, self.D]))
assert_almost_equal(lp1, self.A.logp + self.B.logp + self.D.logp, 10)