def _init_uw_global_shared(start, global_RVs, global_order):
start = {v.name: start[v.name] for v in global_RVs}
bij = DictToArrayBijection(global_order, start)
u_start = bij.map(start)
w_start = np.zeros_like(u_start)
uw_start = np.concatenate([u_start, w_start])
uw_global_shared = theano.shared(uw_start, 'uw_global_shared')
return uw_global_shared, bij
def _init_uw_global_shared(start, global_RVs, global_order):
start = {v.name: start[v.name] for v in global_RVs}
bij = DictToArrayBijection(global_order, start)
u_start = bij.map(start)
w_start = np.zeros_like(u_start)
uw_start = np.concatenate([u_start, w_start])
uw_global_shared = theano.shared(uw_start, 'uw_global_shared')
return uw_global_shared, bij
def test_list_ordering(self):
a = num.random.rand(10).reshape((5, 2))
b = num.random.rand(5).reshape((5, 1))
c = num.random.rand(1).reshape((1, 1))
ta = tt.matrix('a')
tb = tt.matrix('b')
tc = tt.matrix('c')
ta.tag.test_value = a
tb.tag.test_value = b
tc.tag.test_value = c
tvars = [ta, tb, tc]
with self.assertRaises(KeyError):
ordering = utility.ListArrayOrdering(tvars)
ordering['b']
lordering = utility.ListArrayOrdering(tvars, 'tensor')
lordering['b'].slc
for var in ordering:
print var
lpoint = [a, b, c]
lij = utility.ListToArrayBijection(lordering, lpoint)
ref_point = {'a': a, 'b':b, 'c':c}
array = lij.l2a(lpoint)
point = lij.l2d(lpoint)
print 'arr', array
#print 'point, ref_point', point, ref_point
print lij.l2d(lij.a2l(array))
ordering = ArrayOrdering(tvars)
bij = DictToArrayBijection(ordering, point)
array2 = bij.map(point)
print 'bija', array2
def _setup(self, **kwargs):
self._histogram_order = ArrayOrdering(self.global_vars)
self._bij = DictToArrayBijection(self._histogram_order, dict())
class Histogram(Approximation):
"""
Builds Approximation instance from a given trace,
it has the same interface as variational approximation
Parameters
----------
trace : MultiTrace
local_rv : dict
Experimental for Histogram
mapping {model_variable -> local_variable}
Local Vars are used for Autoencoding Variational Bayes
See (AEVB; Kingma and Welling, 2014) for details
model : PyMC3 model
Usage
-----
>>> with model:
... step = NUTS()
... trace = sample(1000, step=step)
... histogram = Histogram(trace[100:])
"""
def __init__(self, trace, local_rv=None, model=None):
super(Histogram, self).__init__(local_rv=local_rv,
model=model,
trace=trace)
def check_model(self, model, **kwargs):
trace = kwargs.get('trace')
if (trace is not None and
not all([var.name in trace.varnames
for var in model.free_RVs])):
raise ValueError('trace has not all FreeRV')
def _setup(self, **kwargs):
self._histogram_order = ArrayOrdering(self.global_vars)
self._bij = DictToArrayBijection(self._histogram_order, dict())
def create_shared_params(self, **kwargs):
trace = kwargs.get('trace')
if trace is None:
histogram = np.atleast_2d(self._bij.map(self.model.test_point))
else:
histogram = np.empty((len(trace), self.global_size))
for i in range(len(trace)):
histogram[i] = self._bij.map(trace[i])
return theano.shared(pm.floatX(histogram), 'histogram')
def randidx(self, size=None):
if size is None:
size = ()
elif isinstance(size, tt.TensorVariable):
if size.ndim < 1:
size = size[None]
elif size.ndim > 1:
raise ValueError('size ndim should be no more than 1d')
else:
pass
else:
size = tuple(np.atleast_1d(size))
return (tt_rng().uniform(size=size,
low=0.0,
high=self.histogram.shape[0] -
1e-16).astype('int64'))
def random_global(self, size=None, no_rand=False):
theano_condition_is_here = isinstance(no_rand, tt.Variable)
if theano_condition_is_here:
return tt.switch(no_rand, self.mean,
self.histogram[self.randidx(size)])
else:
if no_rand:
return self.mean
else:
return self.histogram[self.randidx(size)]
@property
def histogram(self):
"""
Shortcut to flattened Trace
"""
return self.shared_params
@property
@memoize
def histogram_logp(self):
"""
Symbolic logp for every point in trace
"""
node = self.to_flat_input(self.model.logpt)
def mapping(z):
return theano.clone(node, {self.input: z})
x = self.histogram
_histogram_logp, _ = theano.scan(mapping, x, n_steps=x.shape[0])
return _histogram_logp
@property
def mean(self):
return self.histogram.mean(0)
@classmethod
def from_noise(cls,
size,
jitter=.01,
local_rv=None,
start=None,
model=None):
"""
Initialize Histogram with random noise
Parameters
----------
size : number of initial particles
jitter : initial sd
local_rv : dict
mapping {model_variable -> local_variable}
Local Vars are used for Autoencoding Variational Bayes
See (AEVB; Kingma and Welling, 2014) for details
start : initial point
model : pm.Model
PyMC3 Model
Returns
-------
Histogram
"""
hist = cls(None, local_rv=local_rv, model=model)
if start is None:
start = hist.model.test_point
start = hist._bij.map(start)
# Initialize particles
x0 = np.tile(start, (size, 1))
x0 += np.random.normal(0, jitter, x0.shape)
hist.histogram.set_value(x0)
return hist
class Histogram(Approximation):
"""
Builds Approximation instance from a given trace,
it has the same interface as variational approximation
Prameters
----------
trace : MultiTrace
local_rv : dict
Experimental for Histogram
mapping {model_variable -> local_variable}
Local Vars are used for Autoencoding Variational Bayes
See (AEVB; Kingma and Welling, 2014) for details
model : PyMC3 model
Usage
-----
>>> with model:
... step = NUTS()
... trace = sample(1000, step=step)
... histogram = Histogram(trace[100:])
"""
def __init__(self, trace, local_rv=None, model=None):
self.trace = trace
super(Histogram, self).__init__(local_rv=local_rv, model=model)
def check_model(self, model):
if not all([var.name in self.trace.varnames
for var in model.free_RVs]):
raise ValueError('trace has not all FreeRV')
def _setup(self):
self._histogram_order = ArrayOrdering(self.global_vars)
self._bij = DictToArrayBijection(self._histogram_order, dict())
def create_shared_params(self):
trace = self.trace
histogram = np.empty((len(trace), self.global_size))
for i in range(len(trace)):
histogram[i] = self._bij.map(trace[i])
return theano.shared(histogram, 'histogram')
def randidx(self, size=None):
if size is None:
size = ()
elif isinstance(size, tt.TensorVariable):
if size.ndim < 1:
size = size[None]
elif size.ndim > 1:
raise ValueError('size ndim should be no more than 1d')
else:
pass
else:
size = tuple(np.atleast_1d(size))
return (tt_rng()
.uniform(size=size, low=0.0, high=self.histogram.shape[0] - 1e-16)
.astype('int64'))
def random_global(self, size=None, no_rand=False):
theano_condition_is_here = isinstance(no_rand, tt.Variable)
if theano_condition_is_here:
return tt.switch(no_rand,
self.mean,
self.histogram[self.randidx(size)])
else:
if no_rand:
return self.mean
else:
return self.histogram[self.randidx(size)]
@property
def histogram(self):
"""
Shortcut to flattened Trace
"""
return self.shared_params
@property
@memoize
def histogram_logp(self):
"""
Symbolic logp for every point in trace
"""
node = self.to_flat_input(self.model.logpt)
def mapping(z):
return theano.clone(node, {self.input: z})
x = self.histogram
_histogram_logp, _ = theano.scan(
mapping, x, n_steps=x.shape[0]
)
return _histogram_logp
@property
def mean(self):
return self.histogram.mean(0)
@property
def params(self):
return []