def predict(self, Xnew, point=None, diag=False, pred_noise=False, given=None):
R"""
Return the mean vector and covariance matrix of the conditional
distribution as numpy arrays, given a `point`, such as the MAP
estimate or a sample from a `trace`.
Parameters
----------
Xnew : array-like
Function input values. If one-dimensional, must be a column
vector with shape `(n, 1)`.
point : pymc3.model.Point
A specific point to condition on.
diag : bool
If `True`, return the diagonal instead of the full covariance
matrix. Default is `False`.
pred_noise : bool
Whether or not observation noise is included in the conditional.
Default is `False`.
given : dict
Same as `conditional` method.
"""
if given is None:
given = {}
mu, cov = self.predictt(Xnew, diag, pred_noise, given)
return draw_values([mu, cov], point=point)
def astep(self, q0, logp):
"""q0: current state
logp: log probability function
"""
# Draw from the normal prior by multiplying the Cholesky decomposition
# of the covariance with draws from a standard normal
chol = draw_values([self.prior_chol])[0]
nu = np.dot(chol, nr.randn(chol.shape[0]))
y = logp(q0) - nr.standard_exponential()
# Draw initial proposal and propose a candidate point
theta = nr.uniform(0, 2 * np.pi)
theta_max = theta
theta_min = theta - 2 * np.pi
q_new = q0 * np.cos(theta) + nu * np.sin(theta)
while logp(q_new) <= y:
# Shrink the bracket and propose a new point
if theta < 0:
theta_min = theta
else:
theta_max = theta
theta = nr.uniform(theta_min, theta_max)
q_new = q0 * np.cos(theta) + nu * np.sin(theta)
return q_new
def predict(self,
Xnew,
point=None,
diag=False,
pred_noise=False,
given=None):
R"""
Return the mean vector and covariance matrix of the conditional
distribution as numpy arrays, given a `point`, such as the MAP
estimate or a sample from a `trace`.
Parameters
----------
Xnew : array-like
Function input values. If one-dimensional, must be a column
vector with shape `(n, 1)`.
point : pymc3.model.Point
A specific point to condition on.
diag : bool
If `True`, return the diagonal instead of the full covariance
matrix. Default is `False`.
pred_noise : bool
Whether or not observation noise is included in the conditional.
Default is `False`.
given : dict
Same as `conditional` method.
"""
if given is None:
given = {}
mu, cov = self.predictt(Xnew, diag, pred_noise, given)
return draw_values([mu, cov], point=point)
def __init__(self, *args, **kwargs):
add_citations_to_model(self.__citations__, kwargs.get("model", None))
# Make sure that the shape is compatible
shape = kwargs.get("shape", 2)
try:
if list(shape)[0] != 2:
raise ValueError("the first dimension should be exactly 2")
except TypeError:
if shape != 2:
raise ValueError("the first dimension should be exactly 2")
self.min_radius = kwargs.pop("min_radius", 0)
self.max_radius = kwargs.pop("max_radius", 1)
transform = tr.radius_impact(self.min_radius, self.max_radius)
kwargs["shape"] = shape
kwargs["transform"] = transform
super(RadiusImpact, self).__init__(*args, **kwargs)
# Work out some reasonable starting values for the parameters
default = np.zeros(shape)
mn, mx = draw_values([self.min_radius - 0., self.max_radius - 0.])
default[0] = 0.5 * (mn + mx)
default[1] = 0.5
self._default = default
def random(self, point=None, size=None):
"""Random sample, used in `pm.sample_posterior_predictive()`."""
a_const, a_init, a_ar, a_packed_chol, a_chol, a_cov = draw_values(
[
self.t_const,
self.t_init,
self.t_ar,
self.packed_chol,
self.chol,
self.cov,
],
point=point,
size=size,
)
T = int(self.shape[0])
def gen_1():
my_varx_gen = VARXGenerator(
endog=self.n_endog,
lag_endog=self.n_lag_endog,
coef_ar=a_ar,
coef_covariance=a_cov,
coef_const=a_const,
# This seems correct - order is 'oldest' to 'newest':
coef_initial_values=a_init,
# Ignore nonstationary since these are samples:
check_kwargs=dict(action_nonstationary="ignore"),
)
srs = my_varx_gen.generate(T)["output"].values
return srs
# TODO: what to do with `size`?
random_samples = gen_1()
return random_samples
def random(self, point=None, size=None):
(ror, ) = draw_values([self.ror], point=point, size=size)
return generate_samples(
self._random,
dist_shape=self.shape,
broadcast_shape=self.shape,
size=size,
)
def random(self, point=None, size=None):
mn, mx = draw_values([self.min_radius - 0., self.max_radius - 0.],
point=point)
return generate_samples(self._random,
mn,
mx,
dist_shape=self.shape,
broadcast_shape=self.shape,
size=size)
def _draw_from_parents(self, parent, point=None, size=None):
"""
A wrapper over PyMC3's distribution.draw_values() method that
generalizes drawing random samples from the parents of this
distribution to cases where the parents can be PyMC3 RVs or python
vars (ints, floats, numpy arrays, etc).
:param parent: parent RV (or python var)
:param point: value of this RV at which to draw samples.
(Recommendation: Use default None)
:param size: number of samples requested (defaults to 1)
:return: a set of random samples from each parent (RV or python var)
of this RV
"""
if size is None:
size = 1
# static or dynamic?
if self._is_dynamic_parent(parent):
p = parent["node"]
elif self._is_static_parent(parent):
p = parent
# if this is just the previous timeslice value of this node,
# then fill in with self.testval (which is also the initial value)
# and tile to get the correct dimension
if self._is_me(parent):
val = self.testval[np.newaxis, :]
val = np.repeat(val, size, axis=0)
# if this is a numpy array, return the array broadcasted to one
# dimension higher. The new dimension is now the 0th dimension and is
# the number of samples requested, i.e., size
elif isinstance(p, np.ndarray):
val = p[np.newaxis, :]
val = np.repeat(val, size, axis=0)
# for all other cases, uses pymc3's internal draw_values API
else:
val = draw_values([p], point=point, size=size)[0]
return val
def forward_val(self, x, point=None):
p = x[0]
b = x[1]
pl, Ar, dr = draw_values([self.min_radius-0., self.Ar-0., self.dr-0.],
point=point)
m = b <= 1
r = np.empty_like(x)
r[0, m] = (b[m] / (1 + pl) - 1) * (1 - Ar) + 1
r[1, m] = (p[m] - pl) / dr
arg = p[~m] - b[~m] - dr + 1
q1 = (arg / dr) ** 2
q2 = (pl - b[~m] + 1) / arg
r[0, ~m] = q1 * Ar
r[1, ~m] = q2
return np.log(r) - np.log(1 - r)
def __init__(self, vars, proposal="uniform", order="random", model=None):
model = pm.modelcontext(model)
vars = pm.inputvars(vars)
dimcats = []
# The above variable is a list of pairs (aggregate dimension, number
# of categories). For example, if vars = [x, y] with x being a 2-D
# variable with M categories and y being a 3-D variable with N
# categories, we will have dimcats = [(0, M), (1, M), (2, N), (3, N), (4, N)].
for v in vars:
distr = getattr(v.distribution, "parent_dist", v.distribution)
if isinstance(distr, pm.Categorical):
k = draw_values([distr.k])[0]
elif isinstance(distr, pm.Bernoulli) or (v.dtype in pm.bool_types):
k = 2
else:
raise ValueError(
"All variables must be categorical or binary" +
"for CategoricalGibbsMetropolis")
start = len(dimcats)
dimcats += [(dim, k) for dim in range(start, start + v.dsize)]
if order == "random":
self.shuffle_dims = True
self.dimcats = dimcats
else:
if sorted(order) != list(range(len(dimcats))):
raise ValueError("Argument 'order' has to be a permutation")
self.shuffle_dims = False
self.dimcats = [dimcats[j] for j in order]
if proposal == "uniform":
self.astep = self.astep_unif
elif proposal == "proportional":
# Use the optimized "Metropolized Gibbs Sampler" described in Liu96.
self.astep = self.astep_prop
else:
raise ValueError(
"Argument 'proposal' should either be 'uniform' or 'proportional'"
)
super().__init__(vars, [model.fastlogp])
def random(self, point=None, size=None):
"""
Draw random values from HyperGeometric distribution.
Parameters
----------
point : dict, optional
Dict of variable values on which random values are to be
conditioned (uses default point if not specified).
size : int, optional
Desired size of random sample (returns one sample if not
specified).
Returns
-------
array
"""
N, n, k = draw_values([self.N, self.n, self.k], point=point, size=size)
return generate_samples(np.random.hypergeometric,
N,
n,
k,
dist_shape=self.shape,
size=size)
def sample(self,
size=1,
generate_linear=False,
return_logprobs=False,
random_state=None,
dtype=None,
**kwargs):
"""
Generate random samples from the prior.
.. note::
Right now, generating samples with the prior values is slow (i.e.
with ``return_logprobs=True``) because of pymc3 issues (see
discussion here:
https://discourse.pymc.io/t/draw-values-speed-scaling-with-transformed-variables/4076).
This will hopefully be resolved in the future...
Parameters
----------
size : int (optional)
The number of samples to generate.
generate_linear : bool (optional)
Also generate samples in the linear parameters.
return_logprobs : bool (optional)
Generate the log-prior probability at the position of each sample.
**kwargs
Additional keyword arguments are passed to the
`~thejoker.JokerSamples` initializer.
Returns
-------
samples : `thejoker.Jokersamples`
The random samples.
"""
from theano.gof.fg import MissingInputError
from pymc3.distributions import draw_values
import exoplanet.units as xu
if dtype is None:
dtype = np.float64
sub_pars = {
k: p
for k, p in self.pars.items()
if k in self._nonlinear_equiv_units or (
(k in self._linear_equiv_units
or k in self._v0_offsets_equiv_units) and generate_linear)
}
if generate_linear:
par_names = self.par_names
else:
par_names = list(self._nonlinear_equiv_units.keys())
pars_list = list(sub_pars.values())
# MAJOR HACK RELATED TO UPSTREAM ISSUES WITH pymc3:
init_shapes = dict()
for par in pars_list:
if hasattr(par, 'distribution'):
init_shapes[par.name] = par.distribution.shape
par.distribution.shape = (size, )
with random_state_context(random_state):
samples_values = draw_values(pars_list)
raw_samples = {
p.name: samples.astype(dtype)
for p, samples in zip(pars_list, samples_values)
}
if return_logprobs:
logp = []
for par in pars_list:
try:
_logp = par.distribution.logp(raw_samples[par.name]).eval()
except AttributeError:
logger.warning("Cannot auto-compute log-prior value for "
f"parameter {par} because it is defined "
"as a transformation from another "
"variable.")
continue
except MissingInputError:
logger.warning("Cannot auto-compute log-prior value for "
f"parameter {par} because it depends on "
"other variables.")
continue
logp.append(_logp)
log_prior = np.sum(logp, axis=0)
# CONTINUED MAJOR HACK RELATED TO UPSTREAM ISSUES WITH pymc3:
for par in pars_list:
if hasattr(par, 'distribution'):
par.distribution.shape = init_shapes[par.name]
# Apply units if they are specified:
prior_samples = JokerSamples(poly_trend=self.poly_trend,
n_offsets=self.n_offsets,
**kwargs)
for name in par_names:
p = sub_pars[name]
unit = getattr(p, xu.UNIT_ATTR_NAME, u.one)
if p.name not in prior_samples._valid_units.keys():
continue
prior_samples[p.name] = np.atleast_1d(raw_samples[p.name]) * unit
if return_logprobs:
prior_samples['ln_prior'] = log_prior
# TODO: right now, elsewhere, we assume the log_prior is a single value
# for each sample (i.e. the total prior value). In principle, we could
# store all of the individual log-prior values (for each parameter),
# like here:
# log_prior = {k: np.atleast_1d(v)
# for k, v in log_prior.items()}
# log_prior = Table(log_prior)[par_names]
return prior_samples
def random(self, point=None, size=1):
probs, = draw_values([self.probst], point=point, size=size)
sample = multidim_choice(probs=probs, size=size)
if size == 1:
sample = sample[0]
return sample
def forward_val(self, x, point=None):
(opror, ) = draw_values([self.one_plus_ror - 0.0], point=point)
return super(ImpactParameterTransform, self).forward_val(x / opror,
point=point)
def random(self, point=None, size=None, repeat=None):
mu = draw_values([self.mu], point=point)
return generate_samples(self.zpt_cdf,
mu,
dist_shape=self.shape,
size=size)