def find_hessian(point, vars=None, model=None):
"""
Returns Hessian of logp at the point passed.
Parameters
----------
model: Model (optional if in `with` context)
point: dict
vars: list
Variables for which Hessian is to be calculated.
"""
model = modelcontext(model)
H = model.fastd2logp(vars)
return H(Point(point, filter_model_vars=True, model=model))
def find_hessian_diag(point, vars=None, model=None):
"""
Returns Hessian of logp at the point passed.
Parameters
----------
model: Model (optional if in `with` context)
point: dict
vars: list
Variables for which Hessian is to be calculated.
"""
model = modelcontext(model)
H = model.compile_fn(hessian_diag(model.logpt(), vars))
return H(Point(point, model=model))
def _initialize_kernel(self):
"""Create variables and logp function necessary to run kernel
This method should not be overwritten. If needed, use `setup_kernel`
instead.
"""
# Create dictionary that stores original variables shape and size
initial_point = self.model.recompute_initial_point(
seed=self.rng.integers(2**30))
for v in self.variables:
self.var_info[v.name] = (initial_point[v.name].shape,
initial_point[v.name].size)
# Create particles bijection map
if self.start:
init_rnd = self.start
else:
init_rnd = self.initialize_population()
population = []
for i in range(self.draws):
point = Point(
{v.name: init_rnd[v.name][i]
for v in self.variables},
model=self.model)
population.append(DictToArrayBijection.map(point).data)
self.tempered_posterior = np.array(floatX(population))
# Initialize prior and likelihood log probabilities
shared = make_shared_replacements(initial_point, self.variables,
self.model)
self.prior_logp_func = _logp_forw(initial_point, [self.model.varlogpt],
self.variables, shared)
self.likelihood_logp_func = _logp_forw(initial_point,
[self.model.datalogpt],
self.variables, shared)
priors = [
self.prior_logp_func(sample) for sample in self.tempered_posterior
]
likelihoods = [
self.likelihood_logp_func(sample)
for sample in self.tempered_posterior
]
self.prior_logp = np.array(priors).squeeze()
self.likelihood_logp = np.array(likelihoods).squeeze()
def test_missing_data(self):
# Originally from a case described in #3122
X = np.random.binomial(1, 0.5, 10)
X[0] = -1 # masked a single value
X = np.ma.masked_values(X, value=-1)
with pm.Model() as m:
x1 = pm.Uniform("x1", 0.0, 1.0)
x2 = pm.Bernoulli("x2", x1, observed=X)
gf = m.logp_dlogp_function()
gf._extra_are_set = True
assert m["x2_missing"].type == gf._extra_vars_shared["x2_missing"].type
pnt = m.test_point.copy()
del pnt["x2_missing"]
res = [gf(DictToArrayBijection.map(Point(pnt, model=m))) for i in range(5)]
assert reduce(lambda x, y: np.array_equal(x, y) and y, res) is not False
def fixed_hessian(point, vars=None, model=None):
"""
Returns a fixed Hessian for any chain location.
Parameters
----------
model: Model (optional if in `with` context)
point: dict
vars: list
Variables for which Hessian is to be calculated.
"""
model = modelcontext(model)
if vars is None:
vars = model.cont_vars
vars = inputvars(vars)
point = Point(point, model=model)
rval = np.ones(DictToArrayBijection.map(point).size) / 10
return rval
def __init__(
self,
vars=None,
scaling=None,
step_scale=0.25,
is_cov=False,
model=None,
blocked=True,
potential=None,
dtype=None,
Emax=1000,
target_accept=0.8,
gamma=0.05,
k=0.75,
t0=10,
adapt_step_size=True,
step_rand=None,
**aesara_kwargs
):
"""Set up Hamiltonian samplers with common structures.
Parameters
----------
vars: list, default=None
List of Aesara variables. If None, all continuous RVs from the
model are included.
scaling: array_like, ndim={1,2}
Scaling for momentum distribution. 1d arrays interpreted matrix
diagonal.
step_scale: float, default=0.25
Size of steps to take, automatically scaled down by 1/n**(1/4),
where n is the dimensionality of the parameter space
is_cov: bool, default=False
Treat scaling as a covariance matrix/vector if True, else treat
it as a precision matrix/vector
model: pymc.Model
blocked: bool, default=True
potential: Potential, optional
An object that represents the Hamiltonian with methods `velocity`,
`energy`, and `random` methods.
**aesara_kwargs: passed to Aesara functions
"""
self._model = modelcontext(model)
if vars is None:
vars = self._model.cont_vars
else:
vars = [self._model.rvs_to_values.get(var, var) for var in vars]
super().__init__(vars, blocked=blocked, model=self._model, dtype=dtype, **aesara_kwargs)
self.adapt_step_size = adapt_step_size
self.Emax = Emax
self.iter_count = 0
# We're using the initial/test point to determine the (initial) step
# size.
# XXX: If the dimensions of these terms change, the step size
# dimension-scaling should change as well, no?
test_point = self._model.initial_point
nuts_vars = [test_point[v.name] for v in vars]
size = sum(v.size for v in nuts_vars)
self.step_size = step_scale / (size ** 0.25)
self.step_adapt = step_sizes.DualAverageAdaptation(
self.step_size, target_accept, gamma, k, t0
)
self.target_accept = target_accept
self.tune = True
if scaling is None and potential is None:
mean = floatX(np.zeros(size))
var = floatX(np.ones(size))
potential = QuadPotentialDiagAdapt(size, mean, var, 10)
if isinstance(scaling, dict):
point = Point(scaling, model=self._model)
scaling = guess_scaling(point, model=self._model, vars=vars)
if scaling is not None and potential is not None:
raise ValueError("Can not specify both potential and scaling.")
if potential is not None:
self.potential = potential
else:
self.potential = quad_potential(scaling, is_cov)
self.integrator = integration.CpuLeapfrogIntegrator(self.potential, self._logp_dlogp_func)
self._step_rand = step_rand
self._warnings = []
self._samples_after_tune = 0
self._num_divs_sample = 0
def __call__(self, q0: RaveledVars) -> RaveledVars:
"""Returns proposed sample given the current sample
in dictionary form (q0_dict)."""
# Logging is reduced to avoid extensive console output
# during multiple recursive calls of subsample()
_log = logging.getLogger("pymc")
_log.setLevel(logging.ERROR)
# Convert current sample from RaveledVars ->
# dict before feeding to subsample.
q0_dict = DictToArrayBijection.rmap(q0)
with self.model_below:
# Check if the tuning flag has been set to False
# in which case tuning is stopped. The flag is set
# to False (by MLDA's astep) when the burn-in
# iterations of the highest-level MLDA sampler run out.
# The change propagates to all levels.
if self.tune:
# Subsample in tuning mode
trace = subsample(
draws=0,
step=self.step_method_below,
start=q0_dict,
tune=self.subsampling_rate,
)
else:
# Subsample in normal mode without tuning
# If DEMetropolisZMLDA is the base sampler a flag is raised to
# make sure that history is edited after tuning ends
if self.tuning_end_trigger:
if isinstance(self.step_method_below, DEMetropolisZMLDA):
self.step_method_below.tuning_end_trigger = True
self.tuning_end_trigger = False
trace = subsample(
draws=self.subsampling_rate,
step=self.step_method_below,
start=q0_dict,
tune=0,
)
# set logging back to normal
_log.setLevel(logging.NOTSET)
# return sample with index self.subchain_selection from the generated
# sequence of length self.subsampling_rate. The index is set within
# MLDA's astep() function
q_dict = trace.point(self.subchain_selection)
# Make sure output dict is ordered the same way as the input dict.
q_dict = Point(
{key: q_dict[key]
for key in q0_dict.keys()},
model=self.model_below,
filter_model_vars=True,
)
return DictToArrayBijection.map(q_dict)