def step(self, point):
for name, shared_var in self.shared.items():
shared_var.set_value(point[name])
q = DictToArrayBijection.map(
{v.name: point[v.name]
for v in self.vars})
step_res = self.astep(q)
if self.generates_stats:
apoint, stats = step_res
else:
apoint = step_res
if not isinstance(apoint, RaveledVars):
# We assume that the mapping has stayed the same
apoint = RaveledVars(apoint, q.point_map_info)
new_point = DictToArrayBijection.rmap(apoint, start_point=point)
if self.generates_stats:
return new_point, stats
return new_point
def test_leapfrog_reversible_single():
n = 3
start, model, _ = models.non_normal(n)
integrators = ['leapfrog', 'two-stage', 'three-stage']
steps = [BaseHMC(vars=model.vars, model=model, integrator=method, use_single_leapfrog=True)
for method in integrators]
for method, step in zip(integrators, steps):
bij = DictToArrayBijection(step.ordering, start)
q0 = bij.map(start)
p0 = floatX(np.ones(n) * .05)
precision = select_by_precision(float64=1E-8, float32=1E-5)
for epsilon in [0.01, 0.1, 1.2]:
for n_steps in [1, 2, 3, 4, 20]:
dlogp0 = step.dlogp(q0)
q, p = q0, p0
dlogp = dlogp0
energy = step.compute_energy(q, p)
for _ in range(n_steps):
q, p, v, dlogp, _ = step.leapfrog(q, p, dlogp, floatX(np.array(epsilon)))
p = -p
for _ in range(n_steps):
q, p, v, dlogp, _ = step.leapfrog(q, p, dlogp, floatX(np.array(epsilon)))
close_to(q, q0, precision, str(('q', method, n_steps, epsilon)))
close_to(-p, p0, precision, str(('p', method, n_steps, epsilon)))
def create_shared_params(self,
trace=None,
size=None,
jitter=1,
start=None):
if trace is None:
if size is None:
raise opvi.ParametrizationError(
"Need `trace` or `size` to initialize")
else:
if start is None:
start = self.model.initial_point
else:
start_ = self.model.initial_point.copy()
self.model.update_start_vals(start_, start)
start = start_
start = pm.floatX(DictToArrayBijection.map(start))
# Initialize particles
histogram = np.tile(start, (size, 1))
histogram += pm.floatX(
np.random.normal(0, jitter, histogram.shape))
else:
histogram = np.empty((len(trace) * len(trace.chains), self.ddim))
i = 0
for t in trace.chains:
for j in range(len(trace)):
histogram[i] = DictToArrayBijection.map(trace.point(j, t))
i += 1
return dict(histogram=aesara.shared(pm.floatX(histogram), "histogram"))
def test_leapfrog_reversible_single():
n = 3
start, model, _ = models.non_normal(n)
integrators = ['leapfrog', 'two-stage', 'three-stage']
steps = [BaseHMC(vars=model.vars, model=model, integrator=method, use_single_leapfrog=True)
for method in integrators]
for method, step in zip(integrators, steps):
bij = DictToArrayBijection(step.ordering, start)
q0 = bij.map(start)
p0 = np.ones(n) * .05
for epsilon in [0.01, 0.1, 1.2]:
for n_steps in [1, 2, 3, 4, 20]:
dlogp0 = step.dlogp(q0)
q, p = q0, p0
dlogp = dlogp0
energy = step.compute_energy(q, p)
for _ in range(n_steps):
q, p, v, dlogp, _ = step.leapfrog(q, p, dlogp, np.array(epsilon))
p = -p
for _ in range(n_steps):
q, p, v, dlogp, _ = step.leapfrog(q, p, dlogp, np.array(epsilon))
close_to(q, q0, 1e-8, str(('q', method, n_steps, epsilon)))
close_to(-p, p0, 1e-8, str(('p', method, n_steps, epsilon)))
class PyMC3Potential:
def __init__(self, vars=None, model=None, point=None):
self.model = pm.modelcontext(model)
# Work out the full starting coordinates
if point is None:
point = self.model.test_point
else:
pm.util.update_start_vals(point, self.model.test_point, self.model)
# Fit all the parameters by default
if vars is None:
vars = self.model.cont_vars
self.vars = inputvars(vars)
allinmodel(self.vars, self.model)
# Work out the relevant bijection map
point = Point(point, model=self.model)
self.bijection = DictToArrayBijection(ArrayOrdering(self.vars), point)
# Pre-compile the theano model and gradient
nlp = -self.model.logpt
grad = theano.grad(nlp, self.vars, disconnected_inputs="ignore")
self.func = get_theano_function_for_var([nlp] + grad, model=self.model)
def __call__(self, coords):
res = self.func(*get_args_for_theano_function(
self.bijection.rmap(coords), model=self.model))
d = dict(zip((v.name for v in self.vars), res[1:]))
g = self.bijection.map(d)
return res[0], g
def step(self, point: PointType):
partial_funcs_and_point = [
DictToArrayBijection.mapf(x, start_point=point) for x in self.fs
]
if self.allvars:
partial_funcs_and_point.append(point)
apoint = DictToArrayBijection.map(
{v.name: point[v.name]
for v in self.vars})
step_res = self.astep(apoint, *partial_funcs_and_point)
if self.generates_stats:
apoint_new, stats = step_res
else:
apoint_new = step_res
if not isinstance(apoint_new, RaveledVars):
# We assume that the mapping has stayed the same
apoint_new = RaveledVars(apoint_new, apoint.point_map_info)
point_new = DictToArrayBijection.rmap(apoint_new, start_point=point)
if self.generates_stats:
return point_new, stats
return point_new
def __init__(self, vars, out_vars, shared, blocked=True):
self.vars = vars
self.ordering = ArrayOrdering(vars)
self.lordering = utility.ListArrayOrdering(out_vars, intype='tensor')
lpoint = [var.tag.test_value for var in out_vars]
self.shared = {var.name: shared for var, shared in shared.items()}
self.blocked = blocked
self.bij = DictToArrayBijection(self.ordering, self.population[0])
self.lij = utility.ListToArrayBijection(self.lordering, lpoint)
def step(self, point):
for var, share in self.shared.items():
share.set_value(point[var])
self.bij = DictToArrayBijection(self.ordering, point)
if self.generates_stats:
apoint, stats = self.astep(self.bij.map(point))
return self.bij.rmap(apoint), stats
else:
apoint = self.astep(self.bij.map(point))
return self.bij.rmap(apoint)
def test_leapfrog_reversible():
n = 3
np.random.seed(42)
start, model, _ = models.non_normal(n)
size = sum(start[n.name].size for n in model.value_vars)
scaling = floatX(np.random.rand(size))
class HMC(BaseHMC):
def _hamiltonian_step(self, *args, **kwargs):
pass
step = HMC(vars=model.value_vars, model=model, scaling=scaling)
step.integrator._logp_dlogp_func.set_extra_values({})
astart = DictToArrayBijection.map(start)
p = RaveledVars(floatX(step.potential.random()), astart.point_map_info)
q = RaveledVars(floatX(np.random.randn(size)), astart.point_map_info)
start = step.integrator.compute_state(p, q)
for epsilon in [0.01, 0.1]:
for n_steps in [1, 2, 3, 4, 20]:
state = start
for _ in range(n_steps):
state = step.integrator.step(epsilon, state)
for _ in range(n_steps):
state = step.integrator.step(-epsilon, state)
npt.assert_allclose(state.q.data, start.q.data, rtol=1e-5)
npt.assert_allclose(state.p.data, start.p.data, rtol=1e-5)
def initialize_population(self):
"""Create an initial population from the prior distribution."""
population = []
var_info = OrderedDict()
if self.start is None:
init_rnd = sample_prior_predictive(
self.draws,
var_names=[v.name for v in self.model.unobserved_RVs],
model=self.model,
)
else:
init_rnd = self.start
init = self.model.initial_point
for v in self.variables:
var_info[v.name] = (init[v.name].shape, init[v.name].size)
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.posterior = np.array(floatX(population))
self.var_info = var_info
def test_leapfrog_reversible():
n = 3
start, model, _ = models.non_normal(n)
step = BaseHMC(vars=model.vars, model=model)
bij = DictToArrayBijection(step.ordering, start)
q0 = bij.map(start)
p0 = floatX(np.ones(n) * .05)
precision = select_by_precision(float64=1E-8, float32=1E-4)
for epsilon in [.01, .1, 1.2]:
for n_steps in [1, 2, 3, 4, 20]:
q, p = q0, p0
q, p, _ = step.leapfrog(q, p, floatX(np.array(epsilon)), np.array(n_steps, dtype='int32'))
q, p, _ = step.leapfrog(q, -p, floatX(np.array(epsilon)), np.array(n_steps, dtype='int32'))
close_to(q, q0, precision, str((n_steps, epsilon)))
close_to(-p, p0, precision, str((n_steps, epsilon)))
class ArrayStepSharedLLK(BlockedStep):
"""
Modified ArrayStepShared To handle returned larger point including the
likelihood values.
Takes additionally a list of output vars including the likelihoods.
Parameters
----------
vars : list
variables to be sampled
out_vars : list
variables to be stored in the traces
shared : dict
theano variable -> shared variables
blocked : boolen
(default True)
"""
def __init__(self, vars, out_vars, shared, blocked=True):
self.vars = vars
self.ordering = ArrayOrdering(vars)
self.lordering = ListArrayOrdering(out_vars, intype='tensor')
lpoint = [var.tag.test_value for var in out_vars]
self.shared = {var.name: shared for var, shared in shared.items()}
self.blocked = blocked
self.bij = DictToArrayBijection(self.ordering, self.population[0])
blacklist = list(
set(self.lordering.variables) - set([var.name for var in vars]))
self.lij = ListToArrayBijection(self.lordering,
lpoint,
blacklist=blacklist)
def __getstate__(self):
return self.__dict__
def __setstate__(self, state):
self.__dict__.update(state)
def step(self, point):
for var, share in self.shared.items():
share.container.storage[0] = point[var]
apoint, alist = self.astep(self.bij.map(point))
return self.bij.rmap(apoint), alist
def test_leapfrog_reversible():
n = 3
start, model, _ = models.non_normal(n)
step = BaseHMC(vars=model.vars, model=model)
bij = DictToArrayBijection(step.ordering, start)
q0 = bij.map(start)
p0 = np.ones(n) * .05
for epsilon in [.01, .1, 1.2]:
for n_steps in [1, 2, 3, 4, 20]:
q, p = q0, p0
q, p, _ = step.leapfrog(q, p, np.array(epsilon), np.array(n_steps, dtype='int32'))
q, p, _ = step.leapfrog(q, -p, np.array(epsilon), np.array(n_steps, dtype='int32'))
close_to(q, q0, 1e-8, str((n_steps, epsilon)))
close_to(-p, p0, 1e-8, str((n_steps, epsilon)))
def __init__(self, model=None):
# Get the model
self.model = pm.modelcontext(model)
# Get the variables
self.varnames = get_default_varnames(self.model.unobserved_RVs, False)
# Get the starting point
self.start = Point(self.model.test_point, model=self.model)
self.ndim = len(self.start)
self.mean = None
self.cov = None
# Compile the log probability function
self.vars = inputvars(self.model.cont_vars)
self.bij = DictToArrayBijection(ArrayOrdering(self.vars), self.start)
self.func = get_theano_function_for_var(
self.model.logpt, model=self.model
)
class ModelWrapper:
def __init__(self, start=None, vars=None, model=None):
model = self.model = pm.modelcontext(model)
# Work out the full starting coordinates
if start is None:
start = model.test_point
else:
update_start_vals(start, model.test_point, model)
self.start = start
# Fit all the parameters by default
if vars is None:
vars = model.cont_vars
vars = self.vars = inputvars(vars)
allinmodel(vars, model)
# Work out the relevant bijection map
start = Point(start, model=model)
self.bij = DictToArrayBijection(ArrayOrdering(vars), start)
# Pre-compile the theano model and gradient
nlp = -model.logpt
grad = theano.grad(nlp, vars, disconnected_inputs="ignore")
self.func = get_theano_function_for_var([nlp] + grad, model=model)
def __call__(self, vec):
try:
res = self.func(*get_args_for_theano_function(self.bij.rmap(vec),
model=self.model))
except Exception:
import traceback
print("array:", vec)
print("point:", self.bij.rmap(vec))
traceback.print_exc()
raise
d = dict(zip((v.name for v in self.vars), res[1:]))
g = self.bij.map(d)
return res[0], g
def astep(self, q0: RaveledVars) -> Tuple[RaveledVars, List[Dict[str, Any]]]:
point_map_info = q0.point_map_info
q0 = q0.data
if not self.steps_until_tune and self.tune:
if self.tune == "scaling":
self.scaling = tune(self.scaling, self.accepted / float(self.tune_interval))
elif self.tune == "lambda":
self.lamb = tune(self.lamb, self.accepted / float(self.tune_interval))
# Reset counter
self.steps_until_tune = self.tune_interval
self.accepted = 0
epsilon = self.proposal_dist() * self.scaling
# differential evolution proposal
# select two other chains
ir1, ir2 = np.random.choice(self.other_chains, 2, replace=False)
r1 = DictToArrayBijection.map(self.population[ir1])
r2 = DictToArrayBijection.map(self.population[ir2])
# propose a jump
q = floatX(q0 + self.lamb * (r1.data - r2.data) + epsilon)
accept = self.delta_logp(q, q0)
q_new, accepted = metrop_select(accept, q, q0)
self.accepted += accepted
self.steps_until_tune -= 1
stats = {
"tune": self.tune,
"scaling": self.scaling,
"lambda": self.lamb,
"accept": np.exp(accept),
"accepted": accepted,
}
q_new = RaveledVars(q_new, point_map_info)
return q_new, [stats]
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)
bij = DictToArrayBijection(ArrayOrdering(vars), point)
rval = np.ones(bij.map(point).size) / 10
return rval
def __init__(self, vars=None, model=None, point=None):
self.model = pm.modelcontext(model)
# Work out the full starting coordinates
if point is None:
point = self.model.test_point
else:
pm.util.update_start_vals(point, self.model.test_point, self.model)
# Fit all the parameters by default
if vars is None:
vars = self.model.cont_vars
self.vars = inputvars(vars)
allinmodel(self.vars, self.model)
# Work out the relevant bijection map
point = Point(point, model=self.model)
self.bijection = DictToArrayBijection(ArrayOrdering(self.vars), point)
# Pre-compile the theano model and gradient
nlp = -self.model.logpt
grad = theano.grad(nlp, self.vars, disconnected_inputs="ignore")
self.func = get_theano_function_for_var([nlp] + grad, model=self.model)
def step(self, point):
bij = DictToArrayBijection(self.ordering, point)
inputs = [bij.mapf(x) for x in self.fs]
if self.allvars:
inputs.append(point)
if self.generates_stats:
apoint, stats = self.astep(bij.map(point), *inputs)
return bij.rmap(apoint), stats
else:
apoint = self.astep(bij.map(point), *inputs)
return bij.rmap(apoint)
class ArrayStepShared(BlockedStep):
"""Faster version of ArrayStep that requires the substep method that does not wrap
the functions the step method uses.
Works by setting shared variables before using the step. This eliminates the mapping
and unmapping overhead as well as moving fewer variables around.
"""
def __init__(self, vars, shared, blocked=True):
"""
Parameters
----------
vars: list of sampling variables
shared: dict of aesara variable -> shared variable
blocked: Boolean (default True)
"""
self.vars = vars
self.ordering = ArrayOrdering(vars)
self.shared = {
get_var_name(var): shared
for var, shared in shared.items()
}
self.blocked = blocked
self.bij = None
def step(self, point):
for var, share in self.shared.items():
share.set_value(point[var])
self.bij = DictToArrayBijection(self.ordering, point)
if self.generates_stats:
apoint, stats = self.astep(self.bij.map(point))
return self.bij.rmap(apoint), stats
else:
apoint = self.astep(self.bij.map(point))
return self.bij.rmap(apoint)
def create_shared_params(self, start=None):
if start is None:
start = self.model.initial_point
else:
start_ = start.copy()
self.model.update_start_vals(start_, self.model.initial_point)
start = start_
if self.batched:
start = start[self.group[0].name][0]
else:
start = DictToArrayBijection.map(start)
rho = np.zeros((self.ddim, ))
if self.batched:
start = np.tile(start, (self.bdim, 1))
rho = np.tile(rho, (self.bdim, 1))
return {
"mu": aesara.shared(pm.floatX(start), "mu"),
"rho": aesara.shared(pm.floatX(rho), "rho"),
}
def create_shared_params(self, start=None):
if start is None:
start = self.model.initial_point
else:
start_ = start.copy()
self.model.update_start_vals(start_, self.model.initial_point)
start = start_
if self.batched:
start = start[self.group[0].name][0]
else:
start = DictToArrayBijection.map(start)
n = self.ddim
L_tril = np.eye(n)[np.tril_indices(n)].astype(aesara.config.floatX)
if self.batched:
start = np.tile(start, (self.bdim, 1))
L_tril = np.tile(L_tril, (self.bdim, 1))
return {
"mu": aesara.shared(start, "mu"),
"L_tril": aesara.shared(L_tril, "L_tril")
}
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 astep(self, q0):
"""Perform a single HMC iteration."""
perf_start = time.perf_counter()
process_start = time.process_time()
p0 = self.potential.random()
p0 = RaveledVars(p0, q0.point_map_info)
start = self.integrator.compute_state(q0, p0)
if not np.isfinite(start.energy):
model = self._model
check_test_point = model.point_logps()
error_logp = check_test_point.loc[(
np.abs(check_test_point) >= 1e20) | np.isnan(check_test_point)]
self.potential.raise_ok(q0.point_map_info)
message_energy = (
"Bad initial energy, check any log probabilities that "
"are inf or -inf, nan or very small:\n{}".format(
error_logp.to_string()))
warning = SamplerWarning(
WarningType.BAD_ENERGY,
message_energy,
"critical",
self.iter_count,
)
self._warnings.append(warning)
raise SamplingError("Bad initial energy")
adapt_step = self.tune and self.adapt_step_size
step_size = self.step_adapt.current(adapt_step)
self.step_size = step_size
if self._step_rand is not None:
step_size = self._step_rand(step_size)
hmc_step = self._hamiltonian_step(start, p0.data, step_size)
perf_end = time.perf_counter()
process_end = time.process_time()
self.step_adapt.update(hmc_step.accept_stat, adapt_step)
self.potential.update(hmc_step.end.q, hmc_step.end.q_grad, self.tune)
if hmc_step.divergence_info:
info = hmc_step.divergence_info
point = None
point_dest = None
info_store = None
if self.tune:
kind = WarningType.TUNING_DIVERGENCE
else:
kind = WarningType.DIVERGENCE
self._num_divs_sample += 1
# We don't want to fill up all memory with divergence info
if self._num_divs_sample < 100 and info.state is not None:
point = DictToArrayBijection.rmap(info.state.q)
if self._num_divs_sample < 100 and info.state_div is not None:
point = DictToArrayBijection.rmap(info.state_div.q)
if self._num_divs_sample < 100:
info_store = info
warning = SamplerWarning(
kind,
info.message,
"debug",
self.iter_count,
info.exec_info,
divergence_point_source=point,
divergence_point_dest=point_dest,
divergence_info=info_store,
)
self._warnings.append(warning)
self.iter_count += 1
if not self.tune:
self._samples_after_tune += 1
stats = {
"tune": self.tune,
"diverging": bool(hmc_step.divergence_info),
"perf_counter_diff": perf_end - perf_start,
"process_time_diff": process_end - process_start,
"perf_counter_start": perf_start,
}
stats.update(hmc_step.stats)
stats.update(self.step_adapt.stats())
return hmc_step.end.q, [stats]
class MCMCInterface:
"""
An interface for using a ``pymc3`` model with a plain vanilla MCMC sampler.
Args:
model (optional): The ``pymc3`` model. If ``None`` (default), uses the
current model on the stack.
"""
def __init__(self, model=None):
# Get the model
self.model = pm.modelcontext(model)
# Get the variables
self.varnames = get_default_varnames(self.model.unobserved_RVs, False)
# Get the starting point
self.start = Point(self.model.test_point, model=self.model)
self.ndim = len(self.start)
self.mean = None
self.cov = None
# Compile the log probability function
self.vars = inputvars(self.model.cont_vars)
self.bij = DictToArrayBijection(ArrayOrdering(self.vars), self.start)
self.func = get_theano_function_for_var(
self.model.logpt, model=self.model
)
def optimize(self, **kwargs):
"""
Maximize the log probability of a ``pymc3`` model.
This routine wraps ``pymc3_ext.optimize``, which in turn
wraps the ``scipy.optimize.minimize`` function. This method
accepts any of the keywords accepted by either of those
two functions.
Returns:
The array of parameter values at the optimum point.
"""
self.map_soln, self.info = optimize(
model=self.model, return_info=True, **kwargs
)
self.mean = self.info["x"]
self.cov = self.info["hess_inv"]
return self.mean
def get_initial_state(
self, nwalkers=30, var=None, check_finite=True, max_tries=100
):
"""
Generate random initial points for sampling.
If the ``optimize`` method was called beforehand, this method
returns samples from a multidimensional Gaussian centered on
the maximum a posteriori (MAP) solution with covariance equal
to the inverse of the Hessian matrix at that point, unless
``var`` is provided, in which case that is used instead.
If the optimizer was not called, this method
returns samples from a Gaussian with mean equal to the
model's test point (``model.test_point``) and variance equal to
``var``.
Args:
var (float, array, or matrix, optional): Variance of the
multidimensional Gaussian used to draw samples.
This quantity is optional if ``optimize`` was called
beforehand, otherwise it must be provided.
Default is ``None``.
Returns:
An array of shape ``(nwalkers, ndim)`` where ``ndim``
is the number of free model parameters.
"""
if var is None:
if self.mean is not None and self.cov is not None:
# User ran `optimize`, so let's sample from
# the Laplacian approximation at the MAP point
mean = self.mean
cov = self.cov
else:
raise ValueError(
"Please provide a variance `var`, or run `optimize` before calling this method."
)
else:
if self.mean is not None:
# User ran `optimize`, so let's sample around
# the MAP point
mean = self.mean
else:
# Sample around the test value
mean = self.bij.map(self.start)
cov = var * np.eye(len(mean))
# Sample from the Gaussian
p0 = np.random.multivariate_normal(mean, cov, size=nwalkers)
# Ensure the log probability is finite everywhere
if check_finite:
for k in range(nwalkers):
n = 0
while not np.isfinite(self.logp(p0[k])):
if n > max_tries:
raise ValueError(
"Unable to initialize walkers at a point with finite `logp`. "
"Try reducing `var` or running `optimize()`."
)
p0[k] = np.random.multivariate_normal(mean, cov)
return p0
def logp(self, x):
"""
Return the log probability evaluated at a point.
Args:
x (array): The array of parameter values.
Returns:
The value of the log probability function evaluated at ``x``.
"""
try:
res = self.func(
*get_args_for_theano_function(
self.bij.rmap(x), model=self.model
)
)
except Exception:
import traceback
print("array:", x)
print("point:", self.bij.rmap(x))
traceback.print_exc()
raise
return res
def transform(self, samples, varnames=None, progress=True):
"""
Transform samples from the internal to the user parametrization.
Args:
samples (array or matrix): The set of points to transform.
varnames (list, optional): The names of the parameters to
transform to. These may either be strings or the actual
``pymc3`` model variables. If ``None`` (default), these
are determined automatically and may be accessed as the
``varnames`` attribute of this class.
progress (bool, optional): Display a progress bar? Default ``True``.
Returns:
An array of shape ``(..., len(varnames))``, where
``... = samples.shape[:-1]``, containing the transformed
samples.
"""
is_1d = len(np.shape(samples)) == 1
samples = np.atleast_2d(samples)
if varnames is None:
varnames = self.varnames
varnames = [v.name if not type(v) is str else v for v in varnames]
shape = list(samples.shape)
shape[-1] = len(varnames)
x = np.zeros(shape)
for k in tqdm(range(len(samples)), disable=not progress):
point = pmx.optim.get_point(self, samples[k])
for j, name in enumerate(varnames):
x[k, j] = point[name]
if is_1d:
return x.flatten()
else:
return x
def __init__(
self,
draws: int,
tune: int,
step_method,
step_method_pickled,
chain: int,
seed,
start,
mp_ctx,
pickle_backend,
):
self.chain = chain
process_name = "worker_chain_%s" % chain
self._msg_pipe, remote_conn = multiprocessing.Pipe()
self._shared_point = {}
self._point = {}
for name, shape, dtype in DictToArrayBijection.map(start).point_map_info:
size = 1
for dim in shape:
size *= int(dim)
size *= dtype.itemsize
if size != ctypes.c_size_t(size).value:
raise ValueError("Variable %s is too large" % name)
array = mp_ctx.RawArray("c", size)
self._shared_point[name] = (array, shape, dtype)
array_np = np.frombuffer(array, dtype).reshape(shape)
array_np[...] = start[name]
self._point[name] = array_np
self._readable = True
self._num_samples = 0
if step_method_pickled is not None:
step_method_send = step_method_pickled
else:
step_method_send = step_method
self._process = mp_ctx.Process(
daemon=True,
name=process_name,
target=_run_process,
args=(
process_name,
remote_conn,
step_method_send,
step_method_pickled is not None,
self._shared_point,
draws,
tune,
seed,
pickle_backend,
),
)
self._process.start()
# Close the remote pipe, so that we get notified if the other
# end is closed.
remote_conn.close()
def find_MAP(start=None,
vars=None,
method="L-BFGS-B",
return_raw=False,
include_transformed=True,
progressbar=True,
maxeval=5000,
model=None,
*args,
**kwargs):
"""
Finds the local maximum a posteriori point given a model.
find_MAP should not be used to initialize the NUTS sampler. Simply call pymc3.sample() and it will automatically initialize NUTS in a better way.
Parameters
----------
start: `dict` of parameter values (Defaults to `model.test_point`)
vars: list
List of variables to optimize and set to optimum (Defaults to all continuous).
method: string or callable
Optimization algorithm (Defaults to 'L-BFGS-B' unless
discrete variables are specified in `vars`, then
`Powell` which will perform better). For instructions on use of a callable,
refer to SciPy's documentation of `optimize.minimize`.
return_raw: bool
Whether to return the full output of scipy.optimize.minimize (Defaults to `False`)
include_transformed: bool, optional defaults to True
Flag for reporting automatically transformed variables in addition
to original variables.
progressbar: bool, optional defaults to True
Whether or not to display a progress bar in the command line.
maxeval: int, optional, defaults to 5000
The maximum number of times the posterior distribution is evaluated.
model: Model (optional if in `with` context)
*args, **kwargs
Extra args passed to scipy.optimize.minimize
Notes
-----
Older code examples used find_MAP() to initialize the NUTS sampler,
but this is not an effective way of choosing starting values for sampling.
As a result, we have greatly enhanced the initialization of NUTS and
wrapped it inside pymc3.sample() and you should thus avoid this method.
"""
model = modelcontext(model)
if start is None:
start = model.test_point
else:
update_start_vals(start, model.test_point, model)
check_start_vals(start, model)
if vars is None:
vars = model.cont_vars
vars = inputvars(vars)
disc_vars = list(typefilter(vars, discrete_types))
allinmodel(vars, model)
start = Point(start, model=model)
bij = DictToArrayBijection(ArrayOrdering(vars), start)
logp_func = bij.mapf(model.fastlogp_nojac)
x0 = bij.map(start)
try:
dlogp_func = bij.mapf(model.fastdlogp_nojac(vars))
compute_gradient = True
except (AttributeError, NotImplementedError, tg.NullTypeGradError):
compute_gradient = False
if disc_vars or not compute_gradient:
pm._log.warning(
"Warning: gradient not available." +
"(E.g. vars contains discrete variables). MAP " +
"estimates may not be accurate for the default " +
"parameters. Defaulting to non-gradient minimization " +
"'Powell'.")
method = "Powell"
if "fmin" in kwargs:
fmin = kwargs.pop("fmin")
warnings.warn(
"In future versions, set the optimization algorithm with a string. "
'For example, use `method="L-BFGS-B"` instead of '
'`fmin=sp.optimize.fmin_l_bfgs_b"`.')
cost_func = CostFuncWrapper(maxeval, progressbar, logp_func)
# Check to see if minimization function actually uses the gradient
if "fprime" in getargspec(fmin).args:
def grad_logp(point):
return nan_to_num(-dlogp_func(point))
opt_result = fmin(cost_func, x0, fprime=grad_logp, *args, **kwargs)
else:
# Check to see if minimization function uses a starting value
if "x0" in getargspec(fmin).args:
opt_result = fmin(cost_func, x0, *args, **kwargs)
else:
opt_result = fmin(cost_func, *args, **kwargs)
if isinstance(opt_result, tuple):
mx0 = opt_result[0]
else:
mx0 = opt_result
else:
# remove 'if' part, keep just this 'else' block after version change
if compute_gradient:
cost_func = CostFuncWrapper(maxeval, progressbar, logp_func,
dlogp_func)
else:
cost_func = CostFuncWrapper(maxeval, progressbar, logp_func)
try:
opt_result = minimize(cost_func,
x0,
method=method,
jac=compute_gradient,
*args,
**kwargs)
mx0 = opt_result["x"] # r -> opt_result
except (KeyboardInterrupt, StopIteration) as e:
mx0, opt_result = cost_func.previous_x, None
if isinstance(e, StopIteration):
pm._log.info(e)
finally:
last_v = cost_func.n_eval
if progressbar:
assert isinstance(cost_func.progress, ProgressBar)
cost_func.progress.total = last_v
cost_func.progress.update(last_v)
print()
vars = get_default_varnames(model.unobserved_RVs, include_transformed)
mx = {
var.name: value
for var, value in zip(vars,
model.fastfn(vars)(bij.rmap(mx0)))
}
if return_raw:
return mx, opt_result
else:
return mx
def astep(self, q0):
"""One MLDA step, given current sample q0"""
# Check if the tuning flag has been changed and if yes,
# change the proposal's tuning flag and reset self.accepted
# This is triggered by _iter_sample while the highest-level MLDA step
# method is running. It then propagates to all levels.
if self.proposal_dist.tune != self.tune:
self.proposal_dist.tune = self.tune
# set tune in sub-methods of compound stepper explicitly because
# it is not set within sample.py (only the CompoundStep's tune flag is)
if isinstance(self.step_method_below, CompoundStep):
for method in self.step_method_below.methods:
method.tune = self.tune
self.accepted = 0
# Convert current sample from numpy array ->
# dict before feeding to proposal
q0_dict = DictToArrayBijection.rmap(q0)
# Set subchain_selection (which sample from the coarse chain
# is passed as a proposal to the fine chain). If variance
# reduction is used, a random sample is selected as proposal.
# If variance reduction is not used, the last sample is
# selected as proposal.
if self.variance_reduction:
self.subchain_selection = np.random.randint(0, self.subsampling_rate)
else:
self.subchain_selection = self.subsampling_rate - 1
self.proposal_dist.subchain_selection = self.subchain_selection
# Call the recursive DA proposal to get proposed sample
# and convert dict -> numpy array
pre_q = self.proposal_dist(q0_dict)
q = DictToArrayBijection.map(pre_q)
# Evaluate MLDA acceptance log-ratio
# If proposed sample from lower levels is the same as current one,
# do not calculate likelihood, just set accept to 0.0
if (q.data == q0.data).all():
accept = np.float(0.0)
skipped_logp = True
else:
accept = self.delta_logp(q.data, q0.data) + self.delta_logp_below(q0.data, q.data)
skipped_logp = False
# Accept/reject sample - next sample is stored in q_new
q_new, accepted = metrop_select(accept, q, q0)
if skipped_logp:
accepted = False
# if sample is accepted, update self.Q_last with the sample's Q value
# runs only for VR or when store_Q_fine is True
if self.variance_reduction or self.store_Q_fine:
if accepted and not skipped_logp:
self.Q_last = self.model.Q.get_value()
# Variance reduction
if self.variance_reduction:
self.update_vr_variables(accepted, skipped_logp)
# Adaptive error model - runs only during tuning.
if self.tune and self.adaptive_error_model:
self.update_error_estimate(accepted, skipped_logp)
# Update acceptance counter
self.accepted += accepted
stats = {"tune": self.tune, "accept": np.exp(accept), "accepted": accepted}
# Save the VR statistics to the stats dictionary (only happens in the
# top MLDA level)
if (self.variance_reduction or self.store_Q_fine) and not self.is_child:
q_stats = {}
if self.variance_reduction:
m = self
for level in range(self.num_levels - 1, 0, -1):
# save the Q differences for this level and iteration
q_stats[f"Q_{level}_{level - 1}"] = np.array(m.Q_diff)
# this makes sure Q_diff is reset for
# the next iteration
m.Q_diff = []
if level == 1:
break
m = m.step_method_below
q_stats["Q_0"] = np.array(m.Q_base_full)
m.Q_base_full = []
if self.store_Q_fine:
q_stats["Q_" + str(self.num_levels - 1)] = np.array(self.Q_last)
stats = {**stats, **q_stats}
# Capture the base tuning stats from the level below.
self.base_tuning_stats = []
if isinstance(self.step_method_below, MLDA):
self.base_tuning_stats = self.step_method_below.base_tuning_stats
elif isinstance(self.step_method_below, MetropolisMLDA):
self.base_tuning_stats.append({"base_scaling": self.step_method_below.scaling[0]})
elif isinstance(self.step_method_below, DEMetropolisZMLDA):
self.base_tuning_stats.append(
{
"base_scaling": self.step_method_below.scaling[0],
"base_lambda": self.step_method_below.lamb,
}
)
elif isinstance(self.step_method_below, CompoundStep):
# Below method is CompoundStep
for method in self.step_method_below.methods:
if isinstance(method, MetropolisMLDA):
self.base_tuning_stats.append({"base_scaling": method.scaling[0]})
elif isinstance(method, DEMetropolisZMLDA):
self.base_tuning_stats.append(
{"base_scaling": method.scaling[0], "base_lambda": method.lamb}
)
return q_new, [stats] + self.base_tuning_stats
def find_MAP(start=None,
vars=None,
method="L-BFGS-B",
return_raw=False,
include_transformed=True,
progressbar=True,
maxeval=5000,
model=None,
*args,
**kwargs):
"""Finds the local maximum a posteriori point given a model.
`find_MAP` should not be used to initialize the NUTS sampler. Simply call
``pymc3.sample()`` and it will automatically initialize NUTS in a better
way.
Parameters
----------
start: `dict` of parameter values (Defaults to `model.initial_point`)
vars: list
List of variables to optimize and set to optimum (Defaults to all continuous).
method: string or callable
Optimization algorithm (Defaults to 'L-BFGS-B' unless
discrete variables are specified in `vars`, then
`Powell` which will perform better). For instructions on use of a callable,
refer to SciPy's documentation of `optimize.minimize`.
return_raw: bool
Whether to return the full output of scipy.optimize.minimize (Defaults to `False`)
include_transformed: bool, optional defaults to True
Flag for reporting automatically transformed variables in addition
to original variables.
progressbar: bool, optional defaults to True
Whether or not to display a progress bar in the command line.
maxeval: int, optional, defaults to 5000
The maximum number of times the posterior distribution is evaluated.
model: Model (optional if in `with` context)
*args, **kwargs
Extra args passed to scipy.optimize.minimize
Notes
-----
Older code examples used `find_MAP` to initialize the NUTS sampler,
but this is not an effective way of choosing starting values for sampling.
As a result, we have greatly enhanced the initialization of NUTS and
wrapped it inside ``pymc3.sample()`` and you should thus avoid this method.
"""
model = modelcontext(model)
if vars is None:
vars = model.cont_vars
if not vars:
raise ValueError("Model has no unobserved continuous variables.")
vars = inputvars(vars)
disc_vars = list(typefilter(vars, discrete_types))
allinmodel(vars, model)
start = copy.deepcopy(start)
if start is None:
start = model.initial_point
else:
model.update_start_vals(start, model.initial_point)
model.check_start_vals(start)
start = Point(start, model=model)
x0 = DictToArrayBijection.map(start)
# TODO: If the mapping is fixed, we can simply create graphs for the
# mapping and avoid all this bijection overhead
def logp_func(x):
return DictToArrayBijection.mapf(model.fastlogp_nojac)(RaveledVars(
x, x0.point_map_info))
try:
# This might be needed for calls to `dlogp_func`
# start_map_info = tuple((v.name, v.shape, v.dtype) for v in vars)
def dlogp_func(x):
return DictToArrayBijection.mapf(model.fastdlogp_nojac(vars))(
RaveledVars(x, x0.point_map_info))
compute_gradient = True
except (AttributeError, NotImplementedError, tg.NullTypeGradError):
compute_gradient = False
if disc_vars or not compute_gradient:
pm._log.warning(
"Warning: gradient not available." +
"(E.g. vars contains discrete variables). MAP " +
"estimates may not be accurate for the default " +
"parameters. Defaulting to non-gradient minimization " +
"'Powell'.")
method = "Powell"
if compute_gradient:
cost_func = CostFuncWrapper(maxeval, progressbar, logp_func,
dlogp_func)
else:
cost_func = CostFuncWrapper(maxeval, progressbar, logp_func)
try:
opt_result = minimize(cost_func,
x0.data,
method=method,
jac=compute_gradient,
*args,
**kwargs)
mx0 = opt_result["x"] # r -> opt_result
except (KeyboardInterrupt, StopIteration) as e:
mx0, opt_result = cost_func.previous_x, None
if isinstance(e, StopIteration):
pm._log.info(e)
finally:
last_v = cost_func.n_eval
if progressbar:
assert isinstance(cost_func.progress, ProgressBar)
cost_func.progress.total = last_v
cost_func.progress.update(last_v)
print()
mx0 = RaveledVars(mx0, x0.point_map_info)
vars = get_default_varnames(model.unobserved_value_vars,
include_transformed)
mx = {
var.name: value
for var, value in zip(
vars,
model.fastfn(vars)(DictToArrayBijection.rmap(mx0)))
}
if return_raw:
return mx, opt_result
else:
return mx
def optimize(start=None,
vars=None,
model=None,
return_info=False,
verbose=True,
**kwargs):
"""Maximize the log prob of a PyMC3 model using scipy
All extra arguments are passed directly to the ``scipy.optimize.minimize``
function.
Args:
start: The PyMC3 coordinate dictionary of the starting position
vars: The variables to optimize
model: The PyMC3 model
return_info: Return both the coordinate dictionary and the result of
``scipy.optimize.minimize``
verbose: Print the success flag and log probability to the screen
"""
from scipy.optimize import minimize
model = pm.modelcontext(model)
# Work out the full starting coordinates
if start is None:
start = model.test_point
else:
update_start_vals(start, model.test_point, model)
# Fit all the parameters by default
if vars is None:
vars = model.cont_vars
vars = inputvars(vars)
allinmodel(vars, model)
# Work out the relevant bijection map
start = Point(start, model=model)
bij = DictToArrayBijection(ArrayOrdering(vars), start)
# Pre-compile the theano model and gradient
nlp = -model.logpt
grad = theano.grad(nlp, vars, disconnected_inputs="ignore")
func = get_theano_function_for_var([nlp] + grad, model=model)
if verbose:
names = [
get_untransformed_name(v.name)
if is_transformed_name(v.name) else v.name for v in vars
]
sys.stderr.write("optimizing logp for variables: [{0}]\n".format(
", ".join(names)))
bar = tqdm.tqdm()
# This returns the objective function and its derivatives
def objective(vec):
res = func(*get_args_for_theano_function(bij.rmap(vec), model=model))
d = dict(zip((v.name for v in vars), res[1:]))
g = bij.map(d)
if verbose:
bar.set_postfix(logp="{0:e}".format(-res[0]))
bar.update()
return res[0], g
# Optimize using scipy.optimize
x0 = bij.map(start)
initial = objective(x0)[0]
kwargs["jac"] = True
info = minimize(objective, x0, **kwargs)
# Only accept the output if it is better than it was
x = info.x if (np.isfinite(info.fun) and info.fun < initial) else x0
# Coerce the output into the right format
vars = get_default_varnames(model.unobserved_RVs, True)
point = {
var.name: value
for var, value in zip(vars,
model.fastfn(vars)(bij.rmap(x)))
}
if verbose:
bar.close()
sys.stderr.write("message: {0}\n".format(info.message))
sys.stderr.write("logp: {0} -> {1}\n".format(-initial, -info.fun))
if not np.isfinite(info.fun):
logger.warning("final logp not finite, returning initial point")
logger.warning(
"this suggests that something is wrong with the model")
logger.debug("{0}".format(info))
if return_info:
return point, info
return point
def dlogp_func(x):
return DictToArrayBijection.mapf(model.fastdlogp_nojac(vars))(
RaveledVars(x, x0.point_map_info))
