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 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:
ipfn = make_initial_point_fn(
model=self.model,
overrides=start,
jitter_rvs={},
return_transformed=True,
)
start = ipfn(
self.model.rng_seeder.randint(2**30, dtype=np.int64))
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 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 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:
start = self._prepare_start(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)).data
i += 1
return dict(histogram=aesara.shared(pm.floatX(histogram), "histogram"))
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 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 _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 create_shared_params(self, start=None):
ipfn = make_initial_point_fn(
model=self.model,
overrides=start,
jitter_rvs={},
return_transformed=True,
)
start = ipfn(self.model.rng_seeder.randint(2**30, dtype=np.int64))
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 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 create_shared_params(self, start=None):
ipfn = make_initial_point_fn(
model=self.model,
overrides=start,
jitter_rvs={},
return_transformed=True,
)
start = ipfn(self.model.rng_seeder.randint(2**30, dtype=np.int64))
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 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]
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)
def find_MAP(start=None,
vars=None,
method="L-BFGS-B",
return_raw=False,
include_transformed=True,
progressbar=True,
maxeval=5000,
model=None,
*args,
seed: Optional[int] = None,
**kwargs):
"""Finds the local maximum a posteriori point given a model.
`find_MAP` should not be used to initialize the NUTS sampler. Simply call
``pymc.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 ``pymc.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)
ipfn = make_initial_point_fn(
model=model,
jitter_rvs={},
return_transformed=True,
overrides=start,
)
if seed is None:
seed = model.rng_seeder.randint(2**30, dtype=np.int64)
start = ipfn(seed)
model.check_start_vals(start)
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(file=sys.stdout)
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 dlogp_func(x):
return DictToArrayBijection.mapf(model.fastdlogp_nojac(vars))(
RaveledVars(x, x0.point_map_info))
def __init__(
self,
draws: int,
tune: int,
step_method,
step_method_pickled,
chain: int,
seed,
start: Dict[str, np.ndarray],
mp_ctx,
):
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:
if mp_ctx.get_start_method() == "spawn":
raise ValueError(
"please provide a pre-pickled step method when multiprocessing start method is 'spawn'"
)
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,
),
)
self._process.start()
# Close the remote pipe, so that we get notified if the other
# end is closed.
remote_conn.close()