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 initialize_population(self):
"""Create an initial population from the prior distribution."""
population = []
var_info = OrderedDict()
init_rnd = sample_prior_predictive(
self.draws,
var_names=[v.name for v in self.model.unobserved_RVs],
model=self.model,
)
init = self.model.test_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(self.model.dict_to_array(point))
self.nf_samples = np.array(floatX(population))
self.live_points = np.array(floatX(population))
self.var_info = var_info
self.posterior = np.empty((0, np.shape(self.nf_samples)[1]))
def initialize_population(self):
"""Create an initial population from the prior distribution."""
population = []
if self.init_samples is None:
init_rnd = sample_prior_predictive(
self.N,
var_names=[v.name for v in self.model.unobserved_RVs],
model=self.model,
)
for i in range(self.N):
point = Point(
{v.name: init_rnd[v.name][i]
for v in self.variables},
model=self.model)
population.append(self.model.dict_to_array(point))
self.prior_samples = np.array(floatX(population))
elif self.init_samples is not None:
self.prior_samples = np.copy(self.init_samples)
self.samples = np.copy(self.prior_samples)
self.nf_samples = np.copy(self.samples)
self.get_posterior_logp()
self.get_prior_logp()
self.log_weight = self.posterior_logp - self.prior_logp
self.log_evidence = logsumexp(self.log_weight) - np.log(
len(self.log_weight))
self.evidence = np.exp(self.log_evidence)
self.log_weight = self.log_weight - self.log_evidence
self.regularize_weights()
#same as in fitnf but prior~q
self.log_weight_pq_num = self.posterior_logp + 2 * self.prior_logp
self.log_weight_pq_den = 3 * self.prior_logp
self.log_evidence_pq = logsumexp(self.log_weight_pq_num) - logsumexp(
self.log_weight_pq_den)
self.evidence_pq = np.exp(self.log_evidence_pq)
self.log_weight_pq = self.posterior_logp - self.prior_logp - self.log_evidence_pq
self.pq_bw_loss = np.log(
(np.exp(self.posterior_logp) -
np.exp(self.log_evidence_pq +
self.prior_logp))**2) #not actually used yet I think
self.regularize_weights_pq()
#sum of mean loss (p - q*Z_pq)^2 /N for diagnostic purposes
self.log_mean_loss = np.log(
np.mean((np.exp(self.posterior_logp) -
np.exp(self.prior_logp + self.log_evidence_pq))**2))
self.init_weights_cleanup(lambda x: self.prior_logp(x),
lambda x: self.prior_dlogp(x))
self.q_ess = self.calculate_ess(self.log_weight)
self.total_ess = self.calculate_ess(self.sinf_logw)
self.all_logq = np.array([])
self.nf_models = []
self.nf_models_uw = []
def __init__(self, vars=None, scaling=None, step_scale=0.25, is_cov=False,
model=None, blocked=True, potential=None,
integrator="leapfrog", dtype=None, **theano_kwargs):
"""Set up Hamiltonian samplers with common structures.
Parameters
----------
vars : list of theano variables
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)
is_cov : bool, default=False
Treat scaling as a covariance matrix/vector if True, else treat
it as a precision matrix/vector
model : pymc3 Model instance
blocked: bool, default=True
potential : Potential, optional
An object that represents the Hamiltonian with methods `velocity`,
`energy`, and `random` methods.
**theano_kwargs: passed to theano functions
"""
model = modelcontext(model)
if vars is None:
vars = model.cont_vars
vars = inputvars(vars)
super(BaseHMC, self).__init__(vars, blocked=blocked, model=model,
dtype=dtype, **theano_kwargs)
size = self._logp_dlogp_func.size
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=model)
scaling = guess_scaling(point, model=model, vars=vars)
if scaling is not None and potential is not None:
raise ValueError("Can not specify both potential and scaling.")
self.step_size = step_scale / (size ** 0.25)
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)
def __call__(self, q0_dict: dict) -> dict:
"""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("pymc3")
_log.setLevel(logging.ERROR)
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
self.trace = subsample(
draws=0,
step=self.step_method_below,
start=q0_dict,
trace=self.trace,
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
self.trace = subsample(
draws=self.subsampling_rate,
step=self.step_method_below,
start=q0_dict,
trace=self.trace,
)
# 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
new_point = self.trace.point(-self.subsampling_rate + self.subchain_selection)
new_point = Point(new_point, model=self.model_below, filter_model_vars=True)
return new_point
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.fastfn(hessian_diag(model.logpt, vars))
return H(Point(point, model=model))
def _iter_sample(draws,
step,
start=None,
trace=None,
chain=0,
tune=None,
model=None,
random_seed=-1):
"""
Modified from :func:`pymc3.sampling._iter_sample`
tune: int
adaptiv step-size scaling is stopped after this chain sample
"""
model = modelcontext(model)
draws = int(draws)
if draws < 1:
raise ValueError('Argument `draws` should be above 0.')
if start is None:
start = {}
if random_seed != -1:
seed(random_seed)
try:
step = CompoundStep(step)
except TypeError:
pass
point = Point(start, model=model)
step.chain_index = chain
trace.setup(draws, chain)
for i in range(draws):
if i == tune:
step = stop_tuning(step)
logger.debug('Step: Chain_%i step_%i' % (chain, i))
point, out_list = step.step(point)
trace.write(out_list, i)
yield trace
def __init__(self, vars=None, scaling=None, step_scale=0.25, is_cov=False,
model=None, blocked=True, use_single_leapfrog=False, **theano_kwargs):
"""Superclass to implement Hamiltonian/hybrid monte carlo
Parameters
----------
vars : list of theano variables
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)
is_cov : bool, default=False
Treat scaling as a covariance matrix/vector if True, else treat it as a
precision matrix/vector
state
State object
model : pymc3 Model instance. default=Context model
blocked: Boolean, default True
use_single_leapfrog: Boolean, will leapfrog steps take a single step at a time.
default False.
**theano_kwargs: passed to theano functions
"""
model = modelcontext(model)
if vars is None:
vars = model.cont_vars
vars = inputvars(vars)
if scaling is None:
scaling = model.test_point
if isinstance(scaling, dict):
scaling = guess_scaling(Point(scaling, model=model), model=model, vars=vars)
n = scaling.shape[0]
self.step_size = step_scale / (n ** 0.25)
self.potential = quad_potential(scaling, is_cov, as_cov=False)
shared = make_shared_replacements(vars, model)
if theano_kwargs is None:
theano_kwargs = {}
self.H, self.compute_energy, self.leapfrog, self._vars = get_theano_hamiltonian_functions(
vars, shared, model.logpt, self.potential, use_single_leapfrog, **theano_kwargs)
super(BaseHMC, self).__init__(vars, shared, blocked=blocked)
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 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, 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 __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 of aesara variables
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)
is_cov: bool, default=False
Treat scaling as a covariance matrix/vector if True, else treat
it as a precision matrix/vector
model: pymc3 Model instance
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
vars = inputvars(vars)
super().__init__(vars,
blocked=blocked,
model=model,
dtype=dtype,
**aesara_kwargs)
self.adapt_step_size = adapt_step_size
self.Emax = Emax
self.iter_count = 0
size = self._logp_dlogp_func.size
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=model)
scaling = guess_scaling(point, model=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 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 __init__(self,
vars=None,
scaling=None,
step_scale=0.25,
is_cov=False,
model=None,
blocked=True,
use_single_leapfrog=False,
potential=None,
integrator="leapfrog",
**theano_kwargs):
"""Superclass to implement Hamiltonian/hybrid monte carlo
Parameters
----------
vars : list of theano variables
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)
is_cov : bool, default=False
Treat scaling as a covariance matrix/vector if True, else treat it as a
precision matrix/vector
model : pymc3 Model instance. default=Context model
blocked: Boolean, default True
use_single_leapfrog: Boolean, will leapfrog steps take a single step at a time.
default False.
potential : Potential, optional
An object that represents the Hamiltonian with methods `velocity`,
`energy`, and `random` methods.
**theano_kwargs: passed to theano functions
"""
model = modelcontext(model)
if vars is None:
vars = model.cont_vars
vars = inputvars(vars)
if scaling is None and potential is None:
size = sum(np.prod(var.dshape, dtype=int) for var in vars)
mean = floatX(np.zeros(size))
var = floatX(np.ones(size))
potential = QuadPotentialDiagAdapt(size, mean, var, 10)
if isinstance(scaling, dict):
point = Point(scaling, model=model)
scaling = guess_scaling(point, model=model, vars=vars)
if scaling is not None and potential is not None:
raise ValueError("Can not specify both potential and scaling.")
self.step_size = step_scale / (model.ndim**0.25)
if potential is not None:
self.potential = potential
else:
self.potential = quad_potential(scaling, is_cov)
shared = make_shared_replacements(vars, model)
if theano_kwargs is None:
theano_kwargs = {}
self.H, self.compute_energy, self.compute_velocity, self.leapfrog, self.dlogp = get_theano_hamiltonian_functions(
vars, shared, model.logpt, self.potential, use_single_leapfrog,
integrator, **theano_kwargs)
super(BaseHMC, self).__init__(vars, shared, blocked=blocked)
def __init__(self, vars=None, out_vars=None, covariance=None, scale=1.,
n_chains=100, tune=True, tune_interval=100, model=None,
check_bound=True, likelihood_name='like', backend='csv',
proposal_name='MultivariateNormal', **kwargs):
model = modelcontext(model)
if vars is None:
vars = model.vars
vars = inputvars(vars)
if out_vars is None:
out_vars = model.unobserved_RVs
out_varnames = [out_var.name for out_var in out_vars]
self.scaling = utility.scalar2floatX(num.atleast_1d(scale))
self.tune = tune
self.check_bound = check_bound
self.tune_interval = tune_interval
self.steps_until_tune = tune_interval
self.stage_sample = 0
self.cumulative_samples = 0
self.accepted = 0
self.beta = 1.
self.stage = 0
self.chain_index = 0
# needed to use the same parallel implementation function as for SMC
self.resampling_indexes = num.arange(n_chains)
self.n_chains = n_chains
self.likelihood_name = likelihood_name
self._llk_index = out_varnames.index(likelihood_name)
self.backend = backend
self.discrete = num.concatenate(
[[v.dtype in discrete_types] * (v.dsize or 1) for v in vars])
self.any_discrete = self.discrete.any()
self.all_discrete = self.discrete.all()
# create initial population
self.population = []
self.array_population = num.zeros(n_chains)
logger.info('Creating initial population for {}'
' chains ...'.format(self.n_chains))
for i in range(self.n_chains):
self.population.append(
Point({v.name: v.random() for v in vars}, model=model))
self.population[0] = model.test_point
shared = make_shared_replacements(vars, model)
self.logp_forw = logp_forw(out_vars, vars, shared)
self.check_bnd = logp_forw([model.varlogpt], vars, shared)
super(Metropolis, self).__init__(vars, out_vars, shared)
# init proposal
if covariance is None and proposal_name in multivariate_proposals:
t0 = time()
self.covariance = init_proposal_covariance(
bij=self.bij, vars=vars, model=model, pop_size=1000)
t1 = time()
logger.info('Time for proposal covariance init: %f' % (t1 - t0))
scale = self.covariance
elif covariance is None:
scale = num.ones(sum(v.dsize for v in vars))
else:
scale = covariance
self.proposal_name = proposal_name
self.proposal_dist = choose_proposal(
self.proposal_name, scale=scale)
self.proposal_samples_array = self.proposal_dist(n_chains)
self.chain_previous_lpoint = [[]] * self.n_chains
self._tps = None
def _iter_sample(draws,
step,
start=None,
trace=None,
chain=0,
tune=None,
model=None,
random_seed=-1,
overwrite=True,
update_proposal=False,
keep_last=False):
"""
Modified from :func:`pymc3.sampling._iter_sample`
tune: int
adaptiv step-size scaling is stopped after this chain sample
"""
model = modelcontext(model)
draws = int(draws)
if draws < 1:
raise ValueError('Argument `draws` should be above 0.')
if start is None:
start = {}
if random_seed != -1:
seed(random_seed)
try:
step = CompoundStep(step)
except TypeError:
pass
point = Point(start, model=model)
step.chain_index = chain
trace.setup(draws, chain, overwrite=overwrite)
for i in range(draws):
if i == tune:
step = stop_tuning(step)
logger.debug('Step: Chain_%i step_%i' % (chain, i))
point, out_list = step.step(point)
try:
trace.buffer_write(out_list, step.cumulative_samples)
except BufferError: # buffer full
last_sample = deepcopy(trace.buffer[-1])
if update_proposal: # only valid for PT for now
if step.proposal_name in multivariate_proposals:
cov = trace.get_sample_covariance(step)
if cov is not None:
if not isinstance(trace, MemoryChain):
filename = '%s/proposal_cov_chain_%i_%i.%s' % (
trace.dir_path, trace.chain, trace.cov_counter,
'png')
from matplotlib import pyplot as plt
fig, axs = plt.subplots(1, 1)
im = axs.imshow(cov, aspect='auto')
plt.colorbar(im)
fig.savefig(filename, dpi=150)
plt.close(fig)
step.proposal_dist = choose_proposal(
step.proposal_name, scale=cov)
trace.record_buffer()
if keep_last:
# put last sample back
trace.buffer_write(*last_sample)
yield trace
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 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