def get_result_point(stage, config, point_llk='max'):
"""
Return point of a given stage result.
Parameters
----------
stage : :class:`models.Stage`
config : :class:`config.BEATConfig`
point_llk : str
with specified llk(max, mean, min).
Returns
-------
dict
"""
if config.sampler_config.name == 'Metropolis':
if stage.step is None:
raise AttributeError('Loading Metropolis results requires'
' sampler parameters to be loaded!')
sc = config.sampler_config.parameters
from beat.sampler.metropolis import get_trace_stats
pdict, _ = get_trace_stats(stage.mtrace, stage.step, sc.burn, sc.thin)
point = pdict[point_llk]
elif config.sampler_config.name == 'SMC':
llk = stage.mtrace.get_values(varname='like', combine=True)
posterior_idxs = utility.get_fit_indexes(llk)
point = stage.mtrace.point(idx=posterior_idxs[point_llk])
elif config.sampler_config.name == 'PT':
params = config.sampler_config.parameters
llk = stage.mtrace.get_values(varname='like',
burn=int(params.n_samples * params.burn),
thin=params.thin)
posterior_idxs = utility.get_fit_indexes(llk)
point = stage.mtrace.point(idx=posterior_idxs[point_llk])
else:
raise NotImplementedError('Sampler "%s" is not supported!' %
config.sampler_config.name)
return point
def get_trace_stats(mtrace, step, burn=0.5, thin=2):
"""
Get mean value of trace variables and return point.
Parameters
----------
mtrace : :class:`pymc3.backends.base.MultiTrace`
Multitrace sampling result
step : initialised :class:`smc.SMC` sampler object
burn : float
Burn-in parameter to throw out samples from the beginning of the trace
thin : int
Thinning of samples in the trace
Returns
-------
dict with points, covariance matrix
"""
n_steps = len(mtrace)
array_population = num.zeros((step.n_jobs * int(
num.ceil(n_steps * (1 - burn) / thin)),
step.ordering.dimensions))
# collect end points of each chain and put into array
for var, slc, shp, _ in step.ordering.vmap:
if len(shp) == 0:
array_population[:, slc] = num.atleast_2d(
mtrace.get_values(varname=var,
burn=int(burn * n_steps),
thin=thin,
combine=True)).T
else:
array_population[:, slc] = mtrace.get_values(
varname=var,
burn=int(burn * n_steps),
thin=thin,
combine=True)
llks = mtrace.get_values(
varname=step.likelihood_name,
burn=int(burn * n_steps),
thin=thin,
combine=True)
posterior_idxs = utility.get_fit_indexes(llks)
d = {}
for k, v in posterior_idxs.iteritems():
d[k] = step.bij.rmap(array_population[v, :])
d['dist_mean'] = step.bij.rmap(array_population.mean(axis=0))
avar = array_population.var(axis=0)
if avar.sum() == 0.:
logger.warn('Trace std not valid not enough samples! Use 1.')
avar = 1.
cov = num.eye(step.ordering.dimensions) * avar
return d, cov