def sample_pmc(analysis_file, posterior, base_directory='./', step_N=500, steps=10, final_N=5000,
perplexity_threshold=1.0, weight_threshold=1e-10, sigma_test_stat=None, initial_proposal='clusters'):
"""
Samples from a named posterior using the Population Monte Carlo (PMC) methods.
The results of the find-cluster command are expected in EOS_BASE_DIRECTORY/POSTERIOR/clusters.
The output file will be stored in EOS_BASE_DIRECTORY/POSTERIOR/pmc.
:param analysis_file: The name of the analysis file that describes the named posterior, or an object of class `eos.AnalysisFile`.
:type analysis_file: str or `eos.AnalysisFile`
:param posterior: The name of the posterior.
:type posterior: str
:param base_directory: The base directory for the storage of data files. Can also be set via the EOS_BASE_DIRECTORY environment variable.
:type base_directory: str, optional
:param step_N: The number of samples to be used in each adaptation step. These samples will be discarded. Defaults to 500.
:type step_N: int > 0, optional
:param steps: The number of adaptation steps, which are used to adapt the PMC proposal to the posterior. Defaults to 10.
:type steps: int > 0, optional
:param final_N: The number of samples to be stored in the output file. Defaults to 5000,
:type final_N: int > 0, optional
:param perplexity_threshold: The threshold for the perplexity in the last step after which further adaptation steps are to be skipped. Defaults to 1.0.
:type perplexity_threshold: 0.0 < float <= 1.0, optional
:param weight_threshold: Mixture components with a weight smaller than this threshold are pruned.
:type weight_threshold: 0.0 < float <= 1.0, optional.
:param sigma_test_stat: If provided, the inverse CDF of -2*log(PDF) will be evaluated, using the provided values as the respective significance.
:type sigma_test_stat: list or iterable
:param initial_proposal: Specify where the initial proposal should be taken from; 'clusters' (default): use the proposal obtained using `find-clusters`;
'product': use the proposal obtained from `mixture_product`; 'pmc': continue sampling from the previous `sample-pmc` results.
:type initial_proposal: str, optional
"""
output_path = os.path.join(base_directory, posterior, 'pmc')
_set_log_file(output_path, 'log', mode='a' if (initial_proposal == 'pmc') else 'w')
if type(analysis_file) is not eos.AnalysisFile:
_analysis_file = eos.AnalysisFile(analysis_file)
else:
_analysis_file = analysis_file
analysis = _analysis_file.analysis(posterior)
rng = _np.random.mtrand.RandomState(1701)
if initial_proposal == 'clusters':
initial_density = eos.data.MixtureDensity(os.path.join(base_directory, posterior, 'clusters')).density()
elif initial_proposal == 'pmc':
previous_sampler = eos.data.PMCSampler(os.path.join(base_directory, posterior, 'pmc'))
initial_density = previous_sampler.density()
elif initial_proposal == 'product':
initial_density = eos.data.MixtureDensity(os.path.join(base_directory, posterior, 'product')).density()
else:
eos.error("Could not initialize proposal in sample_pmc: argument {} is not supported.".format(initial_proposal))
samples, weights, proposal = analysis.sample_pmc(initial_density, step_N=step_N, steps=steps, final_N=final_N,
rng=rng, final_perplexity_threshold=perplexity_threshold, weight_threshold=weight_threshold)
if initial_proposal == 'pmc':
samples = _np.concatenate((previous_sampler.samples, samples), axis=0)
weights = _np.concatenate((previous_sampler.weights, weights), axis=0)
eos.data.PMCSampler.create(output_path, analysis.varied_parameters, samples, weights, proposal, sigma_test_stat=sigma_test_stat)
def predict_observables(analysis_file, posterior, prediction, base_directory='./', begin=0, end=-1):
'''
Predicts a set of observables based on previously obtained PMC samples.
The input files are expected in EOS_BASE_DIRECTORY/POSTERIOR/pmc.
The output files will be stored in EOS_BASE_DIRECTORY/POSTERIOR/pred-PREDICTION.
:param analysis_file: The name of the analysis file that describes the named posterior, or an object of class `eos.AnalysisFile`.
:type analysis_file: str or `eos.AnalysisFile`
:param posterior: The name of the posterior.
:type posterior: str
:param prediction: The name of the set of observables to predict.
:type prediction: str
:param base_directory: The base directory for the storage of data files. Can also be set via the EOS_BASE_DIRECTORY environment variable.
:type base_directory: str, optional
:param begin: The index of the first sample to use for the predictions. Defaults to 0.
:type begin: int
:param end: The index beyond the last sample to use for the predictions. Defaults to -1.
:type begin: int
'''
_parameters = eos.Parameters()
if type(analysis_file) is not eos.AnalysisFile:
_analysis_file = eos.AnalysisFile(analysis_file)
else:
_analysis_file = analysis_file
observables = _analysis_file.observables(prediction, _parameters)
data = eos.data.PMCSampler(os.path.join(base_directory, posterior, 'pmc'))
try:
from tqdm import tqdm
progressbar = tqdm
except ImportError:
progressbar = lambda x: x
parameters = [_parameters[p['name']] for p in data.varied_parameters]
observable_samples = []
for i, sample in enumerate(progressbar(data.samples[begin:end])):
for p, v in zip(parameters, sample):
p.set(v)
try:
observable_samples.append([o.evaluate() for o in observables])
except RuntimeError as e:
eos.error('skipping prediction for sample {i} due to runtime error ({e}): {s}'.format(i=i, e=e, s=sample))
observable_samples.append([_np.nan for o in observables])
observable_samples = _np.array(observable_samples)
output_path = os.path.join(base_directory, posterior, 'pred-{}'.format(prediction))
eos.data.Prediction.create(output_path, observables, observable_samples, data.weights[begin:end])
def run_steps(analysis_file, base_directory='./'):
"""
Runs a list of predefined steps recorded in the analysis file.
Each step corresponds to a call to one of the following common tasks:
- sample-mcmc
- find-cluster
- sample-pmc
- predict-observables
:param analysis_file: The name of the analysis file that describes the named posterior, or an object of class `eos.AnalysisFile`.
:type analysis_file: str or `eos.AnalysisFile`
"""
if type(analysis_file) is not eos.AnalysisFile:
_analysis_file = eos.AnalysisFile(analysis_file)
else:
_analysis_file = analysis_file
_analysis_file.run()
def sample_mcmc(analysis_file, posterior, chain, base_directory='./', pre_N=150, preruns=3, N=1000, stride=5, cov_scale=0.1, start_point=None):
"""
Samples from a named posterior PDF using Markov Chain Monte Carlo (MCMC) methods.
The output file will be stored in EOS_BASE_DIRECTORY/POSTERIOR/mcmc-CHAIN.
:param analysis_file: The name of the analysis file that describes the named posterior, or an object of class `eos.AnalysisFile`.
:type analysis_file: str or `eos.AnalysisFile`
:param posterior: The name of the posterior PDF from which to draw the samples.
:type posterior: str
:param chain: The index assigned to the Markov chain. This value is used to seed the RNG for a reproducible analysis.
:type chain: int >= 0
:param base_directory: The base directory for the storage of data files. Can also be set via the EOS_BASE_DIRECTORY environment variable.
:type base_directory: str, optional
:param pre_N: The number of samples to be used for an adaptation in each prerun steps. These samples will be discarded.
:type pre_N: int, optional
:param preruns: The number of prerun steps, which are used to adapt the MCMC proposal to the posterior.
:type preruns: int, optional
:param N: The number of samples to be stored in the output file. Defaults to 1000.
:type N: int, optional
:param stride: The ratio of samples drawn over samples stored. For every S samples, S - 1 will be discarded. Defaults to 5.
:type stride: int, optional
:param cov_scale: Scale factor for the initial guess of the covariance matrix.
:type cov_scale: float, optional
:param start_point: Optional starting point for the chain
:type start_point: list-like, optional
"""
output_path = os.path.join(base_directory, posterior, 'mcmc-{:04}'.format(int(chain)))
_set_log_file(output_path, 'log')
if type(analysis_file) is not eos.AnalysisFile:
_analysis_file = eos.AnalysisFile(analysis_file)
else:
_analysis_file = analysis_file
analysis = _analysis_file.analysis(posterior)
rng = _np.random.mtrand.RandomState(int(chain) + 1701)
try:
samples, weights = analysis.sample(N=N, stride=stride, pre_N=pre_N, preruns=preruns, rng=rng, cov_scale=cov_scale, start_point=start_point)
eos.data.MarkovChain.create(output_path, analysis.varied_parameters, samples, weights)
except RuntimeError as e:
eos.error('encountered run time error ({e}) in parameter point:'.format(e=e))
for p in analysis.varied_parameters:
eos.error(' - {n}: {v}'.format(n=p.name(), v=p.evaluate()))