#!/usr/bin/env python """ Provides services for parametric bootstrapping. These include the ability to estimate probabilities or estimate confidence intervals. The estimation of probabilities is done by the EstimateProbability class. Functions that provide ParameterController objects for the 'null' and 'alternative' cases are provided to the constructor. Numerous aspects of the bootstrapping can be controlled such as the choice of numerical optimiser, and the number of samples from which to estimate the probability. This class can be run in serial or in parallel (at the level of each random sample). An observed Likelihood Ratio (LR) statistic is estimated using the provided 'observed' data. Random data sets are simulated under the null model and the LR estimated from these. The probability of the observed LR is taken as the number of sample LR's that were >= to the observed. Confidence interval estimation can be done using the EstimateConfidenceIntervals class. Multiple statistics associated with an analysis can be evaluated simultaneously. Similar setup, and parallelisation options as provided by the EstimateProbability class. """ from __future__ import with_statement from cogent.util import parallel from cogent.maths import optimisers import random __author__ = "Gavin Huttley and Andrew Butterfield" __copyright__ = "Copyright 2007-2009, The Cogent Project" __credits__ = ["Gavin Huttley","Andrew Butterfield", "Matthew Wakefield", "Edward Lang", "Peter Maxwell"] __license__ = "GPL" __version__ = "1.4.1" __maintainer__ = "Gavin Huttley" __email__ = "gavin.huttley@anu.edu.au" __status__ = "Production" class ParametricBootstrapCore(object): """Core parametric bootstrap services.""" def __init__(self): """Constructor for core parametric bootstrap services class.""" self._numreplicates = 10 self.seed = None self.results = [] def setNumReplicates(self, num): self._numreplicates = num def setSeed(self, seed): self.seed = seed def run(self, show_progress=False, **opt_args): # Sets self.observed and self.results (a list _numreplicates long) to # whatever is returned from self.simplify([LF result from each PC]). # self.simplify() is used as the entire LF result might not be picklable # for MPI. Subclass must provide self.alignment and # self.parameter_controllers alignment_randomness = random.Random() alignment_randomness.seed(self.seed) parallel.sync_random(alignment_randomness) opt_args['show_progress'] = show_progress if 'random_series' not in opt_args and not opt_args.get('local', None): opt_args['random_series'] = random.Random() self._observed = [] starting_points = [] for pc in self.parameter_controllers: pc.setAlignment(self.alignment) lf = pc.optimise(return_calculator=True, **opt_args) starting_points.append(lf) self._observed.append(lf) self.observed = self.simplify(*self.parameter_controllers) # making the simulated alignment on every node keeps alignment_randomness in sync # across all cpus. This is only a good way of doing it so long as making the alignment # is much faster than optimising the likelihood function. simaligns = [self.parameter_controllers[0].simulateAlignment( random_series=alignment_randomness) for i in range(self._numreplicates)] def _one_replicate(simalign): for (pc, starting_point) in zip(self.parameter_controllers, starting_points): with pc.real_par_controller.updatesPostponed(): # using a calculator as a memo object to rest the parameter values pc.real_par_controller.updateFromCalculator(starting_point) # but alignment is different each time pc.setAlignment(simalign) # pc needs to be told that it can't use all the CPUs it did initially # which seems less than ideal (it should be able to discover that itself) pc.real_par_controller.setupParallelContext() pc.optimise(**opt_args) return self.simplify(*self.parameter_controllers) self.results = parallel.map(_one_replicate, simaligns, show_progress=False)#opt shows anyway class EstimateProbability(ParametricBootstrapCore): # 2 parameter controllers, LR def __init__(self, null_parameter_controller, alt_parameter_controller, alignment): ParametricBootstrapCore.__init__(self) self.alignment = alignment self.null_parameter_controller = null_parameter_controller self.alt_parameter_controller = alt_parameter_controller self.parameter_controllers = [self.null_parameter_controller, self.alt_parameter_controller] def simplify(self, null_result, alt_result): return (null_result.getLogLikelihood(), alt_result.getLogLikelihood()) def getObservedlnL(self): return self.observed def getSamplelnL(self): return self.results def getSampleLRList(self): LR = [2 * (alt_lnL - null_lnL) for (null_lnL, alt_lnL) in self.results] LR.sort() LR.reverse() return LR def getObservedLR(self): return 2 * (self.observed[1] - self.observed[0]) def getEstimatedProb(self): """Return the estimated probability. Calculated as the number of sample LR's >= observed LR divided by the number of replicates. """ observed_LR = self.getObservedLR() sample_LRs = self.getSampleLRList() for (count, value) in enumerate(sample_LRs): if value <= observed_LR: return float(count) / len(sample_LRs) return 1.0 class EstimateConfidenceIntervals(ParametricBootstrapCore): """Estimate confidence interval(s) for one or many statistics by parametric bootstrapping.""" def __init__(self, parameter_controller, func_calcstats, alignment): # func_calcstats takes a param dict and returns the statistic of interest ParametricBootstrapCore.__init__(self) self.alignment = alignment self.parameter_controller = parameter_controller self.parameter_controllers = [parameter_controller] self.func_calcstats = func_calcstats def simplify(self, result): return (result.getLogLikelihood(), self.func_calcstats(result)) #.getStatisticsAsDict()) def getObservedStats(self): return self.observed[1] def getSampleStats(self): return [s for (lnL, s) in self.results] def getSamplelnL(self): return [lnL for (lnL, s) in self.results] def getObservedlnL(self): return self.observed[0]