Source code for mc3.sampler_driver

# Copyright (c) 2015-2019 Patricio Cubillos and contributors.
# MC3 is open-source software under the MIT license (see LICENSE).

__all__ = ['sample']

import os
import sys
import importlib
import multiprocessing as mpr
from datetime import date

if sys.version_info.major == 2:
    range = xrange

import numpy as np
import matplotlib as mpl
if os.environ.get('DISPLAY', '') == '':
import matplotlib.pyplot as plt

from .fit_driver import fit
from .mcmc_driver import mcmc
from .ns_driver import nested_sampling
from . import utils   as mu
from . import stats   as ms
from . import plots   as mp
from .VERSION import __version__

[docs]@mu.ignore_system_exit def sample(data=None, uncert=None, func=None, params=None, indparams=[], pmin=None, pmax=None, pstep=None, prior=None, priorlow=None, priorup=None, sampler=None, ncpu=None, leastsq=None, chisqscale=False, nchains=7, nsamples=None, burnin=0, thinning=1, grtest=True, grbreak=0.0, grnmin=0.5, wlike=False, fgamma=1.0, fepsilon=0.0, hsize=10, kickoff='normal', plots=False, ioff=False, showbp=True, savefile=None, resume=False, rms=False, log=None, pnames=None, texnames=None, parname=None, nproc=None, stepsize=None, full_output=None, chireturn=None, lm=None, walk=None, **kwargs): """ This beautiful piece of code executes an MCMC or NS posterior sampling. Parameters ---------- data: 1D float ndarray or string Data to be fit by func. If string, path to file containing data. uncert: 1D float ndarray Uncertainties of data. func: Callable or string-iterable The callable function that models data as: model = func(params, *indparams) Or an iterable of 3 strings (funcname, modulename, path) that specifies the function name, function module, and module path. If the module is already in the python-path scope, path can be omitted. params: 1D float ndarray or string Set of initial fitting parameters for func. If string, path to file containing data. indparams: tuple or string Additional arguments required by func. If string, path to file containing indparams. pmin: 1D ndarray Lower boundaries for the posterior exploration. pmax: 1D ndarray Upper boundaries for the posterior exploration. pstep: 1D ndarray Parameter stepping behavior. - Free parameters have pstep>0. - Fixed parameters have pstep=0. - Negative values indicate a shared parameter, with pstep set to the negative index of the sharing parameter (starting the count from 1), e.g.: to share second parameter and first one, do: pstep[1] = -1. For MCMC, the pstep value of free parameters set the scale of the initial jump proposal. prior: 1D ndarray Parameter priors. The type of prior is determined by priorlow and priorup: if both priorlow>0 and priorup>0 Gaussian else Uniform between [pmin,pmax] priorlow: 1D ndarray Lower prior uncertainty values. priorup: 1D ndarray Upper prior uncertainty values. sampler: String Sampling algorithm: - 'mrw': Metropolis random walk. - 'demc': Differential Evolution Markov chain. - 'snooker': DEMC-z with snooker update. - 'dynesty': DynamicNestedSampler() sampler from dynesty. ncpu: Integer Number of processors for the MCMC chains (MC3 defaults to one CPU for each chain plus a CPU for the central hub). leastsq: String If not None, perform a least-square optimization before the MCMC run. Select from: 'lm': Levenberg-Marquardt (most efficient, but does not obey bounds) 'trf': Trust Region Reflective chisqscale: Boolean Scale the data uncertainties such that the reduced chi-square = 1. nchains: Scalar Number of simultaneous chains to run. nsamples: Scalar Total number of samples. burnin: Integer Number of burned-in (discarded) number of iterations at the beginning of the chains. thinning: Integer Thinning factor of the chains (use every thinning-th iteration) used in the GR test and plots. wlike: Bool If True, calculate the likelihood in a wavelet-base. This requires three additional parameters (TBD: this needs documentation). grtest: Boolean If True, run Gelman & Rubin test. grbreak: Float Gelman-Rubin convergence threshold to stop the MCMC (I'd suggest grbreak ~ 1.001--1.005). Do not break if grbreak=0.0 (default). grnmin: Integer or float Minimum number of samples required for grbreak to stop the MCMC. If grnmin > 1: grnmin sets the minimum required number of samples. If 0 < grnmin < 1: grnmin sets the minimum required nsamples fraction. fgamma: Float Proposals jump scale factor for DEMC's gamma. The code computes: gamma = fgamma * 2.38 / sqrt(2*Nfree) fepsilon: Float Jump scale factor for DEMC's support distribution. The code computes: e = fepsilon * Normal(0, pstep) hsize: Integer Number of initial samples per chain. kickoff: String Flag to indicate how to start the chains: 'normal' for normal distribution around initial guess, or 'uniform' for uniform distribution withing the given boundaries. plots: Bool If True plot parameter traces, pairwise-posteriors, and posterior histograms. ioff: Bool If True, set plt.ioff(), i.e., do not display figures on screen. showbp: Bool If True, show best-fitting values in histogram and pairwise plots. savefile: String If not None, filename to store allparams and other MCMC results. resume: Boolean If True resume a previous run. rms: Boolean If True, calculate the RMS of the residuals: data - best_model. log: String or mc3.utils.Log instance Filename (as string) or log handler (as Log instance) handle logging. pnames: 1D string iterable List of parameter names (including fixed and shared parameters) to display on output screen and figures. See also texnames. Screen output trims up to the 11th character. If not defined, default to texnames. texnames: 1D string iterable Parameter names for figures, which may use latex syntax. If not defined, default to pnames. kwargs: Dict Additional keyword arguments passed to the sampler. Deprecated Parameters --------------------- parname: 1D string ndarray Deprecated, use pnames instead. nproc: Integer Deprecated, use ncpu instead. stepsize: 1D ndarray Deprecated, use pstep instead. chireturn: Deprecated. full_output: Bool Deprecated. lm: Bool Deprecated, see leastsq. walk: String Deprecated, use sampler instead. Returns ------- mc3_output: Dict A Dictionary containing the MCMC posterior distribution and related stats, including: - posterior: thinned posterior distribution of shape [nsamples, nfree], including the burn-in phase. - zchain: chain indices for the posterior samples. - zmask: posterior mask to remove the burn-in. - chisq: chi^2 values for the posterior samples. - log_post: log(posterior) for the posterior samples (see Notes). - burnin: number of burned-in samples per chain. - ifree: Indices of the free parameters. - pnames: Parameter names. - texnames: Parameter names in Latex format. - meanp: mean of the marginal posteriors. - stdp: standard deviation of the marginal posteriors. - CRlo: lower boundary of the marginal 68%-highest posterior density (the credible region). - CRhi: upper boundary of the marginal 68%-HPD. - bestp: model parameters for the optimal log(posterior) in the sample. - best_log_post: optimal log(posterior) in the sample (see Notes). - best_model: model evaluated at bestp. - best_chisq: chi^2 for the optimal log(posterior) in the sample. - red_chisq: reduced chi-square: chi^2/(ndata-nfree) for the best-fitting sample. - BIC: Bayesian Information Criterion: chi^2 - nfree*log(ndata) for the best-fitting sample. - chisq_factor: Uncertainties scale factor to enforce chi^2_red = 1. - stddev_residuals: standard deviation of the residuals. - acceptance_rate: sample's acceptance rate. Notes ----- The log_post variable is defined here as: log_post = log(posterior) = log(likelihood) + log(prior) = -0.5*chi-square + log_prior = sum_i -0.5*((data[i] - model[i])/uncert[i])**2 + log_prior with log_prior defined as: log_prior = sum_j -0.5*((params[j] - prior[j])/prior_uncert[j])**2 For each parameter with a Gaussian prior. Note that constant terms have been neglected. Examples -------- >>> import numpy as np >>> import mc3 >>> def quad(p, x): >>> return p[0] + p[1]*x + p[2]*x**2.0 >>> # Preamble, create a noisy synthetic dataset: >>> np.random.seed(3) >>> x = np.linspace(0, 10, 100) >>> p_true = [3, -2.4, 0.5] >>> y = quad(p_true, x) >>> uncert = np.sqrt(np.abs(y)) >>> data = y + np.random.normal(0, uncert) >>> # Initial guess for fitting parameters: >>> params = np.array([ 3.0, -2.0, 0.1]) >>> pstep = np.array([ 1.0, 1.0, 1.0]) >>> pmin = np.array([ 0.0, -5.0, -1.0]) >>> pmax = np.array([10.0, 5.0, 1.0]) >>> # Gaussian prior on first parameter, uniform on second and third: >>> prior = np.array([3.5, 0.0, 0.0]) >>> priorlow = np.array([0.1, 0.0, 0.0]) >>> priorup = np.array([0.1, 0.0, 0.0]) >>> indparams = [x] >>> func = quad >>> ncpu = 7 >>> # MCMC sampling: >>> mcmc_output = mc3.sample(data, uncert, func, params, indparams=indparams, >>> sampler='snooker', pstep=pstep, ncpu=ncpu, pmin=pmin, pmax=pmax, >>> prior=prior, priorlow=priorlow, priorup=priorup, >>> leastsq='lm', nsamples=1e5, burnin=1000, plots=True) >>> # Nested sampling: >>> ns_output = mc3.sample(data, uncert, func, params, indparams=indparams, >>> sampler='dynesty', pstep=pstep, ncpu=ncpu, pmin=pmin, pmax=pmax, >>> prior=prior, priorlow=priorlow, priorup=priorup, >>> leastsq='lm', plots=True) >>> # See more examples and details at: >>> # >>> # """ # Logging object: if isinstance(log, str): log = mu.Log(log, append=resume) closelog = True else: closelog = False if log is None: log = mu.Log() log.msg("\n{:s}\n" " Multi-core Markov-chain Monte Carlo (MC3).\n" " Version {}.\n" " Copyright (c) 2015-{:d} Patricio Cubillos and collaborators.\n" " MC3 is open-source software under the MIT license (see LICENSE).\n" "{:s}\n\n".format(log.sep, __version__,, log.sep)) # Deprecation warnings (to be removed not before summer 2020): if parname is not None: log.warning("parname argument is deprecated. Use pnames instead.") if pnames is None: pnames = parname if nproc is not None: log.warning("nproc argument is deprecated. Use ncpu instead.") if ncpu is None: ncpu = nproc if stepsize is not None: log.warning("stepsize argument is deprecated. Use pstep instead.") if pstep is None: pstep = stepsize if walk is not None: log.warning("walk argument is deprecated. Use sampler instead.") if sampler is None: sampler = walk if chireturn is not None: log.warning("chireturn argument is deprecated.") if full_output is not None: log.warning("full_output argument is deprecated.") if isinstance(leastsq, bool): if leastsq is True: leastsq = 'trf' if lm is False else 'lm' elif leastsq is False: leastsq = None log.warning("leastsq as boolean is deprecated. See docs for new " "usage. Set leastsq={}".format(repr(leastsq))) if isinstance(lm, bool): log.warning('lm argument is deprecated. See new usage of leastsq. ' 'Set leastsq={}'.format(repr(leastsq))) if sampler is None: log.error("'sampler' is a required argument.") if nsamples is None and sampler in ['MRW', 'DEMC', 'snooker']: log.error("'nsamples' is a required argument for MCMC runs.") if leastsq not in [None, 'lm', 'trf']: log.error("Invalid 'leastsq' input ({}). Must select from " "['lm', 'trf'].".format(leastsq)) # Read the model parameters: params = mu.isfile(params, 'params', log, 'ascii', False, not_none=True) # Unpack if necessary: if np.ndim(params) > 1: ninfo, ndata = np.shape(params) if ninfo == 7: # The priors prior = params[4] priorlow = params[5] priorup = params[6] if ninfo >= 4: # The stepsize pstep = params[3] if ninfo >= 3: # The boundaries pmin = params[1] pmax = params[2] else: log.error('Invalid format/shape for params input file.') params = params[0] # The initial guess params = np.array(params) # Process data and uncertainties: data = mu.isfile(data, 'data', log, 'bin', False, not_none=True) if np.ndim(data) > 1: data, uncert = data # Make local 'uncert' a copy, to avoid overwriting: if uncert is None: log.error("'uncert' is a required argument.") uncert = np.copy(uncert) # Process the independent parameters: if indparams != []: indparams = mu.isfile(indparams, 'indparams', log, 'bin', unpack=False) if ioff: plt.ioff() if resume: log.msg("\n\n{:s}\n{:s} Resuming previous MCMC run.\n\n". format(log.sep, log.sep)) # Import the model function: if isinstance(func, (list, tuple, np.ndarray)): if len(func) == 3: sys.path.append(func[2]) else: sys.path.append(os.getcwd()) fmodule = importlib.import_module(func[1]) func = getattr(fmodule, func[0]) elif not callable(func): log.error("'func' must be either a callable or an iterable of strings " "with the model function, file, and path names.") if ncpu is None and sampler in ['snooker', 'demc', 'mrw']: ncpu = nchains elif ncpu is None and sampler == 'dynesty': ncpu = 1 # Cap the number of processors: if ncpu >= mpr.cpu_count(): log.warning("The number of requested CPUs ({:d}) is >= than the number " "of available CPUs ({:d}). Enforced ncpu to {:d}.". format(ncpu, mpr.cpu_count(), mpr.cpu_count()-1)) ncpu = mpr.cpu_count() - 1 nparams = len(params) ndata = len(data) # Setup array of parameter names: if pnames is None and texnames is not None: pnames = texnames elif pnames is not None and texnames is None: texnames = pnames elif pnames is None and texnames is None: pnames = texnames = mu.default_parnames(nparams) pnames = np.asarray(pnames) texnames = np.asarray(texnames) if pmin is None: pmin = np.tile(-np.inf, nparams) if pmax is None: pmax = np.tile( np.inf, nparams) pmin = np.asarray(pmin) pmax = np.asarray(pmax) if (np.any(np.isinf(pmin)) or np.any(np.isinf(pmax))) and sampler=='dynesty': log.error('Parameter space must be constrained by pmin and pmax.') if pstep is None: pstep = 0.1 * np.abs(params) pstep = np.asarray(pstep) # Set prior parameter indices: if prior is None or priorup is None or priorlow is None: prior = priorup = priorlow = np.zeros(nparams) # Override priors for non-free parameters: priorlow[pstep<=0] = 0.0 priorup [pstep<=0] = 0.0 # Check that initial values lie within the boundaries: if np.any(params < pmin) or np.any(params > pmax): pout = "" for pname, par, minp, maxp in zip(pnames, params, pmin, pmax): if par < minp: pout += "\n{:11s} {: 12.5e} < {: 12.5e}".format( pname[:11], minp, par) if par > maxp: pout += "\n{:26s} {: 12.5e} > {: 12.5e}".format( pname[:11], par, maxp) log.error("Some initial-guess values are out of bounds:\n" "Param name pmin value pmax\n" "----------- ------------ ------------ ------------" "{:s}".format(pout)) nfree = int(np.sum(pstep > 0)) # Number of free parameters ifree = np.where(pstep > 0)[0] # Free parameter indices ishare = np.where(pstep < 0)[0] # Shared parameter indices # Check that output path exists: if savefile is not None: fpath, fname = os.path.split(os.path.realpath(savefile)) if not os.path.exists(fpath): log.warning("Output folder path: '{:s}' does not exist. " "Creating new folder.".format(fpath)) os.makedirs(fpath) # At the moment, skip optimization when these dynesty inputs exist: if sampler == 'dynesty' and ('loglikelihood' in kwargs or 'prior_transform' in kwargs): leastsq = None # Least-squares minimization: chisq_factor = 1.0 if leastsq is not None: fit_output = fit(data, uncert, func, np.copy(params), indparams, pstep, pmin, pmax, prior, priorlow, priorup, leastsq) log.msg("Least-squares best-fitting parameters:\n {}\n\n". format(fit_output['bestp']), si=2) # Scale data-uncertainties such that reduced chisq = 1: if chisqscale: chisq_factor = np.sqrt(fit_output['best_chisq']/(ndata-nfree)) uncert *= chisq_factor # Re-calculate best-fitting parameters with new uncertainties: fit_output = fit(data, uncert, func, np.copy(params), indparams, pstep, pmin, pmax, prior, priorlow, priorup, leastsq) log.msg("Least-squares best-fitting parameters (rescaled chisq):" "\n {}\n\n".format(fit_output['bestp']), si=2) params = np.copy(fit_output['bestp']) else: fit_output = None if resume: with np.load(savefile) as oldrun: uncert *= float(oldrun['chisq_factor'])/chisq_factor chisq_factor = float(oldrun['chisq_factor']) # Here's where the magic happens: if sampler in ['mrw', 'demc', 'snooker']: output = mcmc(data, uncert, func, params, indparams, pmin, pmax, pstep, prior, priorlow, priorup, nchains, ncpu, nsamples, sampler, wlike, fit_output, grtest, grbreak, grnmin, burnin, thinning, fgamma, fepsilon, hsize, kickoff, savefile, resume, log) elif sampler == 'dynesty': output = nested_sampling(data, uncert, func, params, indparams, pmin, pmax, pstep, prior, priorlow, priorup, ncpu, thinning, resume, log, **kwargs) if leastsq is not None: if output['best_log_post']-fit_output['best_log_post'] > 3.0e-8: log.warning("MCMC found a better fit than the minimizer:\n" "MCMC best-fitting parameters: (chisq={:.8g})\n{}\n" "Minimizer best-fitting parameters: (chisq={:.8g})\n{}". format(-2*output['best_log_post'], output['bestp'], -2*fit_output['best_log_post'], fit_output['bestp'])) else: output['best_log_post'] = fit_output['best_log_post'] output['best_chisq'] = fit_output['best_chisq'] output['best_model'] = fit_output['best_model'] output['bestp'] = fit_output['bestp'] # And remove burn-in samples: posterior, zchain, zmask = mu.burn(Z=output['posterior'], zchain=output['zchain'], burnin=output['burnin']) # Get some stats: output['chisq_factor'] = chisq_factor output['BIC'] = output['best_chisq'] + nfree*np.log(ndata) if ndata > nfree: output['red_chisq'] = output['best_chisq']/(ndata-nfree) else: output['red_chisq'] = np.nan output['stddev_residuals'] = np.std(output['best_model']-data) # Compute the credible region for each parameter: bestp = output['bestp'] CRlo = np.zeros(nparams) CRhi = np.zeros(nparams) pdf = [] xpdf = [] for post, idx in zip(posterior.T, ifree): PDF, Xpdf, HPDmin = ms.cred_region(post) pdf.append(PDF) xpdf.append(Xpdf) CRlo[idx] = np.amin(Xpdf[PDF>HPDmin]) CRhi[idx] = np.amax(Xpdf[PDF>HPDmin]) # CR relative to the best-fitting value: CRlo[ifree] -= bestp[ifree] CRhi[ifree] -= bestp[ifree] # Get the mean and standard deviation from the posterior: meanp = np.zeros(nparams, np.double) # Parameters mean stdp = np.zeros(nparams, np.double) # Parameter standard deviation meanp[ifree] = np.mean(posterior, axis=0) stdp [ifree] = np.std(posterior, axis=0) for s in ishare: bestp[s] = bestp[-int(pstep[s])-1] meanp[s] = meanp[-int(pstep[s])-1] stdp [s] = stdp [-int(pstep[s])-1] CRlo [s] = CRlo [-int(pstep[s])-1] CRhi [s] = CRhi [-int(pstep[s])-1] output['CRlo'] = CRlo output['CRhi'] = CRhi output['stdp'] = stdp output['meanp'] = meanp output['ifree'] = ifree output['pnames'] = pnames output['texnames'] = texnames log.msg("\nParam name Best fit Lo HPD CR Hi HPD CR Mean Std dev S/N" "\n----------- ----------------------------------- ---------------------- ---------", width=80) for i in range(nparams): snr = "{:.1f}". format(np.abs(bestp[i])/stdp[i]) mean = "{: 11.4e}".format(meanp[i]) lo = "{: 11.4e}".format(CRlo[i]) hi = "{: 11.4e}".format(CRhi[i]) if i in ifree: # Free-fitting value pass elif i in ishare: # Shared value snr = "[share{:02d}]".format(-int(pstep[i])) else: # Fixed value snr = "[fixed]" mean = "{: 11.4e}".format(bestp[i]) log.msg("{:<11s} {:11.4e} {:>11s} {:>11s} {:>11s} {:10.4e} {:>9s}". format(pnames[i][0:11], bestp[i], lo, hi, mean, stdp[i], snr), width=160) fmt = len("{:.4f}".format(output['BIC'])) # Length of string formatting log.msg(" ") if chisqscale: log.msg("sqrt(reduced chi-squared) factor: {:{}.4f}". format(output['chisq_factor'], fmt), indent=2) log.msg("Best-parameter's chi-squared: {:{}.4f}". format(output['best_chisq'], fmt), indent=2) log.msg("Best-parameter's -2*log(posterior): {:{}.4f}". format(-2*output['best_log_post'], fmt), indent=2) log.msg("Bayesian Information Criterion: {:{}.4f}". format(output['BIC'], fmt), indent=2) log.msg("Reduced chi-squared: {:{}.4f}". format(output['red_chisq'], fmt), indent=2) log.msg("Standard deviation of residuals: {:.6g}\n". format(output['stddev_residuals']), indent=2) if savefile is not None or plots or closelog: log.msg("\nOutput sampler files:") # Save results (pop unpickables before saving, then put back): if savefile is not None: unpickables = ['dynesty_sampler'] unpickables = np.intersect1d(unpickables, list(output.keys())) tmp_outputs = {key: output.pop(key) for key in unpickables} np.savez(savefile, **output) output.update(tmp_outputs) log.msg("'{:s}'".format(savefile), indent=2) if plots: # Extract filename from savefile or default to sampler: fname = sampler if savefile is None else os.path.splitext(savefile)[0] # Include bestp in posterior plots: best_freepars = output['bestp'][ifree] if showbp else None # Trace plot: mp.trace(output['posterior'], zchain=output['zchain'], burnin=output['burnin'], pnames=texnames[ifree], savefile=fname+"_trace.png") log.msg("'{:s}'".format(fname+"_trace.png"), indent=2) # Pairwise posteriors: mp.pairwise(posterior, pnames=texnames[ifree], bestp=best_freepars, savefile=fname+"_pairwise.png") log.msg("'{:s}'".format(fname+"_pairwise.png"), indent=2) # Histograms: mp.histogram(posterior, pnames=texnames[ifree], bestp=best_freepars, savefile=fname+"_posterior.png", quantile=0.683, pdf=pdf, xpdf=xpdf) log.msg("'{:s}'".format(fname+"_posterior.png"), indent=2) # RMS vs bin size: if rms: RMS, RMSlo, RMShi, stderr, bs = ms.time_avg(output['best_model']-data) mp.rms(bs, RMS, stderr, RMSlo, RMShi, binstep=len(bs)//500+1, savefile=fname+"_RMS.png") log.msg("'{:s}'".format(fname+"_RMS.png"), indent=2) # Sort of guessing that indparams[0] is the X array for data as in y=y(x): if (indparams != [] and isinstance(indparams[0], (list, tuple, np.ndarray)) and np.size(indparams[0]) == ndata): try: mp.modelfit(data, uncert, indparams[0], output['best_model'], savefile=fname+"_model.png") log.msg("'{:s}'".format(fname+"_model.png"), indent=2) except: pass # Close the log file if necessary: if closelog: log.msg("'{:s}'".format(log.logname), indent=2) log.close() return output