# Copyright (C) 2016 Collin Capano # This program is free software; you can redistribute it and/or modify it # under the terms of the GNU General Public License as published by the # Free Software Foundation; either version 3 of the License, or (at your # option) any later version. # # This program is distributed in the hope that it will be useful, but # WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General # Public License for more details. # # You should have received a copy of the GNU General Public License along # with this program; if not, write to the Free Software Foundation, Inc., # 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA. """ This modules provides classes and functions for using the emcee_pt sampler packages for parameter estimation. """ import shlex import numpy import ptemcee import logging from pycbc.pool import choose_pool from .base import (BaseSampler, setup_output) from .base_mcmc import (BaseMCMC, EnsembleSupport, raw_samples_to_dict, get_optional_arg_from_config) from .base_multitemper import (read_betas_from_hdf, ensemble_compute_acf, ensemble_compute_acl) from ..burn_in import EnsembleMultiTemperedMCMCBurnInTests from pycbc.inference.io import PTEmceeFile from .. import models class PTEmceeSampler(EnsembleSupport, BaseMCMC, BaseSampler): """This class is used to construct the parallel-tempered ptemcee sampler. Parameters ---------- model : model A model from ``pycbc.inference.models``. nwalkers : int Number of walkers to use in sampler. ntemps : int, optional Specify the number of temps to use. Either this, ``Tmax``, or ``betas`` must be specified. Tmax : float, optional Specify the maximum temperature to use. This may be used with ``ntemps``; see :py:func:`ptemcee.make_ladder` for details. Either this, ``ntemps``, or ``betas`` must be specified. betas : list of float, optional Specify the betas to use. Must be provided if ``ntemps`` and ``Tmax`` are not given. Will override ``ntemps`` and ``Tmax`` if provided. adaptive : bool, optional Whether or not to use adaptive temperature levels. Default is False. adaptation_lag : int, optional Only used if ``adaptive`` is True; see :py:mod:`ptemcee.Sampler` for details. If not provided, will use ``ptemcee``'s default. adaptation_time : int, optional Only used if ``adaptive`` is True; see :py:mod:`ptemcee.Sampler` for details. If not provided, will use ``ptemcee``'s default. scale_factor : float, optional Scale factor used for the stretch proposal; see :py:mod:`ptemcee.Sampler` for details. If not provided, will use ``ptemcee``'s default. loglikelihood_function : str, optional Set the function to call from the model for the ``loglikelihood``. Default is ``loglikelihood``. nprocesses : int, optional The number of parallel processes to use. Default is 1 (no paralleliztion). use_mpi : bool, optional Use MPI for parallelization. Default (False) will use python's multiprocessing. """ name = "ptemcee" _io = PTEmceeFile burn_in_class = EnsembleMultiTemperedMCMCBurnInTests def __init__(self, model, nwalkers, ntemps=None, Tmax=None, betas=None, adaptive=False, adaptation_lag=None, adaptation_time=None, scale_factor=None, loglikelihood_function=None, checkpoint_interval=None, checkpoint_signal=None, nprocesses=1, use_mpi=False): self.model = model ndim = len(model.variable_params) # create temperature ladder if needed if ntemps is None and Tmax is None and betas is None: raise ValueError("must provide either ntemps/Tmax or betas") if betas is None: betas = ptemcee.make_ladder(ndim, ntemps=ntemps, Tmax=Tmax) # construct the keyword arguments to pass; if a kwarg is None, we # won't pass it, resulting in ptemcee's defaults being used kwargs = {} kwargs['adaptive'] = adaptive kwargs['betas'] = betas if adaptation_lag is not None: kwargs['adaptation_lag'] = adaptation_lag if adaptation_time is not None: kwargs['adaptation_time'] = adaptation_time if scale_factor is not None: kwargs['scale_factor'] = scale_factor # create a wrapper for calling the model if loglikelihood_function is None: loglikelihood_function = 'loglikelihood' # frustratingly, ptemcee does not support blob data, so we have to # turn it off model_call = models.CallModel(model, loglikelihood_function, return_all_stats=False) # these are used to help paralleize over multiple cores / MPI models._global_instance = model_call model_call = models._call_global_model prior_call = models._call_global_model_logprior self.pool = choose_pool(mpi=use_mpi, processes=nprocesses) # construct the sampler self._sampler = ptemcee.Sampler(nwalkers=nwalkers, ndim=ndim, logl=model_call, logp=prior_call, mapper=self.pool.map, **kwargs) self.nwalkers = nwalkers self._ntemps = ntemps self._checkpoint_interval = checkpoint_interval self._checkpoint_signal = checkpoint_signal # we'll initialize ensemble and chain to None self._chain = None self._ensemble = None @property def io(self): return self._io @property def ntemps(self): """The number of temeratures that are set.""" return self._ntemps @property def base_shape(self): return (self.ntemps, self.nwalkers,) @property def betas(self): """Returns the beta history currently in memory.""" # chain betas has shape niterations x ntemps; transpose to # ntemps x niterations return self._chain.betas.transpose() @property def starting_betas(self): """Returns the betas that were used at startup.""" # the initial betas that were used return self._sampler.betas @property def adaptive(self): """Whether or not the betas are adapted.""" return self._sampler.adaptive @property def adaptation_lag(self): """The adaptation lag for the beta evolution.""" return self._sampler.adaptation_lag @property def adaptation_time(self): """The adaptation time for the beta evolution.""" return self._sampler.adaptation_time @property def scale_factor(self): """The scale factor used by ptemcee.""" return self._sampler.scale_factor @property def ensemble(self): """Returns the current ptemcee ensemble. The ensemble stores the current location of and temperatures of the walkers. If the ensemble hasn't been setup yet, will set one up using p0 for the positions. If set_p0 hasn't been run yet, this will result in a ValueError. """ if self._ensemble is None: if self._p0 is None: raise ValueError("initial positions not set; run set_p0") # use the global numpy random state rstate = numpy.random.mtrand._rand # self._p0 has base_shape x ndim = ntemps x nwalkers x ndim (see # BaseMCMC.set_p0). ptemcee's Ensemble expects # ntemps x nwalkers x ndim... so we're good self._ensemble = self._sampler.ensemble(self._p0, rstate) return self._ensemble @property def _pos(self): """Uses the ensemble for the position.""" # BaseMCMC expects _pos to have shape ntemps x nwalkers x ndim, # which is the same shape as ensemble.x return self.ensemble.x @property def chain(self): """The current chain of samples in memory. The chain is returned as a :py:mod:`ptemcee.chain.Chain` instance. If no chain has been created yet (``_chain`` is None), then will create a new chain using the current ``ensemble``. """ if self._chain is None: # create a chain self._chain = ptemcee.chain.Chain(self.ensemble) return self._chain def clear_samples(self): """Clears the chain and blobs from memory. """ # store the iteration that the clear is occuring on self._lastclear = self.niterations self._itercounter = 0 # set _chain to None; this will both cause the current chain to # get garbage collected, and will cause a new chain to be created # the next time self.chain is called self._chain = None @property def samples(self): """A dict mapping ``variable_params`` to arrays of samples currently in memory. The arrays have shape ``ntemps x nwalkers x niterations``. """ # chain.x has shape niterations x ntemps x nwalkers x ndim # we'll transpose to ntemps x nwalkers x niterations x ndim raw_samples = self._chain.x.transpose((1, 2, 0, 3)) return raw_samples_to_dict(self, raw_samples) @property def model_stats(self): """Returns the log likelihood ratio and log prior as a dict of arrays. The returned array has shape ntemps x nwalkers x niterations. Unfortunately, because ``ptemcee`` does not have blob support, this will only return the loglikelihood and logprior (with the logjacobian set to zero) regardless of what stats the model can return. .. warning:: Since the ``logjacobian`` is not saved by ``ptemcee``, the ``logprior`` returned here is the log of the prior pdf in the sampling coordinate frame rather than the variable params frame. This differs from the variable params frame by the log of the Jacobian of the transform from one frame to the other. If no sampling transforms were used, then the ``logprior`` is the same. """ # log likelihood and prior have shape # niterations x ntemps x nwalkers; we'll tranpose to have shape # ntemps x nwalkers x niterations logl = self._chain.logl.transpose((1, 2, 0)) logp = self._chain.logP.transpose((1, 2, 0)) logjacobian = numpy.zeros(logp.shape) return {'loglikelihood': logl, 'logprior': logp, 'logjacobian': logjacobian} def set_state_from_file(self, filename): """Sets the state of the sampler back to the instance saved in a file. """ with self.io(filename, 'r') as fp: rstate = fp.read_random_state() # set the numpy random state numpy.random.set_state(rstate) # set the ensemble to its last state ensemble = self.ensemble for attr, val in fp.read_ensemble_attrs().items(): setattr(ensemble, attr, val) ensemble.betas = fp.read_betas(iteration=-1) ensemble.time = fp.niterations def run_mcmc(self, niterations): """Advance the ensemble for a number of samples. Parameters ---------- niterations : int Number of samples to get from sampler. """ self.chain.run(niterations) @classmethod def calculate_logevidence(cls, filename, thin_start=None, thin_end=None, thin_interval=None): """Calculates the log evidence from the given file. This uses ``ptemcee``'s thermodynamic integration. Parameters ---------- filename : str Name of the file to read the samples from. Should be an ``PTEmceeFile``. thin_start : int Index of the sample to begin returning stats. Default is to read stats after burn in. To start from the beginning set thin_start to 0. thin_interval : int Interval to accept every i-th sample. Default is to use the `fp.acl`. If `fp.acl` is not set, then use all stats (set thin_interval to 1). thin_end : int Index of the last sample to read. If not given then `fp.niterations` is used. Returns ------- lnZ : float The estimate of log of the evidence. dlnZ : float The error on the estimate. """ with cls._io(filename, 'r') as fp: logls = fp.read_raw_samples(['loglikelihood'], thin_start=thin_start, thin_interval=thin_interval, thin_end=thin_end, temps='all', flatten=False) logls = logls['loglikelihood'] # we need the betas that were used betas = fp.read_betas(thin_start=thin_start, thin_interval=thin_interval, thin_end=thin_end) # we'll separate betas out by their unique temperatures # there's probably a faster way to do this... mean_logls = [] unique_betas = [] ntemps = betas.shape[0] for ti in range(ntemps): ubti, idx = numpy.unique(betas[ti, :], return_inverse=True) unique_idx = numpy.unique(idx) loglsti = logls[ti, :, :] for ii in unique_idx: # average over the walkers and iterations with the same # betas getiters = numpy.where(ii == unique_idx)[0] mean_logls.append(loglsti[:, getiters].mean()) unique_betas.append(ubti[ii]) return ptemcee.util.thermodynamic_integration_log_evidence( numpy.array(unique_betas), numpy.array(mean_logls)) @staticmethod def compute_acf(filename, **kwargs): r"""Computes the autocorrelation function. Calls :py:func:`base_multitemper.ensemble_compute_acf`; see that function for details. Parameters ---------- filename : str Name of a samples file to compute ACFs for. \**kwargs : All other keyword arguments are passed to :py:func:`base_multitemper.ensemble_compute_acf`. Returns ------- dict : Dictionary of arrays giving the ACFs for each parameter. If ``per-walker=True`` is passed as a keyword argument, the arrays will have shape ``ntemps x nwalkers x niterations``. Otherwise, the returned array will have shape ``ntemps x niterations``. """ return ensemble_compute_acf(filename, **kwargs) @staticmethod def compute_acl(filename, **kwargs): r"""Computes the autocorrelation length. Calls :py:func:`base_multitemper.ensemble_compute_acl`; see that function for details. Parameters ---------- filename : str Name of a samples file to compute ACLs for. \**kwargs : All other keyword arguments are passed to :py:func:`base_multitemper.ensemble_compute_acl`. Returns ------- dict A dictionary of ntemps-long arrays of the ACLs of each parameter. """ return ensemble_compute_acl(filename, **kwargs) @classmethod def from_config(cls, cp, model, output_file=None, nprocesses=1, use_mpi=False): """Loads the sampler from the given config file. The following options are retrieved in the ``[sampler]`` section: * ``name = STR`` : Required. This must match the sampler's name. * ``nwalkers = INT`` : Required. The number of walkers to use. * ``ntemps = INT`` : The number of temperatures to use. This may be used in combination with ``Tmax``. Either this, ``Tmax``, ``betas`` or ``betas-file`` must be provided. * ``tmax = FLOAT`` : The maximum temperature to use. This may be used in combination with ``ntemps``, or alone. * ``betas = FLOAT1 FLOAT2 [...]`` : Space-separated list of (intial) inverse temperatures ("betas") to use. This sets both the number of temperatures and the tmax. A ``ValueError`` will be raised if both this and ``ntemps`` or ``Tmax`` are provided. * ``betas-file = STR`` : Path to an hdf file containing the inverse temperatures ("betas") to use. The betas will be retrieved from the file's ``.attrs['betas']``. A ``ValueError`` will be raised if both this and ``betas`` are provided. * ``adaptive =`` : If provided, temperature adaptation will be turned on. * ``adaptation-lag = INT`` : The adaptation lag to use (see ptemcee for details). * ``adaptation-time = INT`` : The adaptation time to use (see ptemcee for details). * ``scale-factor = FLOAT`` : The scale factor to use for the emcee stretch. * ``niterations = INT`` : The number of iterations to run the sampler for. Either this or ``effective-nsamples`` must be provided (but not both). * ``effective-nsamples = INT`` : Run the sampler until the given number of effective samples are obtained. A ``checkpoint-interval`` must also be provided in this case. Either this or ``niterations`` must be provided (but not both). * ``thin-interval = INT`` : Thin the samples by the given value before saving to disk. May provide this, or ``max-samples-per-chain``, but not both. If neither options are provided, will save all samples. * ``max-samples-per-chain = INT`` : Thin the samples such that the number of samples per chain per temperature that are saved to disk never exceeds the given value. May provide this, or ``thin-interval``, but not both. If neither options are provided, will save all samples. * ``checkpoint-interval = INT`` : Sets the checkpoint interval to use. Must be provided if using ``effective-nsamples``. * ``checkpoint-signal = STR`` : Set the checkpoint signal, e.g., "USR2". Optional. * ``logl-function = STR`` : The attribute of the model to use for the loglikelihood. If not provided, will default to ``loglikelihood``. Settings for burn-in tests are read from ``[sampler-burn_in]``. In particular, the ``burn-in-test`` option is used to set the burn in tests to perform. See :py:func:`EnsembleMultiTemperedMCMCBurnInTests.from_config` for details. If no ``burn-in-test`` is provided, no burn in tests will be carried out. Parameters ---------- cp : WorkflowConfigParser instance Config file object to parse. model : pycbc.inference.model.BaseModel instance The model to use. output_file : str, optional The name of the output file to checkpoint and write results to. nprocesses : int, optional The number of parallel processes to use. Default is 1. use_mpi : bool, optional Use MPI for parallelization. Default is False. Returns ------- EmceePTSampler : The sampler instance. """ section = "sampler" # check name assert cp.get(section, "name") == cls.name, ( "name in section [sampler] must match mine") # get the number of walkers to use nwalkers = int(cp.get(section, "nwalkers")) # get the checkpoint interval, if it's specified checkpoint_interval = cls.checkpoint_from_config(cp, section) checkpoint_signal = cls.ckpt_signal_from_config(cp, section) optargs = {} # get the temperature level settings ntemps = get_optional_arg_from_config(cp, section, 'ntemps', int) if ntemps is not None: optargs['ntemps'] = ntemps tmax = get_optional_arg_from_config(cp, section, 'tmax', float) if tmax is not None: optargs['Tmax'] = tmax betas = get_optional_arg_from_config(cp, section, 'betas') if betas is not None: # convert to list sorted in descencding order betas = numpy.sort(list(map(float, shlex.split(betas))))[::-1] optargs['betas'] = betas betas_file = get_optional_arg_from_config(cp, section, 'betas-file') if betas_file is not None: optargs['betas'] = read_betas_from_hdf(betas_file) # check for consistency if betas is not None and betas_file is not None: raise ValueError("provide either betas or betas-file, not both") if 'betas' in optargs and (ntemps is not None or tmax is not None): raise ValueError("provide either ntemps/tmax or betas/betas-file, " "not both") # adaptation parameters adaptive = get_optional_arg_from_config(cp, section, 'adaptive') if adaptive is not None: optargs['adaptive'] = True else: optargs['adaptive'] = False adaptation_lag = get_optional_arg_from_config(cp, section, 'adaptation-lag', int) if adaptation_lag is not None: optargs['adaptation_lag'] = adaptation_lag adaptation_time = get_optional_arg_from_config(cp, section, 'adaptation-time', int) if adaptation_time is not None: optargs['adaptation_time'] = adaptation_time scale_factor = get_optional_arg_from_config(cp, section, 'scale-factor', float) if scale_factor is not None: optargs['scale_factor'] = scale_factor # get the loglikelihood function logl = get_optional_arg_from_config(cp, section, 'logl-function') obj = cls(model, nwalkers, checkpoint_interval=checkpoint_interval, checkpoint_signal=checkpoint_signal, loglikelihood_function=logl, nprocesses=nprocesses, use_mpi=use_mpi, **optargs) # set target obj.set_target_from_config(cp, section) # add burn-in if it's specified obj.set_burn_in_from_config(cp) # set prethin options obj.set_thin_interval_from_config(cp, section) # Set up the output file setup_output(obj, output_file) if not obj.new_checkpoint: obj.resume_from_checkpoint() else: obj.set_start_from_config(cp) return obj def write_results(self, filename): """Writes samples, model stats, acceptance fraction, and random state to the given file. Parameters ---------- filename : str The file to write to. The file is opened using the ``io`` class in an an append state. """ with self.io(filename, 'a') as fp: # write samples fp.write_samples(self.samples, parameters=self.model.variable_params, last_iteration=self.niterations) # write stats fp.write_samples(self.model_stats, last_iteration=self.niterations) # write random state fp.write_random_state() # write betas fp.write_betas(self.betas, last_iteration=self.niterations) # write random state fp.write_random_state() # write attributes of the ensemble fp.write_ensemble_attrs(self.ensemble) def _correctjacobian(self, samples): """Corrects the log jacobian values stored on disk. Parameters ---------- samples : dict Dictionary of the samples. """ # flatten samples for evaluating orig_shape = list(samples.values())[0].shape flattened_samples = {p: arr.ravel() for p, arr in list(samples.items())} # convert to a list of tuples so we can use map function params = list(flattened_samples.keys()) size = flattened_samples[params[0]].size logj = numpy.zeros(size) for ii in range(size): these_samples = {p: flattened_samples[p][ii] for p in params} these_samples = self.model.sampling_transforms.apply(these_samples) self.model.update(**these_samples) logj[ii] = self.model.logjacobian return logj.reshape(orig_shape) def finalize(self): """Calculates the log evidence and writes to the checkpoint file. If sampling transforms were used, this also corrects the jacobian stored on disk. The thin start/interval/end for calculating the log evidence are retrieved from the checkpoint file's thinning attributes. """ if self.model.sampling_transforms is not None: # fix the lobjacobian values stored on disk logging.info("Correcting logjacobian values on disk") with self.io(self.checkpoint_file, 'r') as fp: samples = fp.read_raw_samples(self.variable_params, thin_start=0, thin_interval=1, thin_end=None, temps='all', flatten=False) logjacobian = self._correctjacobian(samples) # write them back out for fn in [self.checkpoint_file, self.backup_file]: with self.io(fn, "a") as fp: fp[fp.samples_group]['logjacobian'][()] = logjacobian logging.info("Calculating log evidence") # get the thinning settings with self.io(self.checkpoint_file, 'r') as fp: thin_start = fp.thin_start thin_interval = fp.thin_interval thin_end = fp.thin_end # calculate logz, dlogz = self.calculate_logevidence( self.checkpoint_file, thin_start=thin_start, thin_end=thin_end, thin_interval=thin_interval) logging.info("log Z, dlog Z: {}, {}".format(logz, dlogz)) # write to both the checkpoint and backup for fn in [self.checkpoint_file, self.backup_file]: with self.io(fn, "a") as fp: fp.write_logevidence(logz, dlogz)