Source code for pymc.sampling.mcmc

#   Copyright 2020 The PyMC Developers
#
#   Licensed under the Apache License, Version 2.0 (the "License");
#   you may not use this file except in compliance with the License.
#   You may obtain a copy of the License at
#
#       http://www.apache.org/licenses/LICENSE-2.0
#
#   Unless required by applicable law or agreed to in writing, software
#   distributed under the License is distributed on an "AS IS" BASIS,
#   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
#   See the License for the specific language governing permissions and
#   limitations under the License.

"""Functions for MCMC sampling."""

import logging
import pickle
import sys
import time
import warnings

from collections import defaultdict
from copy import copy
from typing import Iterator, List, Optional, Sequence, Tuple, Union

import aesara.gradient as tg
import numpy as np

from arviz import InferenceData
from fastprogress.fastprogress import progress_bar
from typing_extensions import TypeAlias

import pymc as pm

from pymc.backends import _init_trace
from pymc.backends.base import BaseTrace, MultiTrace, _choose_chains
from pymc.blocking import DictToArrayBijection
from pymc.exceptions import SamplingError
from pymc.initial_point import PointType, StartDict, make_initial_point_fns_per_chain
from pymc.model import Model, modelcontext
from pymc.sampling.parallel import Draw, _cpu_count
from pymc.sampling.population import _sample_population
from pymc.stats.convergence import log_warning_stats, run_convergence_checks
from pymc.step_methods import NUTS, CompoundStep, DEMetropolis
from pymc.step_methods.arraystep import BlockedStep, PopulationArrayStepShared
from pymc.step_methods.hmc import quadpotential
from pymc.util import (
    RandomSeed,
    RandomState,
    _get_seeds_per_chain,
    drop_warning_stat,
    get_untransformed_name,
    is_transformed_name,
)
from pymc.vartypes import discrete_types

sys.setrecursionlimit(10000)

__all__ = [
    "sample",
    "iter_sample",
    "init_nuts",
]

Step: TypeAlias = Union[BlockedStep, CompoundStep]


_log = logging.getLogger("pymc")


def instantiate_steppers(
    model, steps: List[Step], selected_steps, step_kwargs=None
) -> Union[Step, List[Step]]:
    """Instantiate steppers assigned to the model variables.

    This function is intended to be called automatically from ``sample()``, but
    may be called manually.

    Parameters
    ----------
    model : Model object
        A fully-specified model object.
    steps : list, array_like of shape (selected_steps, )
        A list of zero or more step function instances that have been assigned to some subset of
        the model's parameters.
    selected_steps : dict
        A dictionary that maps a step method class to a list of zero or more model variables.
    step_kwargs : dict, default=None
        Parameters for the samplers. Keys are the lower case names of
        the step method, values a dict of arguments. Defaults to None.

    Returns
    -------
    methods : list or step
        List of step methods associated with the model's variables, or step method
        if there is only one.
    """
    if step_kwargs is None:
        step_kwargs = {}

    used_keys = set()
    for step_class, vars in selected_steps.items():
        if vars:
            args = step_kwargs.get(step_class.name, {})
            used_keys.add(step_class.name)
            step = step_class(vars=vars, model=model, **args)
            steps.append(step)

    unused_args = set(step_kwargs).difference(used_keys)
    if unused_args:
        raise ValueError("Unused step method arguments: %s" % unused_args)

    if len(steps) == 1:
        return steps[0]

    return steps


def assign_step_methods(model, step=None, methods=None, step_kwargs=None):
    """Assign model variables to appropriate step methods.

    Passing a specified model will auto-assign its constituent stochastic
    variables to step methods based on the characteristics of the variables.
    This function is intended to be called automatically from ``sample()``, but
    may be called manually. Each step method passed should have a
    ``competence()`` method that returns an ordinal competence value
    corresponding to the variable passed to it. This value quantifies the
    appropriateness of the step method for sampling the variable.

    Parameters
    ----------
    model : Model object
        A fully-specified model object.
    step : step function or iterable of step functions, optional
        One or more step functions that have been assigned to some subset of
        the model's parameters. Defaults to ``None`` (no assigned variables).
    methods : iterable of step method classes, optional
        The set of step methods from which the function may choose. Defaults
        to the main step methods provided by PyMC.
    step_kwargs : dict, optional
        Parameters for the samplers. Keys are the lower case names of
        the step method, values a dict of arguments.

    Returns
    -------
    methods : list
        List of step methods associated with the model's variables.
    """
    steps = []
    assigned_vars = set()

    if methods is None:
        methods = pm.STEP_METHODS

    if step is not None:
        try:
            steps += list(step)
        except TypeError:
            steps.append(step)
        for step in steps:
            assigned_vars = assigned_vars.union(set(step.vars))

    # Use competence classmethods to select step methods for remaining
    # variables
    selected_steps = defaultdict(list)
    model_logp = model.logp()

    for var in model.value_vars:
        if var not in assigned_vars:
            # determine if a gradient can be computed
            has_gradient = var.dtype not in discrete_types
            if has_gradient:
                try:
                    tg.grad(model_logp, var)
                except (NotImplementedError, tg.NullTypeGradError):
                    has_gradient = False

            # select the best method
            rv_var = model.values_to_rvs[var]
            selected = max(
                methods,
                key=lambda method, var=rv_var, has_gradient=has_gradient: method._competence(
                    var, has_gradient
                ),
            )
            selected_steps[selected].append(var)

    return instantiate_steppers(model, steps, selected_steps, step_kwargs)


def _print_step_hierarchy(s: Step, level: int = 0) -> None:
    if isinstance(s, CompoundStep):
        _log.info(">" * level + "CompoundStep")
        for i in s.methods:
            _print_step_hierarchy(i, level + 1)
    else:
        varnames = ", ".join(
            [
                get_untransformed_name(v.name) if is_transformed_name(v.name) else v.name
                for v in s.vars
            ]
        )
        _log.info(">" * level + f"{s.__class__.__name__}: [{varnames}]")


def all_continuous(vars):
    """Check that vars not include discrete variables"""
    if any([(var.dtype in discrete_types) for var in vars]):
        return False
    else:
        return True


[docs]def sample( draws: int = 1000, step=None, init: str = "auto", n_init: int = 200_000, initvals: Optional[Union[StartDict, Sequence[Optional[StartDict]]]] = None, trace: Optional[BaseTrace] = None, chains: Optional[int] = None, cores: Optional[int] = None, tune: int = 1000, progressbar: bool = True, model=None, random_seed: RandomState = None, discard_tuned_samples: bool = True, compute_convergence_checks: bool = True, callback=None, jitter_max_retries: int = 10, *, return_inferencedata: bool = True, keep_warning_stat: bool = False, idata_kwargs: dict = None, mp_ctx=None, **kwargs, ) -> Union[InferenceData, MultiTrace]: r"""Draw samples from the posterior using the given step methods. Multiple step methods are supported via compound step methods. Parameters ---------- draws : int The number of samples to draw. Defaults to 1000. The number of tuned samples are discarded by default. See ``discard_tuned_samples``. init : str Initialization method to use for auto-assigned NUTS samplers. See `pm.init_nuts` for a list of all options. This argument is ignored when manually passing the NUTS step method. step : function or iterable of functions A step function or collection of functions. If there are variables without step methods, step methods for those variables will be assigned automatically. By default the NUTS step method will be used, if appropriate to the model. n_init : int Number of iterations of initializer. Only works for 'ADVI' init methods. initvals : optional, dict, array of dict Dict or list of dicts with initial value strategies to use instead of the defaults from `Model.initial_values`. The keys should be names of transformed random variables. Initialization methods for NUTS (see ``init`` keyword) can overwrite the default. trace : backend, optional A backend instance or None. If None, the NDArray backend is used. chains : int The number of chains to sample. Running independent chains is important for some convergence statistics and can also reveal multiple modes in the posterior. If ``None``, then set to either ``cores`` or 2, whichever is larger. cores : int The number of chains to run in parallel. If ``None``, set to the number of CPUs in the system, but at most 4. tune : int Number of iterations to tune, defaults to 1000. Samplers adjust the step sizes, scalings or similar during tuning. Tuning samples will be drawn in addition to the number specified in the ``draws`` argument, and will be discarded unless ``discard_tuned_samples`` is set to False. progressbar : bool, optional default=True Whether or not to display a progress bar in the command line. The bar shows the percentage of completion, the sampling speed in samples per second (SPS), and the estimated remaining time until completion ("expected time of arrival"; ETA). model : Model (optional if in ``with`` context) Model to sample from. The model needs to have free random variables. random_seed : int, array-like of int, RandomState or Generator, optional Random seed(s) used by the sampling steps. If a list, tuple or array of ints is passed, each entry will be used to seed each chain. A ValueError will be raised if the length does not match the number of chains. discard_tuned_samples : bool Whether to discard posterior samples of the tune interval. compute_convergence_checks : bool, default=True Whether to compute sampler statistics like Gelman-Rubin and ``effective_n``. callback : function, default=None A function which gets called for every sample from the trace of a chain. The function is called with the trace and the current draw and will contain all samples for a single trace. the ``draw.chain`` argument can be used to determine which of the active chains the sample is drawn from. Sampling can be interrupted by throwing a ``KeyboardInterrupt`` in the callback. jitter_max_retries : int Maximum number of repeated attempts (per chain) at creating an initial matrix with uniform jitter that yields a finite probability. This applies to ``jitter+adapt_diag`` and ``jitter+adapt_full`` init methods. return_inferencedata : bool Whether to return the trace as an :class:`arviz:arviz.InferenceData` (True) object or a `MultiTrace` (False). Defaults to `True`. idata_kwargs : dict, optional Keyword arguments for :func:`pymc.to_inference_data` keep_warning_stat : bool If ``True`` the "warning" stat emitted by, for example, HMC samplers will be kept in the returned ``idata.sample_stat`` group. This leads to the ``idata`` not supporting ``.to_netcdf()`` or ``.to_zarr()`` and should only be set to ``True`` if you intend to use the "warning" objects right away. Defaults to ``False`` such that ``pm.drop_warning_stat`` is applied automatically, making the ``InferenceData`` compatible with saving. mp_ctx : multiprocessing.context.BaseContent A multiprocessing context for parallel sampling. See multiprocessing documentation for details. Returns ------- trace : pymc.backends.base.MultiTrace or arviz.InferenceData A ``MultiTrace`` or ArviZ ``InferenceData`` object that contains the samples. Notes ----- Optional keyword arguments can be passed to ``sample`` to be delivered to the ``step_method``\ s used during sampling. For example: 1. ``target_accept`` to NUTS: nuts={'target_accept':0.9} 2. ``transit_p`` to BinaryGibbsMetropolis: binary_gibbs_metropolis={'transit_p':.7} Note that available step names are: ``nuts``, ``hmc``, ``metropolis``, ``binary_metropolis``, ``binary_gibbs_metropolis``, ``categorical_gibbs_metropolis``, ``DEMetropolis``, ``DEMetropolisZ``, ``slice`` The NUTS step method has several options including: * target_accept : float in [0, 1]. The step size is tuned such that we approximate this acceptance rate. Higher values like 0.9 or 0.95 often work better for problematic posteriors. This argument can be passed directly to sample. * max_treedepth : The maximum depth of the trajectory tree * step_scale : float, default 0.25 The initial guess for the step size scaled down by :math:`1/n**(1/4)`, where n is the dimensionality of the parameter space Alternatively, if you manually declare the ``step_method``\ s, within the ``step`` kwarg, then you can address the ``step_method`` kwargs directly. e.g. for a CompoundStep comprising NUTS and BinaryGibbsMetropolis, you could send :: step=[pm.NUTS([freeRV1, freeRV2], target_accept=0.9), pm.BinaryGibbsMetropolis([freeRV3], transit_p=.7)] You can find a full list of arguments in the docstring of the step methods. Examples -------- .. code:: ipython In [1]: import pymc as pm ...: n = 100 ...: h = 61 ...: alpha = 2 ...: beta = 2 In [2]: with pm.Model() as model: # context management ...: p = pm.Beta("p", alpha=alpha, beta=beta) ...: y = pm.Binomial("y", n=n, p=p, observed=h) ...: idata = pm.sample() In [3]: az.summary(idata, kind="stats") Out[3]: mean sd hdi_3% hdi_97% p 0.609 0.047 0.528 0.699 """ if "start" in kwargs: if initvals is not None: raise ValueError("Passing both `start` and `initvals` is not supported.") warnings.warn( "The `start` kwarg was renamed to `initvals` and can now do more. Please check the docstring.", FutureWarning, stacklevel=2, ) initvals = kwargs.pop("start") if "target_accept" in kwargs: if "nuts" in kwargs and "target_accept" in kwargs["nuts"]: raise ValueError( "`target_accept` was defined twice. Please specify it either as a direct keyword argument or in the `nuts` kwarg." ) if "nuts" in kwargs: kwargs["nuts"]["target_accept"] = kwargs.pop("target_accept") else: kwargs = {"nuts": {"target_accept": kwargs.pop("target_accept")}} if isinstance(trace, list): raise DeprecationWarning( "We have removed support for partial traces because it simplified things." " Please open an issue if & why this is a problem for you." ) model = modelcontext(model) if not model.free_RVs: raise SamplingError( "Cannot sample from the model, since the model does not contain any free variables." ) if cores is None: cores = min(4, _cpu_count()) if chains is None: chains = max(2, cores) if random_seed == -1: random_seed = None random_seed_list = _get_seeds_per_chain(random_seed, chains) if not discard_tuned_samples and not return_inferencedata: warnings.warn( "Tuning samples will be included in the returned `MultiTrace` object, which can lead to" " complications in your downstream analysis. Please consider to switch to `InferenceData`:\n" "`pm.sample(..., return_inferencedata=True)`", UserWarning, stacklevel=2, ) # small trace warning if draws == 0: msg = "Tuning was enabled throughout the whole trace." _log.warning(msg) elif draws < 500: msg = "Only %s samples in chain." % draws _log.warning(msg) draws += tune auto_nuts_init = True if step is not None: if isinstance(step, CompoundStep): for method in step.methods: if isinstance(method, NUTS): auto_nuts_init = False elif isinstance(step, NUTS): auto_nuts_init = False initial_points = None step = assign_step_methods(model, step, methods=pm.STEP_METHODS, step_kwargs=kwargs) if isinstance(step, list): step = CompoundStep(step) elif isinstance(step, NUTS) and auto_nuts_init: if "nuts" in kwargs: nuts_kwargs = kwargs.pop("nuts") [kwargs.setdefault(k, v) for k, v in nuts_kwargs.items()] _log.info("Auto-assigning NUTS sampler...") initial_points, step = init_nuts( init=init, chains=chains, n_init=n_init, model=model, random_seed=random_seed_list, progressbar=progressbar, jitter_max_retries=jitter_max_retries, tune=tune, initvals=initvals, **kwargs, ) if initial_points is None: # Time to draw/evaluate numeric start points for each chain. ipfns = make_initial_point_fns_per_chain( model=model, overrides=initvals, jitter_rvs=set(), chains=chains, ) initial_points = [ipfn(seed) for ipfn, seed in zip(ipfns, random_seed_list)] # One final check that shapes and logps at the starting points are okay. for ip in initial_points: model.check_start_vals(ip) _check_start_shape(model, ip) sample_args = { "draws": draws, "step": step, "start": initial_points, "trace": trace, "chains": chains, "tune": tune, "progressbar": progressbar, "model": model, "cores": cores, "callback": callback, "discard_tuned_samples": discard_tuned_samples, } parallel_args = { "mp_ctx": mp_ctx, } sample_args.update(kwargs) has_population_samplers = np.any( [ isinstance(m, PopulationArrayStepShared) for m in (step.methods if isinstance(step, CompoundStep) else [step]) ] ) parallel = cores > 1 and chains > 1 and not has_population_samplers # At some point it was decided that PyMC should not set a global seed by default, # unless the user specified a seed. This is a symptom of the fact that PyMC samplers # are built around global seeding. This branch makes sure we maintain this unspoken # rule. See https://github.com/pymc-devs/pymc/pull/1395. if parallel: # For parallel sampling we can pass the list of random seeds directly, as # global seeding will only be called inside each process sample_args["random_seed"] = random_seed_list else: # We pass None if the original random seed was None. The single core sampler # methods will only set a global seed when it is not None. sample_args["random_seed"] = random_seed if random_seed is None else random_seed_list t_start = time.time() if parallel: _log.info(f"Multiprocess sampling ({chains} chains in {cores} jobs)") _print_step_hierarchy(step) try: mtrace = _mp_sample(**sample_args, **parallel_args) except pickle.PickleError: _log.warning("Could not pickle model, sampling singlethreaded.") _log.debug("Pickling error:", exc_info=True) parallel = False except AttributeError as e: if not str(e).startswith("AttributeError: Can't pickle"): raise _log.warning("Could not pickle model, sampling singlethreaded.") _log.debug("Pickling error:", exc_info=True) parallel = False if not parallel: if has_population_samplers: has_demcmc = np.any( [ isinstance(m, DEMetropolis) for m in (step.methods if isinstance(step, CompoundStep) else [step]) ] ) _log.info(f"Population sampling ({chains} chains)") initial_point_model_size = sum(initial_points[0][n.name].size for n in model.value_vars) if has_demcmc and chains < 3: raise ValueError( "DEMetropolis requires at least 3 chains. " "For this {}-dimensional model you should use ≥{} chains".format( initial_point_model_size, initial_point_model_size + 1 ) ) if has_demcmc and chains <= initial_point_model_size: warnings.warn( "DEMetropolis should be used with more chains than dimensions! " "(The model has {} dimensions.)".format(initial_point_model_size), UserWarning, stacklevel=2, ) _print_step_hierarchy(step) mtrace = _sample_population(parallelize=cores > 1, **sample_args) else: _log.info(f"Sequential sampling ({chains} chains in 1 job)") _print_step_hierarchy(step) mtrace = _sample_many(**sample_args) t_sampling = time.time() - t_start # count the number of tune/draw iterations that happened # ideally via the "tune" statistic, but not all samplers record it! if "tune" in mtrace.stat_names: stat = mtrace.get_sampler_stats("tune", chains=0) # when CompoundStep is used, the stat is 2 dimensional! if len(stat.shape) == 2: stat = stat[:, 0] stat = tuple(stat) n_tune = stat.count(True) n_draws = stat.count(False) else: # these may be wrong when KeyboardInterrupt happened, but they're better than nothing n_tune = min(tune, len(mtrace)) n_draws = max(0, len(mtrace) - n_tune) if discard_tuned_samples: mtrace = mtrace[n_tune:] # save metadata in SamplerReport mtrace.report._n_tune = n_tune mtrace.report._n_draws = n_draws mtrace.report._t_sampling = t_sampling n_chains = len(mtrace.chains) _log.info( f'Sampling {n_chains} chain{"s" if n_chains > 1 else ""} for {n_tune:_d} tune and {n_draws:_d} draw iterations ' f"({n_tune*n_chains:_d} + {n_draws*n_chains:_d} draws total) " f"took {t_sampling:.0f} seconds." ) mtrace.report._log_summary() idata = None if compute_convergence_checks or return_inferencedata: ikwargs = dict(model=model, save_warmup=not discard_tuned_samples) if idata_kwargs: ikwargs.update(idata_kwargs) idata = pm.to_inference_data(mtrace, **ikwargs) if compute_convergence_checks: if draws - tune < 100: warnings.warn( "The number of samples is too small to check convergence reliably.", stacklevel=2, ) else: convergence_warnings = run_convergence_checks(idata, model) mtrace.report._add_warnings(convergence_warnings) if return_inferencedata: # By default we drop the "warning" stat which contains `SamplerWarning` # objects that can not be stored with `.to_netcdf()`. if not keep_warning_stat: return drop_warning_stat(idata) return idata return mtrace
def _check_start_shape(model, start: PointType): """Checks that the prior evaluations and initial points have identical shapes. Parameters ---------- model : pm.Model The current model on context. start : dict The complete dictionary mapping (transformed) variable names to numeric initial values. """ e = "" try: actual_shapes = model.eval_rv_shapes() except NotImplementedError as ex: warnings.warn(f"Unable to validate shapes: {ex.args[0]}", UserWarning) return for name, sval in start.items(): ashape = actual_shapes.get(name) sshape = np.shape(sval) if ashape != tuple(sshape): e += f"\nExpected shape {ashape} for var '{name}', got: {sshape}" if e != "": raise ValueError(f"Bad shape in start point:{e}") def _sample_many( draws: int, chains: int, start: Sequence[PointType], random_seed: Optional[Sequence[RandomSeed]], step, callback=None, **kwargs, ) -> MultiTrace: """Samples all chains sequentially. Parameters ---------- draws: int The number of samples to draw chains: int Total number of chains to sample. start: list Starting points for each chain random_seed: list of random seeds, optional A list of seeds, one for each chain step: function Step function Returns ------- mtrace: MultiTrace Contains samples of all chains """ traces: List[BaseTrace] = [] for i in range(chains): trace = _sample( draws=draws, chain=i, start=start[i], step=step, random_seed=None if random_seed is None else random_seed[i], callback=callback, **kwargs, ) if trace is None: if len(traces) == 0: raise ValueError("Sampling stopped before a sample was created.") else: break elif len(trace) < draws: if len(traces) == 0: traces.append(trace) break else: traces.append(trace) return MultiTrace(traces) def _sample( *, chain: int, progressbar: bool, random_seed: RandomSeed, start: PointType, draws: int, step=None, trace: Optional[BaseTrace] = None, tune: int, model: Optional[Model] = None, callback=None, **kwargs, ) -> BaseTrace: """Main iteration for singleprocess sampling. Multiple step methods are supported via compound step methods. Parameters ---------- chain : int Number of the chain that the samples will belong to. progressbar : bool Whether or not to display a progress bar in the command line. The bar shows the percentage of completion, the sampling speed in samples per second (SPS), and the estimated remaining time until completion ("expected time of arrival"; ETA). random_seed : single random seed start : dict Starting point in parameter space (or partial point) draws : int The number of samples to draw step : function Step function trace : backend, optional A backend instance or None. If None, the NDArray backend is used. tune : int Number of iterations to tune. model : Model (optional if in ``with`` context) Returns ------- strace : BaseTrace A ``BaseTrace`` object that contains the samples for this chain. """ skip_first = kwargs.get("skip_first", 0) trace = copy(trace) sampling_gen = _iter_sample( draws, step, start, trace, chain, tune, model, random_seed, callback ) _pbar_data = {"chain": chain, "divergences": 0} _desc = "Sampling chain {chain:d}, {divergences:,d} divergences" if progressbar: sampling = progress_bar(sampling_gen, total=draws, display=progressbar) sampling.comment = _desc.format(**_pbar_data) else: sampling = sampling_gen try: strace = None for it, (strace, diverging) in enumerate(sampling): if it >= skip_first and diverging: _pbar_data["divergences"] += 1 if progressbar: sampling.comment = _desc.format(**_pbar_data) except KeyboardInterrupt: pass if strace is None: raise Exception("KeyboardInterrupt happened before the base trace was created.") return strace
[docs]def iter_sample( draws: int, step, start: PointType, trace=None, chain: int = 0, tune: int = 0, model: Optional[Model] = None, random_seed: RandomSeed = None, callback=None, ) -> Iterator[MultiTrace]: """Generate a trace on each iteration using the given step method. Multiple step methods ared supported via compound step methods. Returns the amount of time taken. Parameters ---------- draws : int The number of samples to draw step : function Step function start : dict Starting point in parameter space (or partial point). trace : backend or list This should be a backend instance, or a list of variables to track. If None or a list of variables, the NDArray backend is used. chain : int, optional Chain number used to store sample in backend. tune : int, optional Number of iterations to tune (defaults to 0). model : Model (optional if in ``with`` context) random_seed : single random seed, optional callback : A function which gets called for every sample from the trace of a chain. The function is called with the trace and the current draw and will contain all samples for a single trace. the ``draw.chain`` argument can be used to determine which of the active chains the sample is drawn from. Sampling can be interrupted by throwing a ``KeyboardInterrupt`` in the callback. Yields ------ trace : MultiTrace Contains all samples up to the current iteration Examples -------- :: for trace in iter_sample(500, step): ... """ sampling = _iter_sample(draws, step, start, trace, chain, tune, model, random_seed, callback) for i, (strace, _) in enumerate(sampling): yield MultiTrace([strace[: i + 1]])
def _iter_sample( draws: int, step, start: PointType, trace: Optional[BaseTrace] = None, chain: int = 0, tune: int = 0, model=None, random_seed: RandomSeed = None, callback=None, ) -> Iterator[Tuple[BaseTrace, bool]]: """Generator for sampling one chain. (Used in singleprocess sampling.) Parameters ---------- draws : int The number of samples to draw step : function Step function start : dict Starting point in parameter space (or partial point). Must contain numeric (transformed) initial values for all (transformed) free variables. trace : backend, optional A backend instance or None. If None, the NDArray backend is used. chain : int, optional Chain number used to store sample in backend. tune : int, optional Number of iterations to tune (defaults to 0). model : Model (optional if in ``with`` context) random_seed : single random seed, optional Yields ------ strace : BaseTrace The trace object containing the samples for this chain diverging : bool Indicates if the draw is divergent. Only available with some samplers. """ model = modelcontext(model) draws = int(draws) if draws < 1: raise ValueError("Argument `draws` must be greater than 0.") if random_seed is not None: np.random.seed(random_seed) try: step = CompoundStep(step) except TypeError: pass point = start strace: BaseTrace = _init_trace( expected_length=draws + tune, stats_dtypes=step.stats_dtypes, chain_number=chain, trace=trace, model=model, ) try: step.tune = bool(tune) if hasattr(step, "reset_tuning"): step.reset_tuning() for i in range(draws): stats = None diverging = False if i == 0 and hasattr(step, "iter_count"): step.iter_count = 0 if i == tune: step.stop_tuning() point, stats = step.step(point) strace.record(point, stats) log_warning_stats(stats) diverging = i > tune and stats and stats[0].get("diverging") if callback is not None: callback( trace=strace, draw=Draw(chain, i == draws, i, i < tune, stats, point), ) yield strace, diverging except KeyboardInterrupt: strace.close() raise except BaseException: strace.close() raise else: strace.close() def _mp_sample( draws: int, tune: int, step, chains: int, cores: int, random_seed: Sequence[RandomSeed], start: Sequence[PointType], progressbar: bool = True, trace: Optional[BaseTrace] = None, model=None, callback=None, discard_tuned_samples: bool = True, mp_ctx=None, **kwargs, ) -> MultiTrace: """Main iteration for multiprocess sampling. Parameters ---------- draws : int The number of samples to draw tune : int Number of iterations to tune. step : function Step function chains : int The number of chains to sample. cores : int The number of chains to run in parallel. random_seed : list of random seeds Random seeds for each chain. start : list Starting points for each chain. Dicts must contain numeric (transformed) initial values for all (transformed) free variables. progressbar : bool Whether or not to display a progress bar in the command line. trace : BaseTrace, optional A backend instance, or None. If None, the NDArray backend is used. model : Model (optional if in ``with`` context) callback : Callable A function which gets called for every sample from the trace of a chain. The function is called with the trace and the current draw and will contain all samples for a single trace. the ``draw.chain`` argument can be used to determine which of the active chains the sample is drawn from. Sampling can be interrupted by throwing a ``KeyboardInterrupt`` in the callback. Returns ------- mtrace : pymc.backends.base.MultiTrace A ``MultiTrace`` object that contains the samples for all chains. """ import pymc.sampling.parallel as ps # We did draws += tune in pm.sample draws -= tune traces = [ _init_trace( expected_length=draws + tune, stats_dtypes=step.stats_dtypes, chain_number=chain_number, trace=trace, model=model, ) for chain_number in range(chains) ] sampler = ps.ParallelSampler( draws=draws, tune=tune, chains=chains, cores=cores, seeds=random_seed, start_points=start, step_method=step, progressbar=progressbar, mp_ctx=mp_ctx, ) try: try: with sampler: for draw in sampler: strace = traces[draw.chain] strace.record(draw.point, draw.stats) log_warning_stats(draw.stats) if draw.is_last: strace.close() if callback is not None: callback(trace=trace, draw=draw) except ps.ParallelSamplingError as error: strace = traces[error._chain] for strace in traces: strace.close() multitrace = MultiTrace(traces) multitrace._report._log_summary() raise return MultiTrace(traces) except KeyboardInterrupt: if discard_tuned_samples: traces, length = _choose_chains(traces, tune) else: traces, length = _choose_chains(traces, 0) return MultiTrace(traces)[:length] finally: for strace in traces: strace.close() def _init_jitter( model: Model, initvals: Optional[Union[StartDict, Sequence[Optional[StartDict]]]], seeds: Union[Sequence[int], np.ndarray], jitter: bool, jitter_max_retries: int, ) -> List[PointType]: """Apply a uniform jitter in [-1, 1] to the test value as starting point in each chain. ``model.check_start_vals`` is used to test whether the jittered starting values produce a finite log probability. Invalid values are resampled unless `jitter_max_retries` is achieved, in which case the last sampled values are returned. Parameters ---------- jitter: bool Whether to apply jitter or not. jitter_max_retries : int Maximum number of repeated attempts at initializing values (per chain). Returns ------- start : ``pymc.model.Point`` Starting point for sampler """ ipfns = make_initial_point_fns_per_chain( model=model, overrides=initvals, jitter_rvs=set(model.free_RVs) if jitter else set(), chains=len(seeds), ) if not jitter: return [ipfn(seed) for ipfn, seed in zip(ipfns, seeds)] initial_points = [] for ipfn, seed in zip(ipfns, seeds): rng = np.random.RandomState(seed) for i in range(jitter_max_retries + 1): point = ipfn(seed) if i < jitter_max_retries: try: model.check_start_vals(point) except SamplingError: # Retry with a new seed seed = rng.randint(2**30, dtype=np.int64) else: break initial_points.append(point) return initial_points
[docs]def init_nuts( *, init: str = "auto", chains: int = 1, n_init: int = 500_000, model=None, random_seed: RandomSeed = None, progressbar=True, jitter_max_retries: int = 10, tune: Optional[int] = None, initvals: Optional[Union[StartDict, Sequence[Optional[StartDict]]]] = None, **kwargs, ) -> Tuple[Sequence[PointType], NUTS]: """Set up the mass matrix initialization for NUTS. NUTS convergence and sampling speed is extremely dependent on the choice of mass/scaling matrix. This function implements different methods for choosing or adapting the mass matrix. Parameters ---------- init : str Initialization method to use. * auto: Choose a default initialization method automatically. Currently, this is ``jitter+adapt_diag``, but this can change in the future. If you depend on the exact behaviour, choose an initialization method explicitly. * adapt_diag: Start with a identity mass matrix and then adapt a diagonal based on the variance of the tuning samples. All chains use the test value (usually the prior mean) as starting point. * jitter+adapt_diag: Same as ``adapt_diag``, but use test value plus a uniform jitter in [-1, 1] as starting point in each chain. * jitter+adapt_diag_grad: An experimental initialization method that uses information from gradients and samples during tuning. * advi+adapt_diag: Run ADVI and then adapt the resulting diagonal mass matrix based on the sample variance of the tuning samples. * advi: Run ADVI to estimate posterior mean and diagonal mass matrix. * advi_map: Initialize ADVI with MAP and use MAP as starting point. * map: Use the MAP as starting point. This is discouraged. * adapt_full: Adapt a dense mass matrix using the sample covariances. All chains use the test value (usually the prior mean) as starting point. * jitter+adapt_full: Same as ``adapt_full``, but use test value plus a uniform jitter in [-1, 1] as starting point in each chain. chains : int Number of jobs to start. initvals : optional, dict or list of dicts Dict or list of dicts with initial values to use instead of the defaults from `Model.initial_values`. The keys should be names of transformed random variables. n_init : int Number of iterations of initializer. Only works for 'ADVI' init methods. model : Model (optional if in ``with`` context) random_seed : int, array-like of int, RandomState or Generator, optional Seed for the random number generator. progressbar : bool Whether or not to display a progressbar for advi sampling. jitter_max_retries : int Maximum number of repeated attempts (per chain) at creating an initial matrix with uniform jitter that yields a finite probability. This applies to ``jitter+adapt_diag`` and ``jitter+adapt_full`` init methods. **kwargs : keyword arguments Extra keyword arguments are forwarded to pymc.NUTS. Returns ------- initial_points : list Starting points for each chain. nuts_sampler : ``pymc.step_methods.NUTS`` Instantiated and initialized NUTS sampler object """ model = modelcontext(model) vars = kwargs.get("vars", model.value_vars) if set(vars) != set(model.value_vars): raise ValueError("Must use init_nuts on all variables of a model.") if not all_continuous(vars): raise ValueError("init_nuts can only be used for models with continuous variables.") if not isinstance(init, str): raise TypeError("init must be a string.") init = init.lower() if init == "auto": init = "jitter+adapt_diag" random_seed_list = _get_seeds_per_chain(random_seed, chains) _log.info(f"Initializing NUTS using {init}...") cb = [ pm.callbacks.CheckParametersConvergence(tolerance=1e-2, diff="absolute"), pm.callbacks.CheckParametersConvergence(tolerance=1e-2, diff="relative"), ] initial_points = _init_jitter( model, initvals, seeds=random_seed_list, jitter="jitter" in init, jitter_max_retries=jitter_max_retries, ) apoints = [DictToArrayBijection.map(point) for point in initial_points] apoints_data = [apoint.data for apoint in apoints] potential: quadpotential.QuadPotential if init == "adapt_diag": mean = np.mean(apoints_data, axis=0) var = np.ones_like(mean) n = len(var) potential = quadpotential.QuadPotentialDiagAdapt(n, mean, var, 10) elif init == "jitter+adapt_diag": mean = np.mean(apoints_data, axis=0) var = np.ones_like(mean) n = len(var) potential = quadpotential.QuadPotentialDiagAdapt(n, mean, var, 10) elif init == "jitter+adapt_diag_grad": mean = np.mean(apoints_data, axis=0) var = np.ones_like(mean) n = len(var) if tune is not None and tune > 250: stop_adaptation = tune - 50 else: stop_adaptation = None potential = quadpotential.QuadPotentialDiagAdaptExp( n, mean, alpha=0.02, use_grads=True, stop_adaptation=stop_adaptation, ) elif init == "advi+adapt_diag": approx = pm.fit( random_seed=random_seed_list[0], n=n_init, method="advi", model=model, callbacks=cb, progressbar=progressbar, obj_optimizer=pm.adagrad_window, ) approx_sample = approx.sample( draws=chains, random_seed=random_seed_list[0], return_inferencedata=False ) initial_points = [approx_sample[i] for i in range(chains)] std_apoint = approx.std.eval() cov = std_apoint**2 mean = approx.mean.get_value() weight = 50 n = len(cov) potential = quadpotential.QuadPotentialDiagAdapt(n, mean, cov, weight) elif init == "advi": approx = pm.fit( random_seed=random_seed_list[0], n=n_init, method="advi", model=model, callbacks=cb, progressbar=progressbar, obj_optimizer=pm.adagrad_window, ) approx_sample = approx.sample( draws=chains, random_seed=random_seed_list[0], return_inferencedata=False ) initial_points = [approx_sample[i] for i in range(chains)] cov = approx.std.eval() ** 2 potential = quadpotential.QuadPotentialDiag(cov) elif init == "advi_map": start = pm.find_MAP(include_transformed=True, seed=random_seed_list[0]) approx = pm.MeanField(model=model, start=start) pm.fit( random_seed=random_seed_list[0], n=n_init, method=pm.KLqp(approx), callbacks=cb, progressbar=progressbar, obj_optimizer=pm.adagrad_window, ) approx_sample = approx.sample( draws=chains, random_seed=random_seed_list[0], return_inferencedata=False ) initial_points = [approx_sample[i] for i in range(chains)] cov = approx.std.eval() ** 2 potential = quadpotential.QuadPotentialDiag(cov) elif init == "map": start = pm.find_MAP(include_transformed=True, seed=random_seed_list[0]) cov = pm.find_hessian(point=start) initial_points = [start] * chains potential = quadpotential.QuadPotentialFull(cov) elif init == "adapt_full": mean = np.mean(apoints_data * chains, axis=0) initial_point = initial_points[0] initial_point_model_size = sum(initial_point[n.name].size for n in model.value_vars) cov = np.eye(initial_point_model_size) potential = quadpotential.QuadPotentialFullAdapt(initial_point_model_size, mean, cov, 10) elif init == "jitter+adapt_full": mean = np.mean(apoints_data, axis=0) initial_point = initial_points[0] initial_point_model_size = sum(initial_point[n.name].size for n in model.value_vars) cov = np.eye(initial_point_model_size) potential = quadpotential.QuadPotentialFullAdapt(initial_point_model_size, mean, cov, 10) else: raise ValueError(f"Unknown initializer: {init}.") step = pm.NUTS(potential=potential, model=model, **kwargs) # Filter deterministics from initial_points value_var_names = [var.name for var in model.value_vars] initial_points = [ {k: v for k, v in initial_point.items() if k in value_var_names} for initial_point in initial_points ] return initial_points, step