Source code for pyro.infer.abstract_infer

from __future__ import absolute_import, division, print_function

import numbers
from abc import ABCMeta, abstractmethod
from collections import OrderedDict, defaultdict

import torch
from six import add_metaclass

import pyro.poutine as poutine
from pyro.distributions import Categorical, Empirical
from pyro.ops.stats import waic
from .util import site_is_subsample


[docs]class EmpiricalMarginal(Empirical): """ Marginal distribution over a single site (or multiple, provided they have the same shape) from the ``TracePosterior``'s model. .. note:: If multiple sites are specified, they must have the same tensor shape. Samples from each site will be stacked and stored within a single tensor. See :class:`~pyro.distributions.Empirical`. To hold the marginal distribution of sites having different shapes, use :class:`~pyro.infer.abstract_infer.Marginals` instead. :param TracePosterior trace_posterior: a ``TracePosterior`` instance representing a Monte Carlo posterior. :param list sites: optional list of sites for which we need to generate the marginal distribution. """ def __init__(self, trace_posterior, sites=None, validate_args=None): assert isinstance(trace_posterior, TracePosterior), \ "trace_dist must be trace posterior distribution object" if sites is None: sites = "_RETURN" self._num_chains = 1 self._samples_buffer = defaultdict(list) self._weights_buffer = defaultdict(list) self._populate_traces(trace_posterior, sites) samples, weights = self._get_samples_and_weights() super(EmpiricalMarginal, self).__init__(samples, weights, validate_args=validate_args) def _get_samples_and_weights(self): """ Appends values collected in the samples/weights buffers to their corresponding tensors. """ num_chains = len(self._samples_buffer) samples_by_chain = [] weights_by_chain = [] for i in range(num_chains): samples = torch.stack(self._samples_buffer[i], dim=0) samples_by_chain.append(samples) weights_dtype = samples.dtype if samples.dtype.is_floating_point else torch.float32 weights = torch.as_tensor(self._weights_buffer[i], device=samples.device, dtype=weights_dtype) weights_by_chain.append(weights) if len(samples_by_chain) == 1: return samples_by_chain[0], weights_by_chain[0] else: return torch.stack(samples_by_chain, dim=0), torch.stack(weights_by_chain, dim=0) def _add_sample(self, value, log_weight=None, chain_id=0): """ Adds a new data point to the sample. The values in successive calls to ``add`` must have the same tensor shape and size. Optionally, an importance weight can be specified via ``log_weight`` or ``weight`` (default value of `1` is used if not specified). :param torch.Tensor value: tensor to add to the sample. :param torch.Tensor log_weight: log weight (optional) corresponding to the sample. :param int chain_id: chain id that generated the sample (optional). Note that if this argument is provided, ``chain_id`` must lie in ``[0, num_chains - 1]``, and there must be equal number of samples per chain. """ # Apply default weight of 1.0. if log_weight is None: log_weight = 0.0 if self._validate_args and not isinstance(log_weight, numbers.Number) and log_weight.dim() > 0: raise ValueError("``weight.dim() > 0``, but weight should be a scalar.") # Append to the buffer list self._samples_buffer[chain_id].append(value) self._weights_buffer[chain_id].append(log_weight) self._num_chains = max(self._num_chains, chain_id + 1) def _populate_traces(self, trace_posterior, sites): assert isinstance(sites, (list, str)) for tr, log_weight, chain_id in zip(trace_posterior.exec_traces, trace_posterior.log_weights, trace_posterior.chain_ids): value = tr.nodes[sites]["value"] if isinstance(sites, str) else \ torch.stack([tr.nodes[site]["value"] for site in sites], 0) self._add_sample(value, log_weight=log_weight, chain_id=chain_id)
[docs]class Marginals(object): """ Holds the marginal distribution over one or more sites from the ``TracePosterior``'s model. This is a convenience container class, which can be extended by ``TracePosterior`` subclasses. e.g. for implementing diagnostics. :param TracePosterior trace_posterior: a TracePosterior instance representing a Monte Carlo posterior. :param list sites: optional list of sites for which we need to generate the marginal distribution. """ def __init__(self, trace_posterior, sites=None, validate_args=None): assert isinstance(trace_posterior, TracePosterior), \ "trace_dist must be trace posterior distribution object" if sites is None: sites = ["_RETURN"] elif isinstance(sites, str): sites = [sites] else: assert isinstance(sites, list) self.sites = sites self._marginals = OrderedDict() self._diagnostics = OrderedDict() self._trace_posterior = trace_posterior self._populate_traces(trace_posterior, validate_args) def _populate_traces(self, trace_posterior, validate): self._marginals = {site: EmpiricalMarginal(trace_posterior, site, validate) for site in self.sites}
[docs] def support(self, flatten=False): """ Gets support of this marginal distribution. :param bool flatten: A flag to decide if we want to flatten `batch_shape` when the marginal distribution is collected from the posterior with ``num_chains > 1``. Defaults to False. :returns: a dict with keys are sites' names and values are sites' supports. :rtype: :class:`OrderedDict` """ support = OrderedDict([(site, value.enumerate_support()) for site, value in self._marginals.items()]) if self._trace_posterior.num_chains > 1 and flatten: for site, samples in support.items(): shape = samples.size() flattened_shape = torch.Size((shape[0] * shape[1],)) + shape[2:] support[site] = samples.reshape(flattened_shape) return support
@property def empirical(self): """ A dictionary of sites' names and their corresponding :class:`EmpiricalMarginal` distribution. :type: :class:`OrderedDict` """ return self._marginals
[docs]@add_metaclass(ABCMeta) class TracePosterior(object): """ Abstract TracePosterior object from which posterior inference algorithms inherit. When run, collects a bag of execution traces from the approximate posterior. This is designed to be used by other utility classes like `EmpiricalMarginal`, that need access to the collected execution traces. """ def __init__(self, num_chains=1): self.num_chains = num_chains self._reset() def _reset(self): self.log_weights = [] self.exec_traces = [] self.chain_ids = [] # chain id corresponding to the sample self._idx_by_chain = [[] for _ in range(self.num_chains)] # indexes of samples by chain id self._categorical = None
[docs] def marginal(self, sites=None): """ Generates the marginal distribution of this posterior. :param list sites: optional list of sites for which we need to generate the marginal distribution. :returns: A :class:`Marginals` class instance. :rtype: :class:`Marginals` """ return Marginals(self, sites)
@abstractmethod def _traces(self, *args, **kwargs): """ Abstract method implemented by classes that inherit from `TracePosterior`. :return: Generator over ``(exec_trace, weight)`` or ``(exec_trace, weight, chain_id)``. """ raise NotImplementedError("Inference algorithm must implement ``_traces``.") def __call__(self, *args, **kwargs): # To ensure deterministic sampling in the presence of multiple chains, # we get the index from ``idxs_by_chain`` instead of sampling from # the marginal directly. random_idx = self._categorical.sample().item() chain_idx, sample_idx = random_idx % self.num_chains, random_idx // self.num_chains sample_idx = self._idx_by_chain[chain_idx][sample_idx] trace = self.exec_traces[sample_idx].copy() for name in trace.observation_nodes: trace.remove_node(name) return trace
[docs] def run(self, *args, **kwargs): """ Calls `self._traces` to populate execution traces from a stochastic Pyro model. :param args: optional args taken by `self._traces`. :param kwargs: optional keywords args taken by `self._traces`. """ self._reset() with poutine.block(): for i, vals in enumerate(self._traces(*args, **kwargs)): if len(vals) == 2: chain_id = 0 tr, logit = vals else: tr, logit, chain_id = vals assert chain_id < self.num_chains self.exec_traces.append(tr) self.log_weights.append(logit) self.chain_ids.append(chain_id) self._idx_by_chain[chain_id].append(i) self._categorical = Categorical(logits=torch.tensor(self.log_weights)) return self
[docs] def information_criterion(self, pointwise=False): """ Computes information criterion of the model. Currently, returns only "Widely Applicable/Watanabe-Akaike Information Criterion" (WAIC) and the corresponding effective number of parameters. Reference: [1] `Practical Bayesian model evaluation using leave-one-out cross-validation and WAIC`, Aki Vehtari, Andrew Gelman, and Jonah Gabry :param bool pointwise: a flag to decide if we want to get a vectorized WAIC or not. When ``pointwise=False``, returns the sum. :returns: a dictionary containing values of WAIC and its effective number of parameters. :rtype: :class:`OrderedDict` """ if not self.exec_traces: return {} obs_node = None log_likelihoods = [] for trace in self.exec_traces: obs_nodes = trace.observation_nodes if len(obs_nodes) > 1: raise ValueError("Infomation criterion calculation only works for models " "with one observation node.") if obs_node is None: obs_node = obs_nodes[0] elif obs_node != obs_nodes[0]: raise ValueError("Observation node has been changed, expected {} but got {}" .format(obs_node, obs_nodes[0])) log_likelihoods.append(trace.nodes[obs_node]["fn"] .log_prob(trace.nodes[obs_node]["value"])) ll = torch.stack(log_likelihoods, dim=0) waic_value, p_waic = waic(ll, torch.tensor(self.log_weights, device=ll.device), pointwise) return OrderedDict([("waic", waic_value), ("p_waic", p_waic)])
[docs]class TracePredictive(TracePosterior): """ Generates and holds traces from the posterior predictive distribution, given model execution traces from the approximate posterior. This is achieved by constraining latent sites to randomly sampled parameter values from the model execution traces and running the model forward to generate traces with new response ("_RETURN") sites. :param model: arbitrary Python callable containing Pyro primitives. :param TracePosterior posterior: trace posterior instance holding samples from the model's approximate posterior. :param int num_samples: number of samples to generate. """ def __init__(self, model, posterior, num_samples): self.model = model self.posterior = posterior self.num_samples = num_samples super(TracePredictive, self).__init__() def _traces(self, *args, **kwargs): if not self.posterior.exec_traces: self.posterior.run(*args, **kwargs) data_trace = poutine.trace(self.model).get_trace(*args, **kwargs) for _ in range(self.num_samples): model_trace = self.posterior().copy() self._adjust_to_data(model_trace, data_trace) resampled_trace = poutine.trace(poutine.replay(self.model, model_trace)).get_trace(*args, **kwargs) yield (resampled_trace, 0., 0) def _adjust_to_data(self, trace, data_trace): for name, site in list(trace.nodes.items()): # Adjust subsample sites if site_is_subsample(site): site["fn"] = data_trace.nodes[name]["fn"] site["value"] = data_trace.nodes[name]["value"] # Adjust sites under conditionally independent stacks try: site["cond_indep_stack"] = data_trace.nodes[name]["cond_indep_stack"] site["fn"] = data_trace.nodes[name]["fn"] for cis in site["cond_indep_stack"]: # Select random sub-indices to replay values under conditionally independent stacks. # Otherwise, we assume there is an dependence of indexes between training data # and prediction data. logits = torch.ones(site["value"].size(cis.dim), device=site["value"].device) subidxs = Categorical(logits=logits).sample([cis.size]) site["value"] = site["value"].index_select(cis.dim, subidxs) except KeyError: pass
[docs] def marginal(self, sites=None): """ Gets marginal distribution for this predictive posterior distribution. """ return Marginals(self, sites)