Source code for pyro.poutine.trace_messenger

# Copyright (c) 2017-2019 Uber Technologies, Inc.
# SPDX-License-Identifier: Apache-2.0

import sys
from typing import TYPE_CHECKING, Callable, Generic, Literal, Optional, TypeVar

from typing_extensions import ParamSpec, Self

from pyro.poutine.messenger import Messenger
from pyro.poutine.trace_struct import Trace
from pyro.poutine.util import site_is_subsample

    from pyro.poutine.runtime import Message

_P = ParamSpec("_P")
_T = TypeVar("_T")

[docs]def identify_dense_edges(trace: Trace) -> None: """ Modifies a trace in-place by adding all edges based on the `cond_indep_stack` information stored at each site. """ for name, node in trace.nodes.items(): if site_is_subsample(node): continue if node["type"] == "sample": for past_name, past_node in trace.nodes.items(): if site_is_subsample(past_node): continue if past_node["type"] == "sample": if past_name == name: break past_node_independent = False for query, target in zip( node["cond_indep_stack"], past_node["cond_indep_stack"] ): if ( == and query.counter != target.counter ): past_node_independent = True break if not past_node_independent: trace.add_edge(past_name, name)
[docs]class TraceMessenger(Messenger): """ Return a handler that records the inputs and outputs of primitive calls and their dependencies. Consider the following Pyro program: >>> def model(x): ... s = pyro.param("s", torch.tensor(0.5)) ... z = pyro.sample("z", dist.Normal(x, s)) ... return z ** 2 We can record its execution using ``trace`` and use the resulting data structure to compute the log-joint probability of all of the sample sites in the execution or extract all parameters. >>> trace = pyro.poutine.trace(model).get_trace(0.0) >>> logp = trace.log_prob_sum() >>> params = [trace.nodes[name]["value"].unconstrained() for name in trace.param_nodes] :param fn: a stochastic function (callable containing Pyro primitive calls) :param graph_type: string that specifies the kind of graph to construct :param param_only: if true, only records params and not samples :returns: stochastic function decorated with a :class:`~pyro.poutine.trace_messenger.TraceMessenger` """ def __init__( self, graph_type: Optional[Literal["flat", "dense"]] = None, param_only: Optional[bool] = None, ) -> None: """ :param string graph_type: string that specifies the type of graph to construct (currently only "flat" or "dense" supported) :param param_only: boolean that specifies whether to record sample sites """ super().__init__() if graph_type is None: graph_type = "flat" if param_only is None: param_only = False assert graph_type in ("flat", "dense") self.graph_type = graph_type self.param_only = param_only self.trace = Trace(graph_type=self.graph_type) def __enter__(self) -> Self: self.trace = Trace(graph_type=self.graph_type) return super().__enter__() def __exit__(self, *args, **kwargs) -> None: """ Adds appropriate edges based on cond_indep_stack information upon exiting the context. """ if self.param_only: for node in list(self.trace.nodes.values()): if node["type"] != "param": assert node["name"] is not None self.trace.remove_node(node["name"]) if self.graph_type == "dense": identify_dense_edges(self.trace) return super().__exit__(*args, **kwargs) def __call__(self, fn: Callable[_P, _T]) -> Callable[_P, _T]: """ TODO docs """ return TraceHandler(self, fn)
[docs] def get_trace(self) -> Trace: """ :returns: data structure :rtype: pyro.poutine.Trace Helper method for a very common use case. Returns a shallow copy of ``self.trace``. """ return self.trace.copy()
def _reset(self) -> None: tr = Trace(graph_type=self.graph_type) if "_INPUT" in self.trace.nodes: tr.add_node( "_INPUT", name="_INPUT", type="input", args=self.trace.nodes["_INPUT"]["args"], kwargs=self.trace.nodes["_INPUT"]["kwargs"], ) self.trace = tr super()._reset() def _pyro_post_sample(self, msg: "Message") -> None: if self.param_only: return assert msg["name"] is not None assert msg["infer"] is not None if msg["infer"].get("_do_not_trace"): assert msg["infer"].get("is_auxiliary") assert not msg["is_observed"] return self.trace.add_node(msg["name"], **msg.copy()) def _pyro_post_param(self, msg: "Message") -> None: assert msg["name"] is not None self.trace.add_node(msg["name"], **msg.copy())
[docs]class TraceHandler(Generic[_P, _T]): """ Execution trace poutine. A TraceHandler records the input and output to every Pyro primitive and stores them as a site in a Trace(). This should, in theory, be sufficient information for every inference algorithm (along with the implicit computational graph in the Variables?) We can also use this for visualization. """ def __init__(self, msngr: TraceMessenger, fn: Callable[_P, _T]) -> None: self.fn = fn self.msngr = msngr def __call__(self, *args: _P.args, **kwargs: _P.kwargs) -> _T: """ Runs the stochastic function stored in this poutine, with additional side effects. Resets self.trace to an empty trace, installs itself on the global execution stack, runs self.fn with the given arguments, uninstalls itself from the global execution stack, stores the arguments and return value of the function in special sites, and returns self.fn's return value """ with self.msngr: self.msngr.trace.add_node( "_INPUT", name="_INPUT", type="args", args=args, kwargs=kwargs ) try: ret = self.fn(*args, **kwargs) except (ValueError, RuntimeError) as e: exc_type, exc_value, traceback = sys.exc_info() shapes = self.msngr.trace.format_shapes() assert exc_type is not None exc = exc_type("{}\n{}".format(exc_value, shapes)) exc = exc.with_traceback(traceback) raise exc from e self.msngr.trace.add_node( "_RETURN", name="_RETURN", type="return", value=ret ) return ret @property def trace(self) -> Trace: return self.msngr.trace
[docs] def get_trace(self, *args, **kwargs) -> Trace: """ :returns: data structure :rtype: pyro.poutine.Trace Helper method for a very common use case. Calls this poutine and returns its trace instead of the function's return value. """ self(*args, **kwargs) return self.msngr.get_trace()