Implementation:Pyro ppl Pyro UnconditionMessenger

Attribute	Value
File	pyro/poutine/uncondition_messenger.py
Module	pyro.poutine.uncondition_messenger
Lines	35
Parent Class	Messenger
Purpose	Remove observations from sample sites, forcing them to be sampled from their distributions
License	Apache-2.0 (Uber Technologies, Inc.)

Overview

UnconditionMessenger is the inverse of ConditionMessenger. It forces observed sample sites to be treated as unobserved, causing them to be sampled from their distributions rather than using the observed values.

When an observed site is encountered:

1. is_observed is set to False.
2. The original observed value is saved in msg["infer"]["obs"].
3. msg["infer"]["was_observed"] is set to True for downstream use.
4. msg["value"] is set to None so that a fresh sample will be drawn.
5. msg["done"] is set to False to allow sampling.

This handler is useful for posterior predictive sampling, Reweighted Wake Sleep, and Compiled Sequential Importance Sampling, where the model needs to generate samples at sites that are normally observed.

Code Reference

class UnconditionMessenger(Messenger):
    def __init__(self) -> None:
        super().__init__()

    def _pyro_sample(self, msg: "Message") -> None:
        if msg["is_observed"]:
            msg["is_observed"] = False
            msg["infer"]["was_observed"] = True
            msg["infer"]["obs"] = msg["value"]
            msg["value"] = None
            msg["done"] = False

I/O Contract

Parameter	Type	Description
(none)		This handler takes no parameters

Message Effect	Description
msg["is_observed"]	Set to False for previously observed sites
msg["infer"]["was_observed"]	Set to True to record that the site was originally observed
msg["infer"]["obs"]	Stores the original observed value
msg["value"]	Set to None to trigger fresh sampling
msg["done"]	Set to False to allow sampling

Sites that are not observed are unaffected.

Usage Examples

Posterior Predictive Sampling

def model(data):
    loc = pyro.sample("loc", dist.Normal(0, 1))
    pyro.sample("obs", dist.Normal(loc, 1), obs=data)

# Generate predictions by unconditioning the observations
predictive_model = pyro.poutine.uncondition(model)
# Now "obs" will be sampled from Normal(loc, 1) instead of using data

Combining with Condition

# Uncondition overrides condition
conditioned = pyro.poutine.condition(model, data={"obs": observed_data})
unconditioned = pyro.poutine.uncondition(conditioned)
# "obs" is now sampled, not conditioned

Using as Context Manager

with pyro.poutine.uncondition():
    # All observed sites within this block will be sampled instead
    result = model(data)

Related Pages

• Pyro_ppl_Pyro_Messenger_Base -- Parent class providing the handler protocol
• Pyro_ppl_Pyro_Poutine_Handlers -- The poutine.uncondition() factory function
• Pyro_ppl_Pyro_ConditionMessenger -- The inverse operation: converts sample sites into observed sites
• Pyro_ppl_Pyro_ReplayMessenger -- Related: replays recorded values at sites