[FR] Predictive with deterministic site in the guide · pyro-ppl/pyro#3358

(6 评论) (0 反应) (0 负责人)Python (8,211 star) (981 fork)batch import

enhancementhelp wanted

描述

Hi,

I'm working on a project where we would like to access the output of an NN in the guide when using Predictive. We've implemented it using a deterministic site in the guide. The program boils down to the following.

import pyro
from pyro.infer import Predictive
from pyro.distributions import Normal
import torch

def model():
    pyro.deterministic('m_deter', torch.tensor(1.))
    pyro.sample('x', Normal(torch.zeros(()), torch.ones(())))

def guide():
    pyro.deterministic('g_deter', torch.tensor(1.))
    pyro.sample('x', Normal(torch.zeros(()), torch.ones(())))


Predictive(
  model=model, 
  guide=guide, 
  return_sites=('model_site', 'guide_site', 'x'), 
  num_samples=1)() # Includes m_deter but not g_deter

We would like for both m_deter and g_deter to be included. It looks like Predictive currently only considers model sites for return sites. Would it be possible to expand it so we can include deterministic sites from the guide?

贡献者指南

技术栈: pythonpytorch
领域: machine learning
议题类型: feature
难度: 3
预计时间: 1-3 hours
活动状态: fresh
清晰度: clear
前置要求: Basic Pyro usage
新手友好度: 70
研究方向: The issue is in pyro/infer/predictive.py line 108 where only model sites are considered for return sites. To include deterministic sites from the guide, extend the logic to capture sites from the guide's trace. Refer to how model's deterministic sites are collected via trace and replicate for guide. Test with the provided minimal example to ensure both m deter and g deter appear in output.