Posterior predictive of deterministics of observedRVs are wrong

[lucianopaz]

Following a discourse thread, sample_posterior_predictive does not give consistent results for deterministic nodes that depend on ObservedRV's.

Please provide a minimal, self-contained, and reproducible example.

import numpy as np
from theano import tensor as tt
import pymc3 as pm


def Model_1(meas_in_1, meas_in_2, meas_out):
    with pm.Model() as Assy_model_1:
        mu_in_1 = pm.Normal('mu_in_1', -5., 5.)
        sigma_in_1 = pm.HalfCauchy('sd_in_1', 5.)
        mu_in_2 = pm.Normal('mu_in_2', -5., 5.)
        sigma_in_2 = pm.HalfCauchy('sd__in_2', 5.)

        in_1 = pm.Normal('in_1', mu_in_1, sigma_in_1, observed=meas_in_1)
        in_2 = pm.Normal('in_2', mu_in_2, sigma_in_2, observed=meas_in_2)
        out_diff = in_1 + in_2
        pm.Deterministic('out', out_diff)
        return(Assy_model_1)


np.random.seed(23)
meas_in_1 = 2 + np.random.uniform(0., 1., size=100)
meas_in_2 = 5 + np.random.uniform(0., 1., size=100)
meas_out = meas_in_1 + meas_in_2

Model = Model_1(meas_in_1, meas_in_2, meas_out)

with Model:
    trace = pm.sample()

df_trace = pm.trace_to_dataframe(trace)
print(df_trace.head())

ppc_trace = pm.trace_to_dataframe(trace,
                                  varnames=[n for n in trace.varnames
                                            if n != 'out']).to_dict('records')

ppc = pm.sample_posterior_predictive(model=Model,
                                     trace=ppc_trace,
                                     samples=len(ppc_trace),
                                     vars=(Model.deterministics +
                                           Model.basic_RVs))
print(ppc.keys())
print("ppc['in_1'] + ppc['in_2']")
print(ppc['in_1'] + ppc['in_2'])
print("ppc['out']")
print(ppc['out'])
print("meas_in_1 + meas_in_2")
print(meas_in_1 + meas_in_2)

Please provide any additional information below.

The error is that ppc['out'] != ppc['in_1'] + ppc['in_2'].

The problem seems to come from draw_values. The in_1 and in_2 observed variables are theano.ViewOPs with TensorConstant ancestors:

In [1]: tt.printing.debugprint(Model['out'])
 Elemwise{identity}` [id A] 'out'   
 |Elemwise{add,no_inplace} [id B] ''   
   |ViewOp [id C] 'in_1'   
   | |TensorConstant{[2.5172978...43301769]} [id D]
   |ViewOp [id E] 'in_2'   
     |TensorConstant{[5.6434602...27628245]} [id F]

In draw_values' loop over the named_node relations (while stack:...), TensorConstant variables are ignored and they don't add their parents to the stack. Later on, in the while to_eval or missing_inputs:... loop, the to_eval parameter nodes may be in the wrong order, and out is compiled, and used with the view of the original TensorConstant.value, instead of with the values that get drawn from the posterior predictive distribution.

Versions and main components

• PyMC3 Version: pymc3.6
• Theano Version: 1.0.4
• Python Version: 3.6
• Operating system: Ubuntu
• How did you install PyMC3: (conda/pip) git clone