Resampling unobserved variables in sample_posterior_predictive

[zaxtax]

As part of using pm.as_model, I often end up writing code that looks a like this like:

import pymc as pm
import pymc_experimental as pmx
import numpy as np

n_obs = 100
true_beta = 2.5
true_alpha = 0.25

x = np.random.normal(size=n_obs)
true_p = 1 / (1 + np.exp(-(true_alpha + true_beta * x)))
y = np.random.binomial(n=1, p=true_p)

@pmx.as_model()
def logistic_model(x, y):
    alpha = pm.Normal("alpha")
    beta = pm.Normal("beta")
    p = pm.Deterministic("p", pm.math.sigmoid(alpha + beta * x))
    obs = pm.Bernoulli("obs", p=p, observed=y, shape=x.shape[0])

lm = logistic_model(x, y)
idata = pm.sample(model=lm)

grid_size = 250
x_oos = np.linspace(x.min(), x.max(), grid_size)
lm_oos = logistic_model(x_oos, None)
oos_idata = pm.sample_posterior_predictive(
    idata,
    model=lm_oos,
    var_names=["p", "obs"],
    predictions=True, 
)

This is fairly clean. I don't have to put a dummy value for y when doing oos. But, I have to pass both p and obs to var_names. Setting aside the need to pass p to force resampling of the Deterministic. I don't know why we need to specify obs. While it's not in lm_oos it is in idata and it would be nice to resample it!

What do others thinks?

[twiecki]

I agree, that would be nice.

For now, you could just not specify observerd and only add it with pm.observe() at inference time and otherwise have the first model just be the generative one.

@pmx.as_model()
def logistic_model(x):
    alpha = pm.Normal("alpha")
    beta = pm.Normal("beta")
    p = pm.Deterministic("p", pm.math.sigmoid(alpha + beta * x))
    obs = pm.Bernoulli("obs", p=p) # I don't think you even need shape here

lm = logistic_model(x)
lm_obs = pm.observe(lm, {"obs": y})
idata = pm.sample(model=lm_obs)

[zaxtax]

But will I need to add it while generating predictions?

…

On Fri, 29 Mar 2024, 07:26 Thomas Wiecki, ***@***.***> wrote: I agree, that would be nice. For now, you could just not specify observerd and only add it with pm.observed() at inference time and otherwise have the first model just be the generative one. — Reply to this email directly, view it on GitHub <#7225 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAACCUP5CBJERTBFNBGKI7DY2UCSNAVCNFSM6AAAAABFLKHJ3KVHI2DSMVQWIX3LMV43SRDJONRXK43TNFXW4Q3PNVWWK3TUHM4DSNBYGQ4TM> . You are receiving this because you authored the thread.Message ID: ***@***.***>

[twiecki]

Oh, I think I misunderstood. Are you saying you want sample_posterior_predictive to look for what variable is observed from idata, rather than having it specified in var_names?

[zaxtax]

Yes! Exactly!

[zaxtax]

I should clarify that I want to keep the var_names override, but if the kwarg isn't given it would be nice if the idata was inspected to determine what was observed.

[twiecki]

CC @ricardoV94

[ricardoV94]

I'm not against it