Trouble understanding how to use sample_posterior_predictive to generate a prediction

2

The syntax from the example you linked is somewhat out of date. coords are now deeply integrated into PyMC, so you don’t have to do all this juggling with az.from_pymc. Here’s a simplified model:

coords = {'features': ['treatment', 'cov1', 'cov2']}
coords_mutable = {'obs_id': np.arange(len(target))}

with pm.Model(coords=coords, coords_mutable=coords_mutable) as sk_lm:
    feature_data = pm.MutableData('feature_data', features, dims=('obs_id', 'features'))

    alpha = pm.Normal('alpha', mu=0, sigma=1)
    betas = pm.Normal('betas', mu=[40, 90, 50], sigma=5, dims='features')
    sigma = pm.Exponential("sigma", lam=1)

    mu = pm.Deterministic("mu", alpha + feature_data @ betas, dims='obs_id')

    y = pm.Normal("y", mu=mu, sigma=sigma, observed=target, dims='obs_id')

    idata = pm.sample(init='jitter+adapt_diag')

Since you’re planning to do out-of-sample stuff, you want to set the index coordinate with coords_mutable. That lets you change it down the road. Down the road is here:

new_data = np.random.randn(3, 3)

with sk_lm:
    pm.set_data({"feature_data": new_data}, 
                coords={'obs_id':np.arange(3) + len(target)})
    idata = pm.sample_posterior_predictive(idata, predictions=True, extend_inferencedata=True)

Since I used coords_mutable, I can pass new coords to pm.set_data. This is a common gotcha with out-of-sample prediction – you have to make sure the shapes of the targets somehow get updated to match the new data, even though you’re not using targets. This is discussed a bit more in-depth here.

1 Like