Predict quantities

predict() evaluates named quantities at supplied parameter draws. It propagates their uncertainty through the calculations in the graph. For new random responses, use Simulate responses.

Choose outputs

These examples use model and posterior samples from Fit and predict. Both the mean mu and standard deviation sigma depend on x in that model.

import jax
import jax.numpy as jnp
import tensorflow_probability.substrates.jax.distributions as tfd

import liesel.goose as gs
import liesel.model as lsl

x_grid = jnp.linspace(-1.0, 1.0, 60)
predicted = model.predict(
    samples,
    predict=["mu", "sigma"],
    newdata={"x": x_grid},
)

{name: value.shape for name, value in predicted.items()}

{'mu': (4, 500, 60), 'sigma': (4, 500, 60)}

Each array has shape (4, 500, 60) for the tutorial’s four chains and 500 retained draws. Its last axis indexes grid points. The predict entries are variable or node names, not distribution argument names such as loc. Values saved for weak variables in samples are ignored and recalculated.

Supply new data

newdata replaces named inputs for the evaluation. Omit it to use the current data; omitted inputs keep their current values. Supply compatible arrays for all covariates that change together. If the mean and scale use separate design inputs X and Z, provide both with aligned rows; a one-dimensional slice must state which other covariate is fixed. See Use different covariates in the worked tutorial. Keys must not overlap between samples and newdata. Include draws for every parameter whose posterior uncertainty you intend to propagate: missing strong inputs keep their current values.

Selecting only mu and sigma allows prediction to use their parental submodel, so the old response length does not matter. Predicting response log densities also requires response values matching the new covariates.

With default outputs or model-total nodes, the current implementation uses the full model and applies the new-data state to it. Use model.copy().predict(...) when you need to preserve its state on those paths. Explicit output selections as above operate on a copied parental submodel.

Summarize predictions

summary = gs.SamplesSummary(
    predicted,
    quantiles=(0.05, 0.95),
    which=["quantiles"],
)

mean_interval = summary.quantities["quantile"]["mu"]

mean_interval.shape

(2, 60)

The shape is (2, 60): lower and upper limits for each grid point. These are pointwise 90% credible intervals for the conditional mean. They describe neither new response variation nor simultaneous coverage of the entire curve. Summarize predicted["sigma"] similarly to inspect uncertainty in the spread.

See Use distributions for pointwise log likelihoods. For large prediction jobs, the chunk_size option bounds parallel intermediate calculations; it does not reduce the memory needed for the returned arrays.