Combine models

Contents

Combine models#

A variable belongs to one model at a time. Copy a model or its components when you want an independent version; share variables deliberately when combining two response models.

Copy a submodel#

For model from Build your first model:

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

import liesel.model as lsl

mean_model = model.parental_submodel("mu")

sorted(mean_model.vars)

['beta', 'mu', 'prior_scale', 'x']

The parental submodel contains copies of mu and its ancestors: covariate x, coefficients beta, and their prior scale. It excludes y and the response variance. Changing the copy does not change the original variables.

Add a response#

This complete example creates two groups of observations:

mean_a = lsl.Var.new_param(
    0.0,
    dist=lsl.Dist(tfd.Normal, loc=0.0, scale=2.0),
    name="mean",
)
y_a = lsl.Var.new_obs(
    jnp.array([0.8, 1.2]),
    dist=lsl.Dist(tfd.Normal, loc=mean_a, scale=1.0),
    name="y_a",
)
group_a = lsl.Model(y_a)

mean_b = lsl.Var.new_param(
    0.0,
    dist=lsl.Dist(tfd.Normal, loc=0.0, scale=2.0),
    name="mean",
)
y_b = lsl.Var.new_obs(
    jnp.array([1.4, 1.6]),
    dist=lsl.Dist(tfd.Normal, loc=mean_b, scale=0.5),
    name="y_b",
)
group_b = lsl.Model(y_b)

separate = group_a.copy()
other = group_b.copy()
other.prefix_names("b_")
separate.add(other, copy=True);

separate.plot(width=7, height=4)

Two independent response groups, each with its own mean parameter.

sorted(separate.parameters)

['b_mean', 'mean']

This model has two means, mean and b_mean. Prefixing the second group’s names avoids collisions: add requires unique names. copy=True preserves the source model; the default copy=False moves its contents and empties it.