Change a model

Use model from Build your first model for these examples. Make a copy when you want to compare model variants.

Change a value

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

import liesel.model as lsl

changed = model.copy()
changed.vars["beta"].value = jnp.array([0.5, 1.5])

np.asarray(changed.vars["mu"].value[:3]).round(3)

array([-1.   , -0.962, -0.924], dtype=float32)

round(float(changed.log_lik), 3)

-98.436

The mean and likelihood update automatically. The connections stay the same. Retrieve variables from changed when editing the copy: a Python variable that points to the original beta still belongs to the original model.

Replace a prior

new_width = lsl.Var.new_value(1.0, name="new_prior_scale")
changed.vars["beta"].dist_node["scale"] = new_width

changed.vars["beta"].dist_node["scale"].name

'new_prior_scale'

The coefficient prior now depends on new_prior_scale. To replace the whole prior, assign a new distribution node:

changed.vars["beta"].dist_node = lsl.Dist(tfd.Normal, loc=0.0, scale=1.5)
changed.rebuild_graph();

sorted(changed.vars)

['beta', 'mu', 'sigma', 'variance', 'x', 'y']

The old prior-scale variables disappear because they are no longer inputs of the model’s seed response y. rebuild_graph() rediscovers the graph from its seeds; components added with add_to_seeds=False can also disappear. update() instead recomputes outdated values without pruning the graph.

The same input indexing works for calculations. Here, sigma takes its variance as its first positional input:

changed.vars["sigma"].value_node[0] = lsl.Var.new_value(0.5, name="fixed_variance")
changed.rebuild_graph();

round(float(changed.vars["sigma"].value), 3)

0.707

The response standard deviation is now about 0.707. This changes one connection. It also removes the unused variance parameter and its prior from this model.

Replace a variable

Start from another copy when comparing this change with the previous one:

fixed = model.copy()
fixed.replace("variance", lsl.Var.new_value(0.5, name="variance"));

sorted(fixed.parameters)

['beta']

Only beta remains a parameter. replace() redirects uses of the replaced variable throughout the model, whereas assigning one value_node or dist_node input changes just that connection. Both choices change the statistical model here: variance is fixed and its prior is removed.