Do not unflatten unevaluated lazy properties.

[tillahoffmann]

The tree_flatten function uses self.__dict__.get(name) to obtain the data field with the given name. Unevaluated lazy properties do not appear in self.__dict__, and they are set to None when the representation is unflattened. Here is an example.

>>> import jax
>>> from jax import numpy as jnp
>>> import numpyro
>>> 
>>> 
>>> @jax.jit
>>> def f1(x):
...     return numpyro.distributions.MultivariateNormal(jnp.zeros(3), jnp.eye(3))
>>> 
>>> 
>>> @jax.jit
>>> def f2(x):
...     dist = numpyro.distributions.MultivariateNormal(jnp.zeros(3), jnp.eye(3))
...     dist.precision_matrix
...     return dist

>>> print(f1(0).precision_matrix)
None
>>> print(f2(0).precision_matrix)
[[1. 0. 0.]
 [0. 1. 0.]
 [0. 0. 1.]]

The changes in this PR only set the attribute on the reconstructed instance if the value is not None or if the attribute is not a lazy property. While there is ambiguity between None representing an unevaluated property and None being the value of a lazy property, the implementation remains correct: If the evaluated value is None it is not cached and re-evaluates to None the first time the attribute is accessed on the reconstructed instance.

I've also added a test.

[fehiepsi]

Hi @tillahoffmann, could you point me to an example where the value is None but it is not a lazy property?

[tillahoffmann]

I'm not actually aware of an example where the value is None. We could also just not set the value if it's None without the extra lazy_property check.

I added it to prevent surprises. E.g., if a user implemented a custom distribution where one of the fields has a None value and we didn't check the lazy_property, they would get an AttributeError if they accessed the attribute after a flatten/unflatten cycle. It's probably an unlikely scenario, however.

[fehiepsi]

Oh, I think I understand your implementation now. That makes sense to me. Thanks!!