No trainable variables for real_nvp_default_template

Hi,

Configuration:
TF version '2.4.0'
TFP version '0.12.2'

I am trying to use real_nvp template in RealNVP bijector as given in the example in the docs
https://www.tensorflow.org/probability/api_docs/python/tfp/bijectors/RealNVP

sample() and log_prob() functions work but I do not see any trainable variables which are declared in the template.

tfd = tfp.distributions
tfb = tfp.bijectors

# A common choice for a normalizing flow is to use a Gaussian for the base
# distribution. (However, any continuous distribution would work.) E.g.,
nvp = tfd.TransformedDistribution(
    distribution=tfd.MultivariateNormalDiag(loc=[0., 0., 0.]),
    bijector=tfb.RealNVP(
        num_masked=2,
        shift_and_log_scale_fn=tfb.real_nvp_default_template(
            hidden_layers=[512, 512])))

x = nvp.sample()
nvp.log_prob(x)
nvp.log_prob([0.0, 0.0, 0.0])

print(nvp.trainable_variables)

>>>()

In fact the bijector itself does not seem to inherit any trainable variables

template = tfb.real_nvp_default_template(hidden_layers=[512, 512])
bijector = tfb.RealNVP(num_masked=1, shift_and_log_scale_fn=template)
bijector.forward([0.5, 1., 2.])
bijector.trainable_variables

>>> ()

Is this supposed to be so?

Thanks

Start by running the provided RealNVP and real_nvp_default_template snippets with TensorFlow 2.4.0 and TensorFlow Probability 0.12.2, then inspect how trainable_variables are exposed by the bijector and template. Compare the observed behavior with the RealNVP API example and determine whether the missing variables indicate a bug or expected behavior.