def call(self, inputs):
if not isinstance(inputs, random_variable.RandomVariable):
# Default to a unit normal, i.e., derived from mean squared error loss.
inputs = generated_random_variables.Normal(loc=inputs, scale=1.)
batch_size = tf.shape(inputs)[0] // 2
# TODO(trandustin): Depend on github's ed2 for indexing RVs. This is a hack.
# _, _ = inputs[:batch_size], inputs[batch_size:]
original_inputs = random_variable.RandomVariable(
inputs.distribution[:batch_size], value=inputs.value[:batch_size])
perturbed_inputs = random_variable.RandomVariable(
inputs.distribution[batch_size:], value=inputs.value[batch_size:])
loss = tf.reduce_sum(
tfp.distributions.Normal(self.mean, self.stddev).kl_divergence(
perturbed_inputs.distribution))
loss /= tf.cast(batch_size, dtype=tf.float32)
self.add_loss(loss)
return original_inputs
def TransformedRandomVariable(
rv, # pylint: disable=invalid-name
reversible_layer,
name=None,
sample_shape=(),
value=None):
"""Random variable for f(x), where x ~ p(x) and f is reversible."""
return random_variable.RandomVariable(distribution=TransformedDistribution(
rv.distribution, reversible_layer, name=name),
sample_shape=sample_shape,
value=value)
