TF 2.0 is out! Get hands-on practice at TF World, Oct 28-31. Use code TF20 for 20% off select passes. Register now


View source on GitHub

Class ReparameterizationType

Instances of this class represent how sampling is reparameterized.


  • Class tf.compat.v1.distributions.ReparameterizationType
  • Class tf.contrib.distributions.ReparameterizationType

Two static instances exist in the distributions library, signifying one of two possible properties for samples from a distribution:

FULLY_REPARAMETERIZED: Samples from the distribution are fully reparameterized, and straight-through gradients are supported.

NOT_REPARAMETERIZED: Samples from the distribution are not fully reparameterized, and straight-through gradients are either partially unsupported or are not supported at all. In this case, for purposes of e.g. RL or variational inference, it is generally safest to wrap the sample results in a stop_gradients call and use policy gradients / surrogate loss instead.


View source





View source


Determine if this ReparameterizationType is equal to another.

Since RepaparameterizationType instances are constant static global instances, equality checks if two instances' id() values are equal.


  • other: Object to compare against.


self is other.