Pavel Sountsov

While this particular issue is fixable in principle (see below), the core issue is that this Inference Gym targets are not intended to be used this way. The targets are...

Probably a combination of being an ancient bijector and this line: https://github.com/tensorflow/probability/blob/65f265c62bb1e2d15ef3e25104afb245a6d52429/tensorflow_probability/python/bijectors/power_transform.py#L96 I guess from a performance perspective you'd want to still maintain the power==0 static branch.