Cannot pass a random variable to TransformedDistribution

[AdrienCorenflos]

Hi,

I am using edward2==0.0.1 in combination with tensorflow_probability==0.9.0 and tensorflow==2.1.0, all installed from pip.

I am facing an issue with ed.TransformedDistribution that feel like a regression to me.

import edward2 as ed
import tensorflow_probability as tfp
tfb = tfp.bijectors

normal = ed.Normal(0., 1.)
shifted_normal = ed.TransformedDistribution(normal, tfb.Shift(0.5))
# This raises with AttributeError: 'RandomVariable' object has no attribute 'name'

I can patch the case 1 by passing normal.distribution to ed.TransformedDistribution but I don't think this is desirable.

[dustinvtran]

Thanks for raising this! In the code, we minimally wrap all distributions as ed2 rvs, and the input types are the same as distributions (this is why ed.TransformedDistribution still takes a distribution as input).

One way to resolve this is to augment all higher-order distributions that take a distribution as input to take a RandomVariable as input (that is, raising all tfp.Distribution types to ed.RandomVariable types). It's easy to do this on a one-off basis, but we haven't quite figured out how to automate this which would be better for maintenance.

[AdrienCorenflos]

Ah yes that sounds tricky if you don't want to break the open close principles.

Also even doing it on a case by case basis may prove complicated.
For example, someone coded a distribution by subclassing the Transformed Distribution instead of composing it (I might send a PR for that one...), so you wouldn't be able to check the bases of the classes not their attributes either...