Model predictive sampling methods don't work with Pandas DataFrames

[brendanhasz]

Because you can't tf.expand_dims on a Pandas DataFrame (or any kind of equivalent operation)... Normally with sampling, will add a singleton dimension as the first dimension to the input data x, and then sample from the variational posteriors (which returns tensors of shape [Nsamples, Shape0, ..., ShapeN]), and that way can broadcast between x and the parameter sample tensors.

Can't cast to tensor and then expand dims, b/c that would break the user's code in pf.Model.__call__, which (might) assume it's a Pandas DataFrame object.

Don't really see a good fix here short of just not supporting DataFrames at all...