I guess I have a rather conceptual question you might be able to help me with. So, as far as I understand it, Edward is essentially leveraging TensorFlow to allow gradient informed inference like HMC. Would it not be much more natural to do that in Julia instead of Python to leverage the autodiff functionality developed there? In the Mamba.jl package, for instance, you can use arbitrary Julia transforms and custom functions to define your model…
What other crucial parts besides automatic differentiation do you require from TF for Edward?