[discussion] use transforms in potential_fn about numpyro

[neerajprad]

We could handle the default transforms for both SVI and MCMC via an effect handler, the only difference being that for SVI it affects only the param statements whereas for MCMC it would transform sample statements. Another thing to note is that just transforming in the outside loop is not enough, since this also changes the PE computation.

Regarding:

initialize_model takes an additional argument named transforms.

In practice, I haven't seen models where users had to override the default transforms. I can imagine that providing a way to use custom transforms might be useful for certain models. If we provide that, it shouldn't come at the cost of users having to write boiler-plate code for models where they would just use the default transforms anyway. Let me put up something, and we can discuss the interface issues in the PR.

[neerajprad]

Actually, on second thoughts, I think your approach above would be the simplest since HMC in numpyro doesn't have access to the model directly.