Increased support for non-centred parameterisations

[drbenvincent]

There’s been a small amount of discussion about this, but seeking broader input on this idea... The proposal is to increase support for non-centered parameterisations in PyMC (i.e. for hierarchical models). I think there are 3 possible approaches, but feel free to suggest more.

Idea 1 - Create a pm.NormalNonCentered distribution
Something like

def NormalNonCentered(varname: str, mu, sigma, dims: str):
    _mu = pm.Normal(f"z_{varname}", dims=dims)
    return pm.Deterministic(varname, mu + _mu * sigma, dims=dims)

Idea 2 - Add a parameterisation kwarg for pm.Normal

Idea 3 - Add it as a transformation, eg pymc.distributions.transforms.NonCentered

[theorashid]

I'd rather see idea 2 or 3. Idea 2 is probably best from an end user point of view, but probably bloats the codebase. I've used numpyro's non-centred transformation in the past and it's very easy to use (see here, the tranformed variables are in the block above using LocScaleReparam) – I don't know the internals though. Pyro has many more transforms implemented than numpyro

[ferrine]

Can it be on class structure level? E.g with a LocScale base class?

[theorashid]

Do you mean like numpyro's LocScaleReparam(Reparam) class?