Does model.logp(..., jacobian=False) do anything?

[avm19]

So I tried

l1 = model.logp(jacobian=True)
l2 = model.logp(jacobian=False)

and l1 and l2 are identical graphs except for the name of the head node. I looked into the source code but haven't found how jacobian parameter is used, except for controlling the string name and being passed around into sub-calls.

Is this some sort of legacy from (val, jac) = model.logp(...) signature? Or am I missing something?

Also, the docstring for model.logp is mixed up.

[ricardoV94]

Yes it does something, perhaps you have a model without transforms (like just normal variates). Try with

with pm.Model() as m:
  x = pm.Exponential("x", 1)

Which has a log transform on x and a jacobian correction

[avm19]

Oh, I see. For some reason, I thought Jacobian referred to something like dlogp, not to the determinant in the change-of-variables formula.

[ricardoV94]

Do you want to propose a small modification to the docstrings that would have avoided your confusion in the first place?

[avm19]

Sure. Do I need to create an issue or is a dangling PR okay?