Will Dean
Will Dean
Was this good to go?
Yes, I have noted this and we have it documented in the out of sample section. But I understand it being confusing. What behavior would you expect?
What was the conclusion with this? Did this get settled?
Yes, totally. Makes sense. Let me change up the scenarios then in the initial description. The "consumed" one fails because of internal checks of parent dim names. There'd have to...
Some scenarios that should work: ```python # Broadcasting support p = Prior("Dirichlet", a=np.ones((2, 5)), dims=("geo", "prob")) y = Prior("Categorical", p=p, dims=("geo", "trial")) ``` The information from `rv_op` could potentially warn...
> Is that example with the dims "geo", "trial" in that order in purpose? You would need to transpose them for it to work since it must broadcast to the...
Some things to note: - Current implementation does support consumed dims to some degree they aren't specified ```python p = Prior("Dirichlet", a=[1, 2, 3]) data = Prior("Categorical", p=p, dims="trial") samples...
This issue has been transferred to pymc-devs/pymc-extras
I believe #381 would be easy with this abstraction
One of the difficult aspect here is that the serialization will be different for object. So there will need to be a way to distinguish between objects when they come...