Closed ricardoV94 closed 7 months ago
Attention: Patch coverage is 91.30435%
with 6 lines
in your changes are missing coverage. Please review.
Project coverage is 92.23%. Comparing base (
a2988c7
) to head (45eb6ef
). Report is 1 commits behind head on main.
Just for clarity, we don't have any sections of the code that use a tensor of RNG seeds right?
That's not a thing in PyTensor
Description
This meta-info is necessary to reason about batch dims of SymbolicRandomVariables in the context of https://github.com/pymc-devs/pymc-experimental/pull/300
This is probably what we should use for RandomVariables, instead of defining
ndims_params
andndim_supp
. Those can be properties derived from the gufunc signature.There is however a limitation with gufunc signatures, which has to do with inputs and outputs that are not tensors, such as RNGs and the size vector which obviously cannot have batch dimensions. For now I am treating those as scalars so in the signature they show up as
()
, but perhaps it makes sense to deviate a bit from numpy and use[]
orNone
?For vanilla RandomVariables like Normal. the signature would be
None,None,None,(),()->None,()
, for the inputs: rng, size, dtype, mu, sigma and outputs: next_rng, draws.Related Issue
Checklist
Type of change
📚 Documentation preview 📚: https://pymc--7159.org.readthedocs.build/en/7159/