Seth Axen

> BTW, the reason why I was even using the parallel (single) `pathfinder` thing was because I wanted to initialize several chains for the same posterior in parallel, but I...

> Maybe this is actually something that Pathfinder should (optionally) check? That the log density (gradient) can be evaluated for the returned draws? Oh that's interesting, can you open an...

RE inconsistency in shape conventions between docstrings and what the code expects, I wonder if it would be useful to use [jaxtyping](https://github.com/patrick-kidger/jaxtyping) to provide shapes in type hints. These then...

If there's interest in the jaxtyping approach, I'd be happy to hack together a quick proof-of-concept.

I made a quick attempt to prototype jaxtyping+beartype in SBI, and I don't think it's a good fit for the following reasons: - jaxtyping currently doesn't support multiple variadic shapes...

I've intentionally left it so that `updf(d, x)` is not necessarily `exp(logupdf(d, x))`, because I suspect it's more likely that the exponential of `logupdf` will under-/overflow than for `logpdf`, so...

I'd prefer to get a :+1: feedback on the general approach from the maintainers before messing around with all of the testing functions. Maybe @devmotion ?

Would it make sense to have the following `dims` fallback? ```julia dims(x::AbstractArray) = dims(axes(x)) ``` With this, the following works: ```julia julia> A = UpperTriangular(DimMatrix(randn(5, 5), (X, Y))); julia> dims(A)...

Seems related: ```julia julia> using DimensionalData, Tables julia> ds = DimStack((; x=DimArray(ones(5, 3), (X, Y)), y=DimArray(zeros(5), X))); julia> ds[X(3)] |> Tables.columntable ``` On main, the result is ```julia (Y =...

One coincidental behavior change is that before this PR `dim(fill(1)) === nothing`. After this PR `dim(fill(1)) == ()`. Not certain if this has any bad side effects.