More aggressive concretization

[phipsgabler]

I propose to make concretize turn every : into the value of axis(x, i).

I think this would make the set of VarNames we are working on a nicely defined poset, since then it's basically "sequences of tuples of sets" with the subsumption relation following inductively from sets of indices (leaving out linear indexing and trailing ones).

That should also make this weird corner case obsolete.

And while we're at it: boolean indexing could be concretized to array indexing:

concretize(@varname(x[[true, false]], x) == @varname(x[[1,2]])

[devmotion]

Would this break variables of dynamically changing size, similar to https://turing.ml/dev/tutorials/06-infinite-mixture-model/ where the size of mu changes in the for loop?

[phipsgabler]

Not more than concretization of begin and end would, I assume? All of these depend essentially only on the length/axes of the value at runtime.

But looking at this from the other side -- have we thoroughly tested the existing lens-based varname implementation for this case...?

[phipsgabler]

I mean, even if you have the weirdest case I can think of,

x[end] ~ D
push!(x, ...)

in a loop, it should be save, since the end is concretized at the point where the assignment happens and thus becomes x[<current index in the loop>].

And : you can only use legally once in a tilda, because otherwise, you'd be sampling the same location twice.

Am I missing something?

[phipsgabler]

@torfjelde what do you think about this?

[torfjelde]

On a first glance this looks like a good idea to me! As you said, this shouldn't cause any issues with examples such as mentioned above since it's all done at runtime anyways.

[yebai]

It should be fixed now.