AEInfer happens regardless of the inference program, except when it doesn't

[lenaqr]

As far as I can tell, AEinfer is called here: https://github.com/probcomp/Venturecxx/blob/master/backend/lite/trace.py#L543

In particular:

• It happens once per transition of any primitive inference operator, including weird ones like draw_scaffold and print_scaffold_stats.
• It ignores the scope and block, except if there are no random choices in the block (latents don't count as random choices), in which case it doesn't run it at all.

[axch]

Yeah, it's weird. AE Kernels in Venture are weird, partly because they happened before we understood that we wanted inference programming.

I expect the scope and block to affect AEInfer through the registerAEKernels mechanism, which is, if I recall correctly, supposed to select only the relevant AE Kernels to run. This may also do something reasonable with e.g. draw_scaffold (or may not).

[lenaqr]

I don't think so. From reading the code, it looks like registerAEKernels just maintains a global set of nodes that have AEKernels, and any primitive infer hits all of them regardless of whatever else it does.

Some testing seems to confirm this:

venture[script] > assume x = tag("scope", "block", flip())
False
venture[script] > assume h = make_lazy_hmm(simplex(0.5, 0.5), matrix(array(array(0.5, 0.5), array(0.5, 0.5))), matrix(array(array(0.5, 0.5), array(0.5, 0.5))))
([], {})
venture[script] > observe h(3) = atom<0>
venture[script] > sample h
([matrix([[1, 0]]), matrix([[0, 1]]), matrix([[1, 0]]), matrix([[0, 1]])], {3: [0]})
venture[script] > print_scaffold_stats("scope", "block", 1)
---Scaffold---
# pnodes: 1
# absorbing nodes: 1
# aaa nodes: 0
# brush nodes: 0
border lengths: [2]
# lkernels: 0
[]
venture[script] > sample h
([matrix([[0, 1]]), matrix([[1, 0]]), matrix([[0, 1]]), matrix([[1, 0]])], {3: [0]})

What you described sounds like a reasonable thing that it should do. Maybe there should also be a dedicated inference action for running AEKernels, so that they can be controlled by inference programs.

[axch]

Could argue that this is blocked on the Mite backend being finished, on the grounds that it might light the way for what to do in Lite.