Model parsing discussion

[landmanbester]

As promised in the meeting, I am opening this to discuss all things model parsing related @o-smirnov @JSKenyon @sjperkins @bennahugo. It seems that there is a need to parse complex models, possibly consisting of a mixture of component models (for both morphology and frequency spectrum) as well as pixelated images/facets. The former should be evaluated as accurately as possible using a parametrised DFT while the latter get's processed using a degridder. This functionality already exists (to some extent) in CubiCal but the model specification has become rather convoluted. The dream is to have a single parser that can concatenate different combinations of models and keep track of how each one should be predicted. I think the consensus was that tigger-lsm should be the parser with the API possibly living in codex-africanus. I see we were already thinking along these lines in 2019 ratt-ru/codex-africanus#85. But it may also be preferable to have a separate application to handle the predict. Ideas?

[sjperkins]

I was thinking about this yesterday.

• The way the model is currently specified in codex requires working it into NumPy arrays, which in turn means that sources with different characteristics must be partitioned into groups which must be predicted separately.
• Alternatively, it may be possible to come up with some suitably generic dictionary/list model that can be represented in Numba.
• This means that the dictionary/list will be strictly typed so it won't be quite as loose as a pure Python dictionary, but it may be loose enough that we can process multiple source types in one function.