So, @ABostrom, @chrisholder, @jambo6, @TonyBagnall, @aiwalter, @mloning, I wanted to hear your opinion on a suggestion how to deal with descriptive type hints and type checks in one.

What I would propose is that our base class fit etc look like as follows (below without conversion etc), in the example of the forecaster base class:

@scitype_decorator
def fit(X: Series, Y: Series):
    
    X = check(X)
    Y = check(Y)

    stuff(X,Y)

    self._fit(X.inner, Y.inner)

About the ingredients above:

Series is a custom type, or perhaps better class that denotes scitype, not machine type
what scitype_decorator does is wrap X and Y in the declarative type (ligh-weight wrapper), and does part of what check_series does now: it annotates X, Y, with the machine type. E.g.,: X = Series(X), which moves X to X.inner and adds an X.scitype = "Series" attribute
check inspects X for validity, and adds a X.scitype = "DataFrame" or similar annotation. So after check, we know that X is valid, and we can read the scitype and the mtype using X.scitype and X.mtype.
the remaining stuff has access to these declarative type annotations. It could be compatibility checks, conversions, etc.

To avoid misunderstandings:

we don't expect the user to wrap X and Y in Series already. They can pass a pd.DataFrame or pd.Series or whatever they like. The type hint is to be read like the union type roughly. The wrapping happens in the decorator.

This proposal has some advantages:

type annotation in the base class and the user docstrings will be with the "scientific type"
the scientific type of any object is inspectable, per declarative annotation
we have a consistent way, across modules, to track and inspect alias mtypes for the same scitype
this could be eaily combined with a central "register" location of scitypes, mtypes, and/or a central location for validity checking, and the conversions module as sketched in Forecasting support for multivariate y and multiple input/output types - working prototype #980

Uh oh!

Type aliasing #973

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions