MLJModelInterface.jl

MLJModelInterface.jl copied to clipboard

Explain in docs that model wrappers need to overload save/restore

1

Context: https://github.com/alan-turing-institute/MLJ.jl/issues/1099#issue-2165552624

ablaom

I don't understand why we need a separate `metadata_pkg` when we have `metadata_model`, as both take a model type as first argument. We should just move all the traits into...

ablaom

It seems some macros in metadata_utils cause issues if an MLJ interface is wrapped in module

1

https://github.com/alan-turing-institute/MLJ.jl/issues/744#issuecomment-786746319 See also this comment: "A quirk of submodules and evaluation scopes makes it necessary to load this submodule in the package __init__ function" from #20

ablaom

`params(::MLJType)` method

1

I would like to remove this method, as it's essentially "wrong", because it is field-based, instead of property-based. It is my understanding that the method does not appear in any...

ablaom

invalid model parameters should either error or keep user arguments

4

I noticed today that if a model constructor fails parameter validation it reverts to the default with a warning. I find this behavior confusing. Normally invalid constructors lead to an...

ExpandingMan

Currently returns `(nothing, nothing, nothing)` but reports should be named tuples.

ablaom

We already have `supports_training_losses`. It also seems that documentation does not require the fallback, but this should be checked.

ablaom

ablaom

Perhaps MLJModelInteface should not export any method in the future

1

This has got us into trouble before: #131 #124

ablaom

Due to a recent change in `ScientificTypesBase,jl` from version 2.0 to 3.0, the `ScientifictypesBase.scitype` method is now a two argument method. Packages depending on `MLJModelInterface.jl` previously depended on the previous...

OkonSamuel