Legacy API: Should we migrate its entry-points?

[TomFinley]

The plan is to deprecate then delete the legacy API. While most items in it are of questionable worth, the legacy API does contain some components that might be worth saving, and some we may have to save.

Consider this entry-point definition, in the legacy project.

machinelearning/src/Microsoft.ML.Legacy/Runtime/EntryPoints/ModelOperations.cs

Line 81 in 7b2461c

[TlcModule.EntryPoint(Name = "Transforms.ModelCombiner", Desc = "Combines a sequence of TransformModels into a single model")]

As it is an entry-point, it was duly published in NimbusML as we see here, and this internal entry-point definition actually wound up being used here.

There are a couple questions we might want to ask. The legacy pipeline API was entry-point based, so anything that was part of a pipeline had to be an entry-point (I think). So were these entry-points defined in legacy specifically intended to be useful beyond legacy?

If we want to keep this entry-point, and others like it, we should have a migration plan somewhere, since legacy is to be deleted.

[sfilipi]

@asthana86 had plans to use the entry points. @asthana86 is that still the case?

[Zruty0]

@ganik , I suggest we make the necessary update to NimbusML to not rely on legacy namespace, and migrate out whatever entry points are needed for that

[ganik]

@Zruty0 agreed. I ll check what entrypoints need to be migrated for NimbusML to work

[shmoradims]

@Zruty0 agreed. I ll check what entrypoints need to be migrated for NimbusML to work

@ganik could you please add the NimbusML issue here?

[sfilipi]

Assigning to @ganik based on the discussion.