Return new model when calling `fit`

[lars-reimann]

Is your feature request related to a problem?

Methods of data containers like Table always return a new data container. The fit method of models, however, mutates a model in place.

Desired solution

For the sake of consistency, we should consider whether fit should also return a new model. If we implement that, we should also think about differentiating between trained and untrained models. Only the latter should have a predict methods or methods for metrics (#64).

Possible names (likewise for regressors):

• ClassifierDraft
• ClassifierTemplate
• ClassifierBlueprint
• ClassifierPrototype - It's not quite a prototype though, since we don't just clone an instance when we fit. We also initialize weights etc. So that's not my favorite
• ClassifierBuilder - We don't use the builder pattern though, values can only be set in the constructor. So that is not my favorite either.

Possible alternatives (optional)

No response

Screenshots (optional)

No response

Additional Context (optional)

No response

[lars-reimann]

I've decided against differentiating trained and untrained models in Python for several reasons:

• The trained models must be part of the public interface because they should be usable as a type. However, we cannot prevent in this case that a user can also instantiate them directly, so it's still possible to call predict without calling fit. Both classes also show up in the documentation and it is confusing for users whether they need to work with a Classifier or a ClassiferDraft. Many would pick the wrong thing (Classifier) due to the shorter name.
• Even with one class we can still differentiate between trained and untrained models in the DSL. There we can omit the constructor of untrained models.

Because of this, I'll only change that fit returns a new model.

[lars-reimann]

🎉 This issue has been resolved in version 0.6.0 🎉

The release is available on:

• v0.6.0
• GitHub release

Your semantic-release bot 📦🚀