Add a `Transformer` wrapper for models that supervised that transform

[ablaom]

In pipelines the output of a supervised model that gets propagated to the next component in the pipeline is the output of predict. However, some supervised models also learn a transformation. For example, MLJFlux's NeuralNetworkClassifier and NeuralNetworkRegressor learn entity embeddings to handle categorical inputs, and transform gives access just to these embeddings. We want to use these embeddings as a preprocessing step for some other supervised learner, as in

NeuralNetworkClassifier |> LogisticClassifier

but of course this doesn't work, because the first model is propagating the output of predict instead of transform, because the pipeline apparatus identifies NeuralNetworkClassifier as a Supervised model.

We actually solved this problem in MLJFlux by introducing the EntityEmbedder wrapper, so that the following works:

EntityEmbedder(NeuralNetworkClassifier) |> LogisticClassifier

However, it has struck me rather late, that this wrapper likely works (or should in principle work) for any Supervised model with a transform. So we should really call EntityEmbedder something like Transformer, and perhaps make it immediately available (e.g. by moving it to MLJTransforms.jl).

Thoughts anyone?

@EssamWisam

[EssamWisam]

@ablaom for the name isn't EntityEmbedder a decent name because a definition I found when I first searched for the term embedding is Embedding is a means of representing objects like text, images and audio as points in a continuous vector space. so it seems to be a type of transformation by definition.

As for moving it here, I remember spending a good amount of time to do it here (#22) before doing it on MLJFlux but due to how inconvenient package extensions were you recommended that I don't do so and I was convinced (and likely still convinced).

[ablaom]

Thanks @EssamWisam for pointing out the previous code attempt. That could come in handy.

@ablaom for the name isn't EntityEmbedder a decent name because a definition I found when I first searched for the term embedding is Embedding is a means of representing objects like text, images and audio as points in a continuous vector space. so it seems to be a type of transformation by definition.

I don't see how we could call the generic object EntityEmbedder - as it may have nothing to do with "embedding"; it's just something that wraps any Supervised model that has a transform method, so we it can be treated as a transformer in pipelines (ie. transform not predict gets propagated. But you do raise a discoverability issue. Perhaps we keep EntityEmbedder but just refactor its implementation after we have the generic Transformer.

Another use-case for a Transformer wrapper I ran into today is instances of RecursiveFeatureElimination. In pipelines we may want to view them as transformers, but they have Supervised type.

[ablaom]

I have implemented a Transformer wrapper here. Be great if someone would like to review that. To move things along I've requested registration of the new package MLJWrappers.jl that provides the new Transformer. Next steps:

• MLJWrappers.jl (providing the new Transformer) a dependency of MLJ
• Update the MLJ manual
• Refactor the underlying implementation of MLJFlux EntityEmbedder. I recommend this for maintenance reasons. If there are bugs in this or the other, it would be good to fix them in one place. Also, I noticed some methods/traits don't pass through the EntityEmbedder as expected (if I recall, correctly, training_losses and iteration_parameter are two) - so these get fixed as well.