malfunctional metrics?

[pmeier]

The documentation of MEGNetProbModel warns in two locations, that the metrics are "malfunctional". If that is the case, why provide the ability to use malfunctional ones and not the working ones?

On a side note, pointing the users the unlocknn.metrics is a little peculiar in the light of #26.

[a-ws-m]

The code was designed to try to make the most effective possible usage of Keras's API, which provides an interface for computing model metrics during training (which ones to choose would be selected with the metrics attribute) and evaluation (using the model.evaluate method); these are the two use cases that have the warnings attributed to them. They were implemented as the most natural way to extend the Keras API; however, I found that they gave vastly different (worse) values from the ones I computed myself using the predictions from model.predict and the true target values.
The exact problem is hard to identify. I suspect it's a combination of the postprocessing target scaling that MEGNet implements, and the fact that probabilistic layers (from TensorFlow Probability, TFP) seem poorly supported within Keras, which has forced me to implement a couple of workarounds. I hope that as TFP gets more established, it will become better integrated with Keras and therefore that eventually unlocknn.metrics will be made redundant and the aforementioned methods can be fixed.
For now, I'd like to keep the methods in the code so that they don't need re-implementing when the chance arises, and to make it clear to other potential contributors that the implementation has at least been tried.
I suspect the best thing to do for now may be to make the method and attribute private (which would remove evaluate from the docs, at least); does that sound reasonable?