Add intRVFL to the models module

[denkle]

Add intRVFL to the models module

[denkle]

Thanks for the refactoring, Mike! I like how the new code looks like. Functionality-wise, it is all good, I only added a few minor edits.

Few minor edits made:

1. train_loader in “UCI_benchmark_intRVFL.py” is not needed anymore with the current code. [Fixed].
2. In “UCI_benchmark_intRVFL.py”, added transform() to data when calling “fit_ridge_regression()” - was affecting performance a lot.
3. In “models.py”, kappa is not mentioned as a part of description in “Args”. [Fixed]
4. In “UCI_benchmark_intRVFL.py”, ridge regression was missing (not a big effect on results though) when calling “fit_ridge_regression()”. [Fixed]

[denkle]

Few additional considerations while going through the code:

1. It could make sense to have alpha as an input to IntRVFLRidge() and then avoid calling it explicitly with “fit_ridge_regression()”.
2. I would hide “one_hot_targets” under the hood in “fit_ridge_regression” as it is unclear what does it give to the user by explicitly seeing formation of one-hot encodings in “UCI_benchmark_intRVFL.py” but I see the extra hassle with sending in "num_classes".
3. After refactoring, “fit_ridge_regression()” is a function within IntRVFLRidge() class. Before it was meant to be a separate function just because it is a primitive for a particular type of classifier (like centroids) so it does not have to stay exclusively within IntRVFLRidge(). But perhaps it is not so complicated so one could copy-paste this piece of code every time it is needed in some other place. All these considerations are not critical and perhaps not worth extra discussion so the final word about them is yours.

[mikeheddes]

Thanks for spotting my bugs!

• From what I could tell the alpha parameter is specific to the learning method ridge regression so if we end up adding more methods they could have their own parameters which are independent of the class instance. Let me know if I'm misinterpreting that?
• The reason I moved the one-hot encoding outside ridge regression is because you had a comment that said something like you could also have multiple class vectors per class. By letting the user handle the one-hot encoding it could split one class over multiple class vectors. Do you think this makes sense?
• I will refactor your last point

[denkle]

• I agree that the alpha is specific to the learning method ridge regression. I was probably assuming that if we have other methods for training the classifier we would create another class instance something like IntRVFLAbc() where "Abc" would refer to the specific classifier used for the model (currently IntRVFLRidge suggests that ridge regression is used as the classifier). Maybe it does not make sense to have multiple class instances and, instead, one should specify different ways of forming the classifier within the existing class so specifying "fit_abc()" method instead.
• Okay, that is right. I have not made myself clear enough in that comment. It targeted prototype-based classifiers (like GLVQ) for ridge regression based classier, I am not sure if one could benefit from multiple class vectors per class. So for fit_ridge_regression(), I think we can safely have one-hot encodings inside the function, thus, having less details exposed to the user.
• Great!