Issue 139: MLP multi-class classifier

[Hector-hedb12]

Resolves #139

[Hector-hedb12]

@csala as softmax is just a value in the last_dense_activation parameter, I was thinking in change the name of this primitive from keras.Sequential.MLPSoftmaxClassifier.json to keras.Sequential.MLPClassifier.json. In that way, we could cover the binary classification case as well. I think it is the same network structure 🤔 . What do you think?

[csala]

@Hector-hedb12 I think that we can keep them separated, as there are a couple of other differences, and they clearly respond to different needs (binary/multi-class).

What I would do, though, is renaming this to MLPMultiClassClassifier, so we can have Binary/MultiClass.

[csala]

Since input_dim represents the number of vectors being passed as input and depends on the fit data, this should be set as a fit argument instead of a hyperparameter.
It should also not have a default, as we want to user to input this.

[csala]

When setting this as a fit argument, I would call it features (see #147), and add a description explaining that this is the number of features in X

[csala]

This should be separated as dense_1_units and dense_2_units and set as tunable hyperparameters.
I'm not sure about the ideal range, but in some other cases we have used (1, 500).

[csala]

This should be separated as droput_1_rate and dropout_2_rate

[csala]

Since this example is very quick, let's remove this and let the default (20) be used.

[Hector-hedb12]

Updated @csala

[csala]

Looks good @Hector-hedb12 !