Allow Accuracy to return metric per class

[SkafteNicki]

🚀 Feature

Implement the average argument like in Precision and Recall such that accuracy metric can return the metric per class label.

Motivation

Sometimes it may be beneficially to look at the accuracy per label, especially when working with very unbalanced datasets

Pitch

Alternatives

Additional context

[arv-77]

Dear @SkafteNicki , can I work on this?

[SkafteNicki]

@arvindmuralie77 yes, I will assign you. Thanks for wanting to contribe :]
Please make sure that the feature follows the same pattern as with precision, recall ect. by adding an argument average which normally can take the values 'micro', 'macro', 'weighted', 'none' and None where the two last corresponds to returning the score per class.

[arv-77]

Sure, will do. Thanks @SkafteNicki

[arv-77]

I was thinking of changing the parent class from Metric to StatScores in Accuracy. This way, I can use tp, fp, tn, and fn to calculate accuracy for all values of average. Would that be alright or should I try without changing the base class @SkafteNicki ?

[SkafteNicki]

@arvindmuralie77 Fine by me, that is basically also what I would do :]

[arv-77]

Hey @SkafteNicki , I have created a draft for this issue. When pytest for accuracy is run, 23 tests are failing out of 102 because the mdmc_average parameter does not have default value. If it is set to global all test cases are passing.
I think new test cases will have to be added since there are new parameters.