Support multi-class classification

[Firenze11]

Summary

The project currently doesn't support multiclass classification models, this should be valuable and relatively simple to implement.

Possible Solution

• Each yPred files have multiple columns, each indicating predicted probability of one class
• Automatic clustering algorithm uses nModels * (nClasses - 1) columns to do segmentation
• Visualization doesn't have to change since one data point is still associated with one metric per model, which is the log-loss.
• If users want to segment by performance on one single class, we can implement that based on user-defined performance metric (#105), where user specify the metric to be some sort of loss function comparing the prediction for class N with the ground truth for class N

[imneonizer]

Does it support object detection models? If yes, how do i prepare my prediction data for manifold to be able to visualize model performace. what would be the contents of features dataset, pixel values?