Multilabel classification

[peichins]

Currently, inference/classify is only set up to do single-label classification. It's quite common both in training and running a classifier that more than one class will appear in the same example.

The current way to build a labelled dataset is through a folder-of-folders, where each subfolder represents a class. Therefore, each example must belong to only one class. If an example contains more than one label, and therefore belongs to more than one class, we need to either choose one of the folders, put a copy of the same example in both folders or exclude it from training. All of these are problematic.