Multidimensional output -> need loss and doc

[meilame-tayebjee]

Multidimensional output (for a text, prediction is a vector instead of an int)

Use cases:

• Multilabel ground truth (Request: Support for Multilabel #10) -> (0 1 ... 0 1.. 0) (size of this vector is vocab size)
• Hierarchical classification / simultaneous multiclass classif: one int per "level" of the nomenclature, or per output "type" (for instance if we want to simultaneously predict the topic + sentiment of a text...) -> (5 32 56 1) (size of this vector is number of levels/types)
• Soft classification : instead of having one single label as ground truth, any label has a certain probability to be true -> (0.1 0.8 0 0 .. 0.1). It can be either in a multiclass (competition between label) or multilabel mode (several labels can be true)
  • in the multiclass mode, you want to use nn.CrossEntropyLoss() and your output will sum to 1 (probability distribution) -> you may also want to enforce that the ground truth should sum to 1
  • in the multilabel mode, you will apply nn.BCEwithLogitsLoss()to each coordinate of your logits vector - no sum to 1 constraint at all (i.e. you could typically have (0.1 0.8 0 0.7 0.2) as prediction (or ground truth)

Examples illustrating of all these use cases would be nice.