Defined in File loss.h
public marian::MultiRationalLoss(Class MultiRationalLoss)
We represent loss as pair of expressions, where loss_ is usually a sum of all accumulated loss values per label and count_ is the total number of labels over which the loss was collected.
These two values can then be used to represent various cost variants - for instance label-wise cross-entropy or perplexity. Optimization is only performed with regard to the summed loss_.
Since both, loss_ and count_ are dynamic graph nodes they can be further combined into larger structures. See multi-objective losses below.
@TODO: This seems also used to hold a pair of (logits, mask)
Subclassed by marian::MultiRationalLoss