Defined in File loss.h
public marian::MultiRationalLoss(Class MultiRationalLoss)
MeanMultiRationalLoss: public marian::MultiRationalLoss¶
Sum of mean losses.
Not really a rational loss as labels are factored into loss. Contribution of losses is equal, same as for ScaledMultiRationalLoss, just divided by different number of labels. See:
L = (1/N sum_i^N L_i + 1/M sum_j^M L_j) = (sum_i^N L_i + N/M sum_j^M L_j) / N
We set labels to 1. During reporting, we would see the same numbers, but gradients are scaled differently which may result in different learning curves.