# Class ScaledMultiRationalLoss¶

## Class Documentation¶

class ScaledMultiRationalLoss : public marian::MultiRationalLoss

Scaled sum of losses.

This can weigh losses equally by choosing the first loss_0 as a reference and scaling all remaining losses loss_i by count_0 / count_i. Labels are summed up by the same rule. By this we simulate a sum of losses at similar scales. Dividing by scaled label counts yields a value close to an equally weighted sum of means.

L = sum_i^N L_i + N/M sum_j^M L_j

We set labels to N. When reporting L/N this is equivalent to sum of means. Compare to sum of means below where N is factored into the loss, but labels are set to 1.

Public Functions

ScaledMultiRationalLoss()
ScaledMultiRationalLoss(const RationalLoss &rl)