In the current form, the reduction type should be "pos_pair". Alternatively, the loss calculation could be changed to be per element. See issue #329