In PyTorch, since class labels are zero-indexed, actually only non “-1” target values are considered, instead of "The criterion only considers the first non zero y[j] targets."