Inconsistency between the paper and codes

[zjngjng]

the paper says that NNCCS value range is shifted from [-1,+1] to [ε, 2+ε] by adding a 1+ε term, however, the codes didn't add this term max_negative_cloned[index,label.view(-1)]= -1e-12, do we need to add 1+ε for quality score?

[fdbtrs]

you need to +1 if you train the model. For prediction/testing, it is not needed

[davidmcarreira]

@fdbtrs your answer is not clear. Can you please reference where, in your code, you shift the NNCCS values as reported in the paper?