Question about loss function

[Geeks-Z]

In engine.py line 68
loss = loss - args.pull_constraint_coeff * output['reduce_sim']
In the paper is
loss = loss + args.pull_constraint_coeff * output['reduce_sim']?

[prachigarg23]

Hi @Geeks-Z , I believe this is a typing error in the paper because the objective is to ' pull selected keys closer to corresponding query features' by maximising the cosine similarity of top_k selected keys. Hence, the minus sign in the code should be correct.