A question about the code in sg_net.py about sg_pr

[kxhit]

Thanks for your interest to our work!

We found the NTN module asymmetric to the input pair that means if we exchange the pair order (e.g., [input1: scene0001 and input2: scene0080] VS [input1:scene0080 and input2: scene0001]), the output similarity score is different. This is wrong as we hope for a pair the similarity should be consistent regardless their input order. So we append the feature and target again to train the network to fix this problem.

Well I think a better way is to modify the network structure to ensure the output is order-invariant to the input rather than do data augmentation by appending. Let me know if I make it clear.

[cloudcjf]

Thanks for your response! Yeah, your explanation is quite clear. I will close this issue.

[YangSiri]

Hi, @kxhit

Is there a duplication of func eval_batch_pair in the 459 and 503 row of sg_net.py ?

def eval_batch_pair(self, batch):

[kxhit]

Hi @YangSiri

Thanks for pointing it out. Yeah, there is some dead code. Sorry for not cleaning the code well.