Quick question about Implementation on cifar100

[Dingseewhole]

Thanks for this interesting paper. I have a quick question after comparing the paper and the code.
In the paper, I see "For the proposed Distillation of Colliding Effect (DCE), we used cosine distance to measure the similarity between images." But in the incremental_train_and_eval_DCE_MR_LF_TDE.py I see the code use L2 distance to generate the match_id at line 146 and function pdist
Is this dis_z was used to select knn inputs? If it is that, why in code is L2_dist but not cos_dist?

[JoyHuYY1412]

Thanks for this interesting paper. I have a quick question after comparing the paper and the code.
In the paper, I see "For the proposed Distillation of Colliding Effect (DCE), we used cosine distance to measure the similarity between images." But in the incremental_train_and_eval_DCE_MR_LF_TDE.py I see the code use L2 distance to generate the match_id at line 146 and function pdist
Is this dis_z was used to select knn inputs? If it is that, why in code is L2_dist but not cos_dist?

Thank you for pointing it out! I meant to write "euclidean" distance...

Sorry for bringing your confusion! You can also try cosine distance, and I think maybe you can get slightly better results.

[Dingseewhole]

Thanks for the quick reply!
So in the code incremental_train_and_eval_DCE_MR_LF_TDE.py and the function pdist, you actually use euclidean distance to select the KNN match_id, which means the topk smallest euclidean distance inputs will be sellected as match_id.
Am I right?

[JoyHuYY1412]

Thanks for the quick reply!
So in the code incremental_train_and_eval_DCE_MR_LF_TDE.py and the function pdist, you actually use euclidean distance to select the KNN match_id, which means the topk smallest euclidean distance inputs will be sellected as match_id.
Am I right?

Yes! You are right ^_^