bg2crw / top_k_optimization Goto Github PK

Thanks for the interesting paper!
I was trying to replicate your results, specifically, on the caltech birds dataset.
I'm running your code as is, using method 8 (Another_precision_@k loss), but the metrics seem to significantly underperform the published results.

Are there any hyper parameter differences between what was used to generate the results in the paper versus what is published in this repo?

Thanks!

loss = torch.sum(- T * F.log_softmax(D, -1), -1), should D be -D since we want to minimize the distance between the example with its proxy with same label.

Thanks for sharing your codes.
I have read your codes and paper.
Is the proposed method implemented in SmoothPrec_at_K.py?
I think there are some differences between the paper and SmoothPrec_at_K.py.

Hi,
I've been wondering about the regularization term that you introduced in the code, but isn't mentioned anywhere in the paper (see here, lines 82-88). Is there any specific motivation for it, or can you quantify how much this contributed to the performance of your loss, compared to a non-regularized version?