np.square(knn.kneighbors(clsts, 2)[1]) Should be changed?

[wpumain]

pytorch-NetVlad/netvlad.py

Line 51 in 8f7c37b

dsSq = np.square(knn.kneighbors(clsts, 2)[1])

Should this code be changed to this? "：
dsSq = np.square(knn.kneighbors(clsts, 2)[0])

knn.kneighbors(clsts, 2) returns two ndarrays, the first ndarray is the distance value, the second is the index value.

What we need here is the distance, not the index, right?

[Nanne]

Thanks for spotting this! Looks like it should indeed be changed to that; would you mind submitting a PR?

[wpumain]

Sorry, my understanding of NetVlad is not deep enough. For example, I don't understand why the value of self.alpha should be set in this way in your program?

pytorch-NetVlad/netvlad.py

Line 53 in 8f7c37b

self.alpha = (-np.log(0.01) / np.mean(dsSq[:,1] - dsSq[:,0])).item()

[fantasy567]

@wpumain

Sorry, my understanding of NetVlad is not deep enough. For example, I don't understand why the value of self.alpha should be set in this way in your program?

pytorch-NetVlad/netvlad.py

Line 53 in 8f7c37b

self.alpha = (-np.log(0.01) / np.mean(dsSq[:,1] - dsSq[:,0])).item()

Here，the appendix A of original paper mentioned that, "The α parameter used for initialization is chosen to be large, such that the soft assignment weights ak(xi) are very sparse in order to mimic the conventional VLAD well. Specifically, α is computed so that the the ratio of the largest and the second largest soft assignment weight ak(xi) is on average equal to 100.", and you can derived the formula to set α for initialization based on equation 2 in original paper when the initial centres and traindescriptors were obtained.
By the way, the code self.alpha = (-np.log(0.01) / np.mean(dsSq[:,1] - dsSq[:,0])).item() is right but dsSq should be return of searching the nearest and second nearest centers for each traindesc, only in this way can get sparse weights to mimic conventional assign weight that only nearest weight is 1 other 0

pytorch-NetVlad/netvlad.py

Line 51 in 8f7c37b

dsSq = np.square(knn.kneighbors(clsts, 2)[1])

This is my understaning，wish it's helpful to you.
So for convenience, @Nanne, appreciate the nice job sincerely.