I try to reproduce the SOSR based on the Harnet, but some questions are induced. Can you give me some advice?

[Ahandsomenaive]

I try to reproduce the SOSR based on the Harnet, but some questions are induced. Can you give me some advice?

the code about loss is here:

`import torch import torch.nn as nn import sys

def distance_matrix_vector(anchor, positive): """Given batch of anchor descriptors and positive descriptors calculate distance matrix"""

d1_sq = torch.sum(anchor * anchor, dim=1).unsqueeze(-1)
d2_sq = torch.sum(positive * positive, dim=1).unsqueeze(-1)

eps = 1e-6
return torch.sqrt((d1_sq.repeat(1, positive.size(0)) + torch.t(d2_sq.repeat(1, anchor.size(0)))
                  - 2.0 * torch.bmm(anchor.unsqueeze(0), torch.t(positive).unsqueeze(0)).squeeze(0))+eps) 

def inner_dot_matrix(anchor, postive):

inner = torch.mm(anchor, torch.t(postive))
mask = torch.eye(inner.size(1)).cuda() 
inner = inner - 1e-6*mask
dist_m = torch.sqrt( 2.0*(1.0-inner) + 1e-8)
return dist_m

def loss_SosNet(anchor, positive, anchor_swap = False, anchor_ave = False,\ margin = 1.0, batch_reduce = 'min', loss_type = "triplet_margin",k = 8): """HardNet margin loss - calculates loss based on distance matrix based on positive distance and closest negative distance. """

assert anchor.size() == positive.size(), "Input sizes between positive and negative must be equal."
assert anchor.dim() == 2, "Inputd must be a 2D matrix."
eps = 1e-8
dist_matrix = distance_matrix_vector(anchor, positive) +eps
eye = torch.autograd.Variable(torch.eye(dist_matrix.size(1))).cuda()

# steps to filter out same patches that occur in distance matrix as negatives
pos1 = torch.diag(dist_matrix)
dist_without_min_on_diag = dist_matrix+eye*10
mask = (dist_without_min_on_diag.ge(0.008).float()-1.0)*(-1)
mask = mask.type_as(dist_without_min_on_diag)*10
dist_without_min_on_diag = dist_without_min_on_diag+mask

if batch_reduce == 'min':
    min_neg = torch.min(dist_without_min_on_diag,1)[0]
    if anchor_swap:
        min_neg2 = torch.min(dist_without_min_on_diag,0)[0]
        min_neg = torch.min(min_neg,min_neg2)
    pos = pos1
else: 
    print ('Unknown batch reduce mode. Try min, average or random')
    sys.exit(1)

if loss_type == "triplet_margin":

    loss = torch.clamp(margin + pos - min_neg, min=0.0)**2

else: 
    print ('Unknown loss type. Try triplet_margin, softmax or contrastive')
    sys.exit(1)

dist_matrix_a = inner_dot_matrix(anchor, anchor)+ eps
dist_without_min_on_diag_a = dist_matrix_a+eye*10
mask_a = (dist_without_min_on_diag_a.ge(0.008).float()-1.0)*(-1)
mask_a = mask_a.type_as(dist_without_min_on_diag_a)*10
dist_without_min_on_diag_a = dist_without_min_on_diag_a+mask_a

cur_a = torch.topk(dist_without_min_on_diag_a, k=k, dim=1, largest=False).indices  
#print(cur_a)      
mask_aa = torch.zeros_like(dist_without_min_on_diag_a)
#print(mask_aa)
mask_aa = mask_aa.scatter(1,cur_a,1)

dist_matrix_p = inner_dot_matrix(positive, positive)+ eps
dist_without_min_on_diag_p = dist_matrix_p+eye*10
mask_p = (dist_without_min_on_diag_p.ge(0.008).float()-1.0)*(-1)
mask_p = mask_p.type_as(dist_without_min_on_diag_p)*10
dist_without_min_on_diag_p = dist_without_min_on_diag_p+mask_p

cur_p = torch.topk(dist_without_min_on_diag_p, k=k, dim=1, largest=False).indices        
mask_pp = torch.zeros_like(dist_without_min_on_diag_p)
mask_pp = mask_pp.scatter(1,cur_p,1)

sosr = torch.sqrt( 1e-8 + torch.sum( (dist_without_min_on_diag_a * mask_aa - dist_without_min_on_diag_p * mask_pp)**2, dim = 0 ))
sosr = torch.mean(sosr)

loss = torch.mean(loss)
return loss + sosr , loss, sosr`

[Ahandsomenaive]

loss is not converge！ FPR95 is very high！

[Ahandsomenaive]

now, i find the coding bug~ But, the best score is 1.15, and have not got 1.03 described in your paper! would you like to give me some advice about training setting?

[Ahandsomenaive]

now, i find the coding bug~ But, the best score is 1.15, and have not got 1.03 described in your paper! would you like to give me some advice about training setting?

[ACuOoOoO]

@Ahandsomenaive, hi. I have embeded the SOSR into HardNet (https://github.com/DagnyT/hardnet). And changed the optimizer as the paper advised and the model published here. In this way, the best score I reproduced reaches ~1.10 but it is still far to catch 1.03. However, it's interestingly to find that my reproduction trained on Liberty outperform the paper declared in Hpatch by a little.

[Ahandsomenaive]

@Ahandsomenaive, hi. I have embeded the SOSR into HardNet (https://github.com/DagnyT/hardnet). And changed the optimizer as the paper advised and the model published here. In this way, the best score I reproduced reaches ~1.10 but it is still far to catch 1.03. However, it's interestingly to find that my reproduction trained on Liberty outperform the paper declared in Hpatch by a little.

Hi, Thank you~ a little question, i can't find your code according to the link you provided. Would you like to provide detailed information about your code? Thank you~