In Equation 8: What do you mean by \hat{M}_{i,j} is the moving average of M_ij

[jpainam]

Hi, in Equation 8.

You said, \hat{M} is the moving average of M. Please, what do you mean by that? And how did you implement \hat{M}. I can't really locate that part on your code. Thank you.

[chenxiangzZ]

if i!=j and l_g==l_q: cs_ij_=(inputs[i]*inputs[j]).sum(0) #cross_similarity P*P cs_batch = cs_batch+cs_ij_.detach() cs_ij=torch.diag(cs_ij_) #cross_similarity P*P s_constr_= (ss_diff_norm[i]*ss_diff_norm[j]).sum(0) ss_batch=ss_batch+s_constr_.detach() W = cs_ij+(s_constr_-self.ss_mean) if use_matching_loss: x_optim, _= self.IQP_solver(W,self.lambd) x_optim = torch.from_numpy(x_optim).cuda(device_id) i think the code is here!

[jpainam]

 if  i!=j and l_g==l_q:
      cs_ij_=(inputs[i]*inputs[j]).sum(0)   #cross_similarity P*P
      cs_batch = cs_batch+cs_ij_.detach()
      cs_ij=torch.diag(cs_ij_)   #cross_similarity P*P
      s_constr_= (ss_diff_norm[i]*ss_diff_norm[j]).sum(0)
      ss_batch=ss_batch+s_constr_.detach()
      W = cs_ij+(s_constr_-self.ss_mean)
      if use_matching_loss:
         x_optim, _= self.IQP_solver(W,self.lambd)
         x_optim = torch.from_numpy(x_optim).cuda(device_id)
      else:
         x_optim = torch.from_numpy(np.ones(W.size(0))).cuda(device_id)
      matching_targets.append(x_optim)
      matching_logit.append(matching_inputs[i]*matching_inputs[j])
      loss += -torch.matmul(torch.matmul(x_optim,W),x_optim) + ((lambd[i]+lambd[j])*x_optim).sum()/2

loss = loss/len(matching_targets)
cs_batch=cs_batch/len(matching_targets)
ss_batch=ss_batch/len(matching_targets)
self.ss_mean=self.momentum*self.ss_mean+(1-self.momentum)*ss_batch

Thanks .Ok, i guess, \hat{M}_{i,j} is represented as the

self.register_buffer('ss_mean', torch.zeros((part_num, part_num)))

which you update as

self.ss_mean=self.momentum*self.ss_mean+(1-self.momentum)*ss_batch

[hh23333]

Yes

[jpainam]

Thanks for your reply