jcy132 / hneg_src Goto Github PK

Sorry to bother, I used your pretrained seg models and cut models to test, but I found the acc pixAcc of cityscapes was 83.2, much larger than that of in the paper.
Did I do thing s wrong?

class PatchHDCELoss(nn.Module):
   ...
   ...
        diagonal = torch.eye(npatches, device=feat_q.device, dtype=self.mask_dtype)[None, :, :]
        l_neg_curbatch.masked_fill_(diagonal, -10.0) 
        l_neg = l_neg_curbatch.view(-1, npatches) # [npatches, npatches]

        logits = (l_neg-l_pos)/self.opt.nce_T # [npatches, npatches]
        v = torch.logsumexp(logits, dim=1)    # [npatches]
        loss_vec = torch.exp(v-v.detach())    # [npatches] all elements are 1.0

        # for monitoring
        out_dummy = torch.cat((l_pos, l_neg), dim=1) / self.opt.nce_T # [npatches, npatches+1]
        CELoss_dummy = self.cross_entropy_loss(out_dummy, torch.zeros(out_dummy.size(0), dtype=torch.long, device=feat_q.device))

        loss = loss_vec.mean()-1+CELoss_dummy.detach()

        return loss

Hi, I don't understand 'loss = loss_vec.mean()-1+CELoss_dummy.detach()', which seems to output the same results of 'CELoss_dummy.detach()'. Because 'loss_vec = torch.exp(v-v.detach())' is a tensor with all values of 1, I think ‘loss_vec.mean()-1’ is a tensor with all values of 0. I don't know if this loss is same with infoNCE?

Hello, the research work in this article is particularly meaningful. The batch size in the code is set to 1 by default, but the following problems occur when I set the batch size to 2 or 3. Can you help me solve this problem.
Looking forward to your reply.

File "train.py", line 44, in
model.data_dependent_initialize(data)
File "/home/cvers/Joe/SM/models/cut_model.py", line 176, in data_dependent_initialize
self.compute_G_loss().backward() # calculate graidents for G
File "/home/cvers/Joe/SM/models/cut_model.py", line 289, in compute_G_loss
self.loss_SRC, weight = self.calculate_R_loss(real_A_pool, fake_B_pool, epoch=self.train_epoch)
File "/home/cvers/Joe/SM/models/cut_model.py", line 335, in calculate_R_loss
loss_SRC, weight = crit(f_q, f_k, only_weight, epoch)
File "/home/cvers/anaconda3/envs/airnet/lib/python3.8/site-packages/torch/nn/modules/module.py", line 722, in call_impl
result = self.forward(*input, **kwargs)
File "/home/cvers/Joe/SM/models/SRC.py", line 55, in forward
weight_seed.masked_fill(diagonal, -10.0)
RuntimeError: The expanded size of the tensor (768) must match the existing size (256) at non-singleton dimension 2. Target sizes: [1, 768, 768]. Tensor sizes: [1, 256, 256]

Dear author, where can I get the supplementary materials mentioned in your question?

Hi, thanks for your great work!

I have a question about the PatchHDCELoss implementation in the Single-modal setting.

In models/hDCE.py,

class PatchHDCELoss(nn.Module):
   ...
   ...
        diagonal = torch.eye(npatches, device=feat_q.device, dtype=self.mask_dtype)[None, :, :]
        l_neg_curbatch.masked_fill_(diagonal, -10.0) 
        l_neg = l_neg_curbatch.view(-1, npatches) # [npatches, npatches]

        logits = (l_neg-l_pos)/self.opt.nce_T # [npatches, npatches]
        v = torch.logsumexp(logits, dim=1)    # [npatches]
        loss_vec = torch.exp(v-v.detach())    # [npatches] all elements are 1.0

        # for monitoring
        out_dummy = torch.cat((l_pos, l_neg), dim=1) / self.opt.nce_T # [npatches, npatches+1]
        CELoss_dummy = self.cross_entropy_loss(out_dummy, torch.zeros(out_dummy.size(0), dtype=torch.long, device=feat_q.device))

        loss = loss_vec.mean()-1+CELoss_dummy.detach()

        return loss

I can't understand your intentions in this implementation.
Which equation in your paper did you implement?
I cannot match this loss to a equation in the paper.
I think I missed something. Please give me some explanation.
Thank you!