The code does not match the description of the paper

[USTC-Keyanjie]

Thanks for the awesome contribution!

When I read the code, I found that the code did not match the paper.

https://github.com/XinJCheng/CSPN/blob/24eff1296c282196f1c919714d8f32a9b0dbe7fb/cspn_pytorch/models/torch_resnet_cspn_nyu.py#L269

According to the description of your paper, I think we should change

out = self.relu(self.bn1(self.conv1(x)))
out = torch.cat((out, side_input), 1)

to

x = self.relu(self.bn1(self.conv1(x)))
out = torch.cat((x, side_input), 1)

Looking forward to your reply!

[JUGGHM]

Thanks for the awesome contribution! When I read the code, I found that the code did not match the paper.

  CSPN/cspn_pytorch/models/torch_resnet_cspn_nyu.py

     Line 269
  in
  24eff12

       out = self.relu(self.bn1(self.conv1(x))) 

According to the description of your paper, I think we should change out = self.relu(self.bn1(self.conv1(x))) out = torch.cat((out, side_input), 1)

to x = self.relu(self.bn1(self.conv1(x))) out = torch.cat((x, side_input), 1)

Looking forward to your reply!

I wonder is there any difference for the variable

out

in two versions?

[USTC-Keyanjie]

The out indeed has no difference in two version. However, x has difference.

According to the description of the paper, x must pass the conv, bn and relu operations before entering the shortcut section.

In my opinion, the two version looks like and

[JUGGHM]

The out indeed has no difference in two version. However, x has difference. According to the description of the paper, x must pass the conv, bn and relu operations before entering the shortcut section. In my opinion, the two version looks like

and

You're right

[XinJCheng]

Thanks for your attention and detailed review. Yes, this is a mistake, I will update it later. Thanks