Global loss and shared individually connected layers

[LazyLiunian]

I read your paper, but I can't find the global loss and shared individually connected layer in your code.
Can you tell me where can I use these module?

[LXN711]

I also have this doubt. I cannot find the expression of the relevant formula in the code, only unshared individually connection layer.If you understand this problem, can you discuss it with me?

[LazyLiunian]

https://github.com/bruceyo/GLA-GCN/blob/master/common/s_agcn.py
I mischaracterized the problem earlier, I only found the shared individually layer, in s_agcn.py.
I consider self.fc in this module to be the shared individually layer.

class S_AGCN(TemporalModelBase):

      def __init__(self, num_joints_in, in_features, num_joints_out,
                   filter_widths, causal=False, dropout=0.25, channels=96, dataset='h36m'):
  
          super().__init__(num_joints_in, in_features, num_joints_out, filter_widths, causal, dropout, channels, dataset)
  
          layers_tcngcn = []
          num_person = 1
          in_channels = 2
          num_point = num_joints_in
          self.graph = Graph(dataset)
          self.data_bn = nn.BatchNorm1d(num_person * in_channels * num_point)
          A = self.graph.A
          self.expand_gcn = TCN_GCN_unit(2, channels, A)
  
          self.causal_shift = []
          next_dilation = filter_widths[0]
          for i in range(0, len(filter_widths)):
              self.pad.append((filter_widths[i] - 1)*next_dilation // 2)
              self.causal_shift.append((filter_widths[i]//2 * next_dilation) if causal else 0)
              layers_tcngcn.append(TCN_GCN_unit(channels, channels, A))
              layers_tcngcn.append(TCN_GCN_unit(channels, channels, A, stride=filter_widths[i], residual=False))
              next_dilation = next_dilation * filter_widths[i]
  
          self.layers_tcngcn = nn.ModuleList(layers_tcngcn)
          self.fc = nn.Conv1d(channels, 3, 1)
  
      def set_bn_momentum(self, momentum):
          self.data_bn.momentum = momentum
          self.expand_gcn.gcn1.bn.momentum = momentum
          self.expand_gcn.tcn1.bn.momentum = momentum
          for layer in self.layers_tcngcn:
              layer.gcn1.bn.momentum = momentum
              layer.tcn1.bn.momentum = momentum
  
      def _forward_blocks(self, x):
  
          N, V, T = x.size()
          v = torch.div(V, 2, rounding_mode='floor') # number of 2D pose joints
          x = self.data_bn(x)
          x = x.view(N, 1, v, 2, T)
          x = x.permute(0, 1, 3, 4, 2).contiguous()
          x = x.view(N, 2, T, v)
  
          x = self.expand_gcn(x)
          for i in range( len(self.pad) -1):
              res = x[:, :, self.causal_shift[i] + self.filter_widths[i]//2 :: self.filter_widths[i], :]
  
              x = self.drop(self.layers_tcngcn[2*i](x))
              x = self.drop(self.layers_tcngcn[2*i+1](x))
              x = res + x
          pose_3d_ = x
  
          pose_3d = torch.zeros(N, 3, v).cuda(0)
          for i in range(0,v):
              pose_joint_3d = pose_3d_[:,:,:,i].mean(2)
              pose_joint_3d = pose_joint_3d.view(N, -1, 1)
              pose_joint_3d = self.fc(pose_joint_3d)
              pose_3d[:,:,i] = pose_joint_3d.view(N,-1)
  
          return pose_3d

I also have this doubt. I cannot find the expression of the relevant formula in the code, only unshared individually connection layer.If you understand this problem, can you discuss it with me?

[LXN711]

Thank you for your reply. Perhaps you are right. That self.fc is a shared connection layer, because I have only looked at this code and haven't run it yet. and study it carefully. My question is that the description and code implementation of the individual connection layers in the paper do not seem to correspond. I am troubled by this issue.Do you have this question?

[LazyLiunian]

Thank you for your reply. Perhaps you are right. That self.fc is a shared connection layer, because I have only looked at this code and haven't run it yet. and study it carefully. My question is that the description and code implementation of the individual connection layers in the paper do not seem to correspond. I am troubled by this issue.Do you have this question?

Yes，I also have this issue.