About the loss of SDE_3Dto2D

[MorningEatDinner]

I observed that in code SDEModel3Dto2D_node_adj_dense_02, loss is calculated by:
losses_x = torch.square(score_x + z_x) # [B, max_num_nodes, num_class_X] or [B, max_num_nodes, 1]
losses_adj = torch.square(score_adj + z_adj) # [B, max_num_nodes, max_num_nodes]

But in SDE_model_2d_to_3d, the code to calculate loss is
loss_pos = torch.sum((scores - pos_noise) ** 2, -1) # (num_node)

I'm confused why the 3D_to_2D code isn't score - x.

[chao1224]

Hi @MorningEatDinner
I am not sure what you mean by why the 3D_to_2D code isn't score - x. Can you explain this further?