MungoMeng/Registration-CorrMLP

gMLP

Opened this issue · 1 comments

Why does the gMLP model put v+1, what does it do here?

return u*(v + 1.0)

We followed the gMLP implementation used in MAXIM:
https://github.com/vztu/maxim-pytorch/blob/main/maxim_pytorch/maxim_torch.py#L125

I think this might provide a residual connection on u, which might make the training more stable.