Hi,
It seems line 28 of the gem_cnn.transform.py ( projected = torch.einsum('pij, pj -> pi', projectors[edges[0]], pos[edges[1]]) ) is different from the paper.
The argument for pj seems to be diffs that is (q-b) according to the paper.
BTW, did this model achieve the same performance as the papers'? I have trained it for several epochs, but the speed is low and the outcome is undesirable.
Hi,
It seems line 28 of the gem_cnn.transform.py ( projected = torch.einsum('pij, pj -> pi', projectors[edges[0]], pos[edges[1]]) ) is different from the paper.
The argument for pj seems to be diffs that is (q-b) according to the paper.
BTW, did this model achieve the same performance as the papers'? I have trained it for several epochs, but the speed is low and the outcome is undesirable.