minimalistic_horizontal_edge_features
Closed this issue · 1 comments
MenglinQiu commented
Hi, bro. Is there something wrong with the _minimalistic_horizontal_edge_features function in the src/transforms/graph.py?
base = base_vectors_3d(se_mean_off)[se_id]
u = (offset * base[:, 0]).sum(dim=1).view(-1, 1)
v = (offset * base[:, 0]).sum(dim=1).view(-1, 1)
w = (offset * base[:, 0]).sum(dim=1).view(-1, 1)
se_std_off = scatter_std(torch.cat((u, v, w), dim=1), se_id, dim=0)
The results of u, v, and w will be exactly equal.
drprojects commented
Ooh nice catch ! You are perfectly right, there is indeed something fishy there, I just fixed it in the latest commit. This error was computing incorrect 'std_off'
egde features. This might have a small positive impact on model performance but I do not expect major changes.
Thanks for finding this error and thanks for the support you have been providing in issue #134 😉