about graph conv

[chenbys]

An interesting work, but I have some doubts about the graph conv.

The adj obtained at

DVBE/main.py

Line 133 in dd87981

adj = adj_matrix(nc)

seems a class_num*class_num identity matrix;
and, the adj will torch.mm with the FC features at

DVBE/models/operations.py

Line 211 in dd87981

outputs = torch.mm(self.adj,outputs)

;
so, what's the meaning of adj, and how do their dims match?

[mboboGO]

The graph operation is not used in this code version, thus the adj matrix generation code is incorrect. I will soon fix this issue in few days, and the correct adj matrix should not be an identity matrix.

In the correct version, the adj matix is ncnc dimensional, and outputs is ncC dimensional, where C is the feature channel.
Thus, their dims are matched for torch.mm(adj,output).