src.build_graphs.reshape_edge_feature

src.build_graphs.reshape_edge_feature(F: torch.Tensor, G: torch.Tensor, H: torch.Tensor, device=None) → torch.Tensor

Given point-level features extracted from images, reshape it into edge feature matrix \(X\), where features are arranged by the order of \(G\), \(H\).

\[\mathbf{X}_{e_{ij}} = concat(\mathbf{F}_i, \mathbf{F}_j)\]

where \(e_{ij}\) means an edge connecting nodes \(i, j\)

Parameters

• F – \((b\times d \times n)\) extracted point-level feature matrix. \(b\): batch size. \(d\): feature dimension. \(n\): number of nodes.

• G – \((b\times n \times e)\) factorized adjacency matrix, where \(\mathbf A = \mathbf G \cdot \mathbf H^\top\). \(e\): number of edges.

• H – \((b\times n \times e)\) factorized adjacency matrix, where \(\mathbf A = \mathbf G \cdot \mathbf H^\top\)

• device – device. If not specified, it will be the same as the input

Returns

edge feature matrix X \((b \times 2d \times e)\)