Simplicial Attention Network (SAN) implementation for binary edge classification.
- class topomodelx.nn.simplicial.san.SAN(in_channels, hidden_channels, out_channels=None, n_filters=2, order_harmonic=5, epsilon_harmonic=0.1, n_layers=2)[source]#
Simplicial Attention Network (SAN) implementation for binary edge classification.
- Parameters:
- in_channelsint
Dimension of input features.
- hidden_channelsint
Dimension of hidden features.
- out_channelsint
Dimension of output features.
- n_filtersint, default = 2
Approximation order for simplicial filters.
- order_harmonicint, default = 5
Approximation order for harmonic convolution.
- epsilon_harmonicfloat, default = 1e-1
Epsilon value for harmonic convolution.
- n_layersint, default = 2
Number of message passing layers.
- compute_projection_matrix(laplacian)[source]#
Compute the projection matrix.
The matrix is used to calculate the harmonic component in SAN layers.
- Parameters:
- laplaciantorch.Tensor, shape = (n_edges, n_edges)
Hodge laplacian of rank 1.
- Returns:
- torch.Tensor, shape = (n_edges, n_edges)
Projection matrix.
- forward(x, laplacian_up, laplacian_down)[source]#
Forward computation.
- Parameters:
- xtorch.Tensor, shape = (n_nodes, channels_in)
Node features.
- laplacian_uptorch.Tensor, shape = (n_edges, n_edges)
Upper laplacian of rank 1.
- laplacian_downtorch.Tensor, shape = (n_edges, n_edges)
Down laplacian of rank 1.
- Returns:
- torch.Tensor, shape = (n_edges, out_channels)
Final hidden representations of edges.