spamosaic.architectures.wlgcn

Weighted Light graph convolution (WLGCN) layers for SpaMosaic.

Implements a parameter-free propagation layer on pre-normalized sparse adjacencies and an MLP head for embedding + reconstruction.

class spamosaic.architectures.wlgcn.HEAD(*args: Any, **kwargs: Any)

Bases: Module

MLP head that outputs an L2-normalized embedding and reconstructs input features.

Parameters:

• input_size (int) – Input feature dimension.

• output_size (int) – Embedding dimension.

• dec_l (int, default=1) – Number of decoder layers (1 = linear decoder).

• hidden_size (int, default=512) – Hidden size of the MLP trunk.

• dropout (float, default=0.2) – Dropout rate.

• slope (float, default=0.2) – Negative slope for LeakyReLU.

forward(feature)

Forward pass.

Parameters:

feature (torch.Tensor) – Input features of shape [N, input_size].

Returns:

• z: L2-normalized embedding [N, output_size].

• r: Reconstructed features [N, input_size] from pre-norm representation.

Return type:

tuple[torch.Tensor, torch.Tensor]

class spamosaic.architectures.wlgcn.WLGCN_vanilla(*args: Any, **kwargs: Any)

Bases: Module

Parameter-free LightGCN-style propagation.

Accepts a precomputed normalized sparse adjacency \(\hat{A} \approx D^{-1/2}(A + I\cdot\text{fill})D^{-1/2}\) as a torch_sparse.SparseTensor.

Parameters:

• K (int, default=1) – Number of propagation steps (returns K+1 representations including the 0-th order input).

• agg ({'cat', 'sum', 'mean'}, default='cat') – How to aggregate representations across steps.

forward(x: torch.Tensor, A_hat: torch_sparse.SparseTensor) → torch.Tensor

Propagate features via A_hat for K steps and aggregate.

Parameters:

• x (torch.Tensor) – Node features of shape [N, F].

• A_hat (torch_sparse.SparseTensor) – Pre-normalized sparse adjacency on the correct device.

Returns:

Aggregated output. Shapes depend on agg: - 'cat' → [N, F * (K+1)] - 'sum' → [N, F] - 'mean' → [N, F]

Return type:

torch.Tensor

Classes

spamosaic.architectures.wlgcn.HEAD	MLP head that outputs an L2-normalized embedding and reconstructs input features.
spamosaic.architectures.wlgcn.WLGCN_vanilla	Parameter-free LightGCN-style propagation.