rllm.nn.models.RECT_L¶

class rllm.nn.models.RECT_L(in_dim: int, hidden_dim: int, dropout: float = 0.0)[source]¶

Bases: Module

The RECT model, or more specifically its supervised part RECT-L, from the “Network Embedding with Completely-imbalanced Labels” paper. In particular, a GCN model is trained that reconstructs semantic class knowledge.

Parameters:

in_dim (int) – Size of each input sample.
hidden_dim (int) – Intermediate size of each sample.
dropout (float, optional) – The dropout probability. (default: 0.0)

Example

>>> import torch
>>> model = RECT_L(in_dim=16, hidden_dim=8)
>>> x = torch.randn(4, 16)
>>> adj = torch.tensor([[0, 1, 2], [1, 2, 3]])
>>> model(x, adj).shape
torch.Size([4, 16])

embed(x: Tensor, adj: Tensor)[source]¶

Compute hidden embeddings without gradient updates.

Parameters:

x (Tensor) – Input node features.
adj (Tensor) – Graph connectivity.

Returns:

Hidden node embeddings.

Return type:

Tensor

forward(x: Tensor, adj: Tensor)[source]¶

Encode node features and reconstruct semantic targets.

Parameters:

x (Tensor) – Input node features.
adj (Tensor) – Graph connectivity.

Returns:

Reconstructed output features.

Return type:

Tensor

get_semantic_labels(x: Tensor, y: Tensor, mask: Tensor)[source]¶

Replace labels with corresponding class-center embeddings.

Parameters:

x (Tensor) – Node embeddings.
y (Tensor) – Class labels.
mask (Tensor) – Mask selecting labeled nodes.

Returns:

Class-center embeddings for labeled nodes.

Return type:

Tensor

reset_parameters()[source]¶: Resets all learnable parameters of the module.