torch_geometric.nn.aggr.SortAggregation

class SortAggregation(k: int)[source]

Bases: Aggregation

The pooling operator from the “An End-to-End Deep Learning Architecture for Graph Classification” paper, where node features are sorted in descending order based on their last feature channel. The first \(k\) nodes form the output of the layer.

Note

SortAggregation requires sorted indices index as input. Specifically, if you use this aggregation as part of MessagePassing, ensure that edge_index is sorted by destination nodes, either by manually sorting edge indices via sort_edge_index() or by calling torch_geometric.data.Data.sort().

Parameters:

k (int) – The number of nodes to hold for each graph.

forward(x: Tensor, index: Optional[Tensor] = None, ptr: Optional[Tensor] = None, dim_size: Optional[int] = None, dim: int = -2, max_num_elements: Optional[int] = None) → Tensor[source]

Forward pass.

Parameters:

  • x (torch.Tensor) – The source tensor.

  • index (torch.Tensor, optional) – The indices of elements for applying the aggregation. One of index or ptr must be defined. (default: None)

  • ptr (torch.Tensor, optional) – If given, computes the aggregation based on sorted inputs in CSR representation. One of index or ptr must be defined. (default: None)

  • dim_size (int, optional) – The size of the output tensor at dimension dim after aggregation. (default: None)

  • dim (int, optional) – The dimension in which to aggregate. (default: -2)

  • max_num_elements (Optional[int], default: None) – (int, optional): The maximum number of elements within a single aggregation group. (default: None)

Return type:

Tensor