torch_geometric.nn.pool.ApproxL2KNNIndex

class ApproxL2KNNIndex(num_cells: int, num_cells_to_visit: int, bits_per_vector: int, emb: Optional[Tensor] = None, reserve: Optional[int] = None)[source]

Bases: KNNIndex

Performs fast approximate \(k\)-nearest neighbor search (\(k\)-NN) based on the the \(L_2\) metric via the faiss library. Hyperparameters needs to be tuned for speed-accuracy trade-off.

Parameters:

  • num_cells (int) – The number of cells.

  • num_cells_to_visit (int) – The number of cells that are visited to perform to search.

  • bits_per_vector (int) – The number of bits per sub-vector.

  • emb (torch.Tensor, optional) – The data points to add. (default: None)

  • reserve (int, optional) – The number of elements to reserve memory for before re-allocating (GPU only). (default: None)