pyg_lib.ops

grouped_matmul(inputs: List[Tensor], others: List[Tensor], biases: Optional[List[Tensor]] = None) List[Tensor][source]

Performs dense-dense matrix multiplication according to groups, utilizing dedicated kernels that effectively parallelize over groups.

inputs = [torch.randn(5, 16), torch.randn(3, 32)]
others = [torch.randn(16, 32), torch.randn(32, 64)]

outs = pyg_lib.ops.grouped_matmul(inputs, others)
assert len(outs) == 2
assert outs[0].size() == (5, 32)
assert outs[0] == inputs[0] @ others[0]
assert outs[1].size() == (3, 64)
assert outs[1] == inputs[1] @ others[1]
Parameters:
  • inputs (List[Tensor]) – List of left operand 2D matrices of shapes [N_i, K_i].

  • others (List[Tensor]) – List of right operand 2D matrices of shapes [K_i, M_i].

  • biases (Optional[List[Tensor]], default: None) – Optional bias terms to apply for each element.

Returns:

List[Tensor] – List of 2D output matrices of shapes [N_i, M_i].

segment_matmul(inputs: Tensor, ptr: Tensor, other: Tensor, bias: Optional[Tensor] = None) Tensor[source]

Performs dense-dense matrix multiplication according to segments along the first dimension of inputs as given by ptr, utilizing dedicated kernels that effectively parallelize over groups.

inputs = torch.randn(8, 16)
ptr = torch.tensor([0, 5, 8])
other = torch.randn(2, 16, 32)

out = pyg_lib.ops.segment_matmul(inputs, ptr, other)
assert out.size() == (8, 32)
assert out[0:5] == inputs[0:5] @ other[0]
assert out[5:8] == inputs[5:8] @ other[1]
Parameters:
  • inputs (Tensor) – The left operand 2D matrix of shape [N, K].

  • ptr (Tensor) – Compressed vector of shape [B + 1], holding the boundaries of segments. For best performance, given as a CPU tensor.

  • other (Tensor) – The right operand 3D tensor of shape [B, K, M].

  • bias (Optional[Tensor], default: None) – The bias term of shape [B, M].

Returns:

Tensor – The 2D output matrix of shape [N, M].

sampled_add(left: Tensor, right: Tensor, left_index: Optional[Tensor] = None, right_index: Optional[Tensor] = None) Tensor[source]

Performs a sampled addition of left and right according to the indices specified in left_index and right_index.

\[\textrm{out} = \textrm{left}[\textrm{left_index}] + \textrm{right}[\textrm{right_index}]\]

This operation fuses the indexing and addition operation together, thus being more runtime and memory-efficient.

Parameters:
  • left (Tensor) – The left tensor.

  • right (Tensor) – The right tensor.

  • left_index (Optional[Tensor], default: None) – The values to sample from the left tensor.

  • right_index (Optional[Tensor], default: None) – The values to sample from the right tensor.

Returns:

Tensor – The output tensor.

sampled_sub(left: Tensor, right: Tensor, left_index: Optional[Tensor] = None, right_index: Optional[Tensor] = None) Tensor[source]

Performs a sampled subtraction of left by right according to the indices specified in left_index and right_index.

\[\textrm{out} = \textrm{left}[\textrm{left_index}] - \textrm{right}[\textrm{right_index}]\]

This operation fuses the indexing and subtraction operation together, thus being more runtime and memory-efficient.

Parameters:
  • left (Tensor) – The left tensor.

  • right (Tensor) – The right tensor.

  • left_index (Optional[Tensor], default: None) – The values to sample from the left tensor.

  • right_index (Optional[Tensor], default: None) – The values to sample from the right tensor.

Returns:

Tensor – The output tensor.

sampled_mul(left: Tensor, right: Tensor, left_index: Optional[Tensor] = None, right_index: Optional[Tensor] = None) Tensor[source]

Performs a sampled multiplication of left and right according to the indices specified in left_index and right_index.

\[\textrm{out} = \textrm{left}[\textrm{left_index}] * \textrm{right}[\textrm{right_index}]\]

This operation fuses the indexing and multiplication operation together, thus being more runtime and memory-efficient.

Parameters:
  • left (Tensor) – The left tensor.

  • right (Tensor) – The right tensor.

  • left_index (Optional[Tensor], default: None) – The values to sample from the left tensor.

  • right_index (Optional[Tensor], default: None) – The values to sample from the right tensor.

Returns:

Tensor – The output tensor.

sampled_div(left: Tensor, right: Tensor, left_index: Optional[Tensor] = None, right_index: Optional[Tensor] = None) Tensor[source]

Performs a sampled division of left by right according to the indices specified in left_index and right_index.

\[\textrm{out} = \textrm{left}[\textrm{left_index}] / \textrm{right}[\textrm{right_index}]\]

This operation fuses the indexing and division operation together, thus being more runtime and memory-efficient.

Parameters:
  • left (Tensor) – The left tensor.

  • right (Tensor) – The right tensor.

  • left_index (Optional[Tensor], default: None) – The values to sample from the left tensor.

  • right_index (Optional[Tensor], default: None) – The values to sample from the right tensor.

Returns:

Tensor – The output tensor.

index_sort(inputs: Tensor, max_value: Optional[int] = None) Tuple[Tensor, Tensor][source]

Sorts the elements of the inputs tensor in ascending order. It is expected that inputs is one-dimensional and that it only contains positive integer values. If max_value is given, it can be used by the underlying algorithm for better performance.

Note

This operation is optimized only for tensors associated with the CPU device.

Parameters:
  • inputs (Tensor) – A vector with positive integer values.

  • max_value (Optional[int], default: None) – The maximum value stored inside inputs. This value can be an estimation, but needs to be greater than or equal to the real maximum.

Returns:

Tuple[Tensor, Tensor] – A tuple containing sorted values and indices of the elements in the original input tensor.

softmax_csr(src: Tensor, ptr: Tensor, dim: int = 0) Tensor[source]

Computes a sparsely evaluated softmax. Given a value tensor src, this function first groups the values along the given dimension dim, based on the indices specified via ptr, and then proceeds to compute the softmax individually for each group.

Examples

>>> src = torch.randn(4, 4)
>>> ptr = torch.tensor([0, 4])
>>> softmax(src, ptr)
tensor([[0.0157, 0.0984, 0.1250, 0.4523],
        [0.1453, 0.2591, 0.5907, 0.2410],
        [0.0598, 0.2923, 0.1206, 0.0921],
        [0.7792, 0.3502, 0.1638, 0.2145]])
Parameters:
  • src (Tensor) – The source tensor.

  • ptr (Tensor) – Groups defined by CSR representation.

  • dim (int, default: 0) – The dimension in which to normalize.

Return type:

Tensor

spline_basis(pseudo: Tensor, kernel_size: Tensor, is_open_spline: Tensor, degree: int = 1) Tuple[Tensor, Tensor][source]

Computes the B-spline basis functions.

Parameters:
  • pseudo (Tensor) – Pseudo-coordinates of shape [E, D].

  • kernel_size (Tensor) – Kernel size in each dimension of shape [D].

  • is_open_spline (Tensor) – Whether to use open B-splines of shape [D].

  • degree (int, default: 1) – B-spline degree (1, 2, or 3).

Returns:

Tuple[Tensor, Tensor] – Basis values of shape [E, S] and weight indices of shape [E, S].

spline_weighting(x: Tensor, weight: Tensor, basis: Tensor, weight_index: Tensor) Tensor[source]

Computes the spline weighting of input features.

Parameters:
  • x (Tensor) – Input features of shape [E, M_in].

  • weight (Tensor) – Weight tensor of shape [K, M_in, M_out].

  • basis (Tensor) – B-spline basis values of shape [E, S].

  • weight_index (Tensor) – Weight indices of shape [E, S].

Returns:

Tensor – Output features of shape [E, M_out].

grid_cluster(pos: Tensor, size: Tensor, start: Optional[Tensor] = None, end: Optional[Tensor] = None) Tensor[source]

Clusters all points in pos into voxels of size size.

Each point is assigned a cluster index based on which voxel it falls into. The voxel grid is defined by the size parameter and optionally bounded by start and end.

Parameters:
  • pos (Tensor) – Point positions of shape [N, D].

  • size (Tensor) – Voxel size in each dimension of shape [D].

  • start (Optional[Tensor], default: None) – Start of the voxel grid in each dimension of shape [D]. If None, uses the minimum of pos.

  • end (Optional[Tensor], default: None) – End of the voxel grid in each dimension of shape [D]. If None, uses the maximum of pos.

Returns:

Tensor – Cluster index for each point of shape [N].

fps(src: Tensor, ptr: Tensor, ratio: float = 0.5, random_start: bool = True) Tensor[source]

Performs greedy farthest point sampling.

Starting from a random point (or the first point), iteratively selects the point that is farthest from the already selected set.

Parameters:
  • src (Tensor) – Point positions of shape [N, D].

  • ptr (Tensor) – Batch boundaries as a CSR pointer of shape [B + 1].

  • ratio (float, default: 0.5) – Fraction of points to sample from each batch (in (0, 1]).

  • random_start (bool, default: True) – If True, starts from a random point.

Returns:

Tensor – Indices of the sampled points of shape [M].

knn(x: Tensor, y: Tensor, k: int = 1, ptr_x: Optional[Tensor] = None, ptr_y: Optional[Tensor] = None, cosine: bool = False, num_workers: int = 1) Tensor[source]

Finds for each element in y the k nearest points in x.

Parameters:
  • x (Tensor) – Reference points of shape [N, D].

  • y (Tensor) – Query points of shape [M, D].

  • k (int, default: 1) – Number of nearest neighbors.

  • ptr_x (Optional[Tensor], default: None) – Batch boundaries for x as a CSR pointer.

  • ptr_y (Optional[Tensor], default: None) – Batch boundaries for y as a CSR pointer.

  • cosine (bool, default: False) – If True, uses cosine distance (CUDA only).

  • num_workers (int, default: 1) – Number of workers (unused, for API compat).

Returns:

Tensor – Edge indices of shape [2, M*k] where row 0 is query indices and row 1 is reference indices.

radius(x: Tensor, y: Tensor, r: float = 1.0, ptr_x: Optional[Tensor] = None, ptr_y: Optional[Tensor] = None, max_num_neighbors: int = 32, num_workers: int = 1, ignore_same_index: bool = False) Tensor[source]

Finds all points in x within distance r of points in y.

Parameters:
  • x (Tensor) – Reference points of shape [N, D].

  • y (Tensor) – Query points of shape [M, D].

  • r (float, default: 1.0) – Radius.

  • ptr_x (Optional[Tensor], default: None) – Batch boundaries for x as a CSR pointer.

  • ptr_y (Optional[Tensor], default: None) – Batch boundaries for y as a CSR pointer.

  • max_num_neighbors (int, default: 32) – Maximum number of neighbors per query point.

  • num_workers (int, default: 1) – Number of workers (unused, for API compat).

  • ignore_same_index (bool, default: False) – If True, ignores pairs with same index.

Returns:

Tensor – Edge indices of shape [2, E] where row 0 is query indices and row 1 is reference indices.

nearest(x: Tensor, y: Tensor, ptr_x: Optional[Tensor] = None, ptr_y: Optional[Tensor] = None) Tensor[source]

Finds the nearest point in y for each point in x.

Parameters:
  • x (Tensor) – Query points of shape [N, D].

  • y (Tensor) – Reference points of shape [M, D].

  • ptr_x (Optional[Tensor], default: None) – Batch boundaries for x as a CSR pointer.

  • ptr_y (Optional[Tensor], default: None) – Batch boundaries for y as a CSR pointer.

Returns:

Tensor – Index tensor of shape [N] with the index of the nearest point in y for each point in x.

graclus_cluster(rowptr: Tensor, col: Tensor, weight: Optional[Tensor] = None) Tensor[source]

Computes a greedy graph clustering via the Graclus algorithm.

Nodes are matched greedily in random order. The cluster ID for a matched pair (u, v) is min(u, v). Unmatched nodes are assigned their own index as cluster ID.

Parameters:
  • rowptr (Tensor) – CSR row pointer of shape [N + 1].

  • col (Tensor) – Column indices of shape [E].

  • weight (Optional[Tensor], default: None) – Optional edge weights of shape [E].

Returns:

Tensor – Cluster assignment of shape [N].

edge_sample(start: Tensor, rowptr: Tensor, count: int = 0, factor: float = 1.0) Tensor[source]

Samples edges incident to the given start nodes.

For each start node, samples up to count edges. If count < 1, samples ceil(factor * degree) edges instead.

Parameters:
  • start (Tensor) – Start node indices of shape [S].

  • rowptr (Tensor) – CSR row pointer of shape [N + 1].

  • count (int, default: 0) – Fixed number of edges to sample per node. If < 1, uses factor instead.

  • factor (float, default: 1.0) – Fraction of edges to sample when count < 1.

Returns:

Tensor – Sampled edge indices (into the edge list).