pylibcugraphops.operators.agg_concat_n2n_e2n_bwd#

pylibcugraphops.operators.agg_concat_n2n_e2n_bwd = <nanobind.nb_func object>#

Computes the backward pass for both a simple aggregation (agg_simple) using node features in an: node-to-node reduction (n2n) and a simple aggregation using edge features in an edge-to-node reduction (e2n) on a full graph where the results are concatenated. Moreover, the original features of output nodes are concatenated at the end (agg_concat).

agg_simple_n2n_e2n_bwd(
    grad_input_node: device array, grad_input_edge: device array,
    grad_output: device array, graph: pylibcugraphops.csc_int[32|64],
    dim_node: int, dim_edge: int,
    aggregation_operation: pylibcugraphops.operators.AggOp = pylibcugraphops.operators.AggOp.Sum,
    output_extrema_location: Optional[device array] = None, stream_id: int = 0
) -> None

Parameters:

grad_input_nodedevice array type: Device array containing the output gradient on input node embeddings of forward. Shape: (graph.n_dst_nodes, dim_node).
grad_input_edgedevice array type: Device array containing the output gradient on input edge embeddings of forward. Shape: (n_edges, dim_edge).
grad_outputdevice array type: Device array containing the input gradient on output embeddings of forward. Shape: (graph.n_dst_nodes, dim_node + dim_edge + dim_node).
graphopaque graph type: The graph used for the operation.
dim_nodeint: Node feature dimensionality.
dim_edgeint: Edge Feature dimensionality.
aggregation_operationAggOp, default=AggOp.Sum: The kind of aggregation operation.
output_extrema_locationdevice array type | None: Device array containing the location of the min/max embeddings. This is required for min/max aggregation only, and can be None otherwise. Shape: (graph.n_dst_nodes, dim_node + dim_edge) if set.
stream_idint, default=0: CUDA stream pointer as a python int.