TGraph

class TGraph

Bases: Graph

Wrapper class for pytorch-geometric edge_index providing fast adjacency look-up.

__init__(*args, ensure_sorted=False, **kwargs)

Initialise graph

Parameters:

• edge_index – edge index such that edge_index[0] lists the source and edge_index[1] the target node for each edge

• edge_attr – optionally provide edge weights

• num_nodes – specify number of nodes (default: max(edge_index)+1)

• ensure_sorted – if False, assume that the edge_index input is already sorted

• undir – boolean indicating if graph is directed. If not provided, the edge_index is checked to determine this value.

Methods

__init__	Initialise graph
adj	list neighbours of node
adj_weighted	list neighbours of node and corresponding edge weight :param node: source node
bfs_order	return nodes in breadth-first-search order
connected_component_ids	Find the (weakly)-connected components.
edges	return list of edges where each edge is a tuple (source, target)
edges_weighted	return list of edges where each edge is a tuple (source, target, weight)
from_networkx
from_tg
has_node_labels
is_edge
lcc
neighbourhood	find the neighbourhood of a set of source nodes
nodes_in_lcc	List all nodes in the largest connected component
partition_graph
sample_negative_edges
sample_positive_edges
subgraph	find induced subgraph for a set of nodes
to	Convert to different graph type or move to device
to_networkx	convert graph to NetworkX format

Attributes

device	device holding graph data
nodes
num_edges
num_features
weighted	boolean indicating if graph is weighted
num_nodes	number of nodes
degree	tensor of node degrees
adj_index	adjacency index such that edges starting at node i are given by edge_index[:, adj_index[i]:adj_index[i+1]]
strength	tensor of node strength
weights

edges()

return list of edges where each edge is a tuple (source, target)

edges_weighted()

return list of edges where each edge is a tuple (source, target, weight)

is_edge(source, target)

neighbourhood(nodes: Tensor, hops: int = 1)

find the neighbourhood of a set of source nodes

note that the neighbourhood includes the source nodes themselves

Parameters:

• nodes – indices of source nodes

• hops – number of hops for neighbourhood

Returns:

neighbourhood

subgraph(nodes: Tensor, relabel=False, keep_x=True, keep_y=True)

find induced subgraph for a set of nodes

Parameters:

nodes – node indeces

Returns:

subgraph

connected_component_ids()

Find the (weakly)-connected components. Component ids are sorted by size, such that id=0 corresponds to the largest connected component

nodes_in_lcc()

List all nodes in the largest connected component

to_networkx()

convert graph to NetworkX format

to(*args, graph_cls=None, **kwargs)

Convert to different graph type or move to device

Parameters:

• graph_cls – convert to graph class

• device – convert to device

Can only specify one argument. If positional, type of move is determined automatically.

bfs_order(start=0)

return nodes in breadth-first-search order

Parameters:

start – index of starting node (default: 0)

Returns:

tensor of node indeces

partition_graph(partition, self_loops=True)

sample_negative_edges(num_samples)

sample_positive_edges(num_samples)