I am trying to convert my PyG graph to a NetworkX graph using to_networkx
According to the docs I can optionally pass node and edge attributes as str iterables, in addition to the Data object.
Below are by node and edge attribute lists, with values converted to strings:
Nodes: ['3.3375725746154785', '2.0086510181427',..., '1.5960148572921753', '3.621992349624634']
Edges: ['0.9940207804344958', '0.48573804411542043', ..., '0.7245483440145621', '0.24117984598949904']
to_networkx runs fine when I only pass it the Data object. However, when I also pass these attribute lists, I get the below error:
G[u][v][key] = values[key][i]
KeyError: '0.30194718370332896'
I've looked at the source code, but can't make out what it is doing. Could someone please help explain what is wrong with my attribute lists and what I need to change for them to be accepted.
What I can make out is that this error is specifically referring to my edge attributes. If I remove them, I get the following similar error related to the node attributes:
feat_dict.update({key: values[key][i]})
KeyError: '0.0'
How I construct my graph and pass it to to_networkx:
n1 = np.repeat(np.array([0,1,2,3,4,5,6]),5)
n2 = np.array([0,1,2,3,4,0,1,2,3,4,0,1,2,3,4,0,1,2,3,4,0,1,2,3,4,0,1,2,3,4,0,1,2,3,4])
cat = np.stack((n1,n2), axis=0)
e = torch.tensor(cat, dtype=torch.long)
edge_index = e.t().clone().detach()
edge_attr = torch.tensor(np.random.rand(35,1))
x = torch.tensor([[0], [0], [0], [0], [0], [1], [1]], dtype=torch.float)
data = Data(x=x, edge_index=edge_index.t().contiguous(), edge_attr = edge_attr)
Before I pass the node and edge attributes, I do the string conversion to conform to the str iterable requirment:
networkx_node_values = list(map(str, data.x.t()[0].tolist()))
networkx_edge_values = list(map(str, edge_attr.t()[0].tolist()))
networkX_graph = to_networkx(data, node_attrs = networkx_node_values, edge_attrs = networkx_edge_values)
You need to pass the names of the attributes as a list:
to_networkx(<PyTorchGeometricDataObject>, node_attrs=[<Name of Node Attribute 1>, <Name of Node Attributes 2>, ... ], edge_attr=[<Edge Attribute 1>, ...])
Or in context, based on your given minimal example:
import numpy as np
import torch
from torch_geometric.data import Data
from torch_geometric.utils import to_networkx
n1 = np.repeat(np.array([0,1,2,3,4,5,6]),5)
n2 = np.array([0,1,2,3,4,0,1,2,3,4,0,1,2,3,4,0,1,2,3,4,0,1,2,3,4,0,1,2,3,4,0,1,2,3,4])
cat = np.stack((n1,n2), axis=0)
e = torch.tensor(cat, dtype=torch.long)
edge_index = e.t().clone().detach()
edge_attr = torch.tensor(np.random.rand(35,1))
x = torch.tensor([[0], [0], [0], [0], [0], [1], [1]], dtype=torch.float)
data = Data(x=x, edge_index=edge_index.t().contiguous(), edge_attr = edge_attr)
print(data)
# Data(edge_attr=[35, 1], edge_index=[2, 35], x=[7, 1])
networkX_graph = to_networkx(data, node_attrs=["x"], edge_attrs=["edge_attr"])
print(networkX_graph.nodes(data=True))
# [(0, {'x': 0.0}), (1, {'x': 0.0}),...
print(networkX_graph.edges(data=True))
# [(0, 0, {'edge_attr': 0.3412137594357493}), ...