How is feature matrix mapped to nodes?

[SaffronWolf]

Sorry for the silly question, but when initializing Data object, how PyG knows what row in feature matrix belongs to which node?

import torch
from torch_geometric.data import Data

edge_index = torch.tensor([[0, 1],
                           [1, 0],
                           [1, 2],
                           [2, 1]], dtype=torch.long)
x = torch.tensor([[-1], [0], [1]], dtype=torch.float)

data = Data(x=x, edge_index=edge_index.t().contiguous())

In the above example, how do I ensure that.

[-1] is mapped to node 0
[0] is mapped to node 1
[1] is mapped to node 2

[rusty1s]

The node indices correspond to the row index of the feature matrix. That is, x[0] denotes the features of node 0, x[1] denotes the features of node 1, and so on. As such, node indices in PyG need to be mapped to the 0 to num_nodes - 1 range. Let me know if this resolves your issue :)

[SaffronWolf]

Thanks a lot. Just curious, what happens if node indices are not mapped to 0 to num_nodes -1. How would feature matrix be mapped to nodes then?

[rusty1s]

Interesting question, although not sure I fully understand it. In case you allow arbitrary node indices, you would need to store feature vectors in a hash map rather than in a single feature matrix, which complicates memory access patterns and slows down neural network execution.