Get single Node feature instead of global pooling

[oliverluo1993]

Hi,

I am trying to predict the node level instead of graph level properties.

Except from global pooling layers to get graph level information. Is there a way to fetch a single node feature from batched node feature matrix?

Best regards

[rusty1s]

You can use traditional indexing/index_select to select specific node feature vectors, e.g.

node_idx = torch.tensor([0, 10, 15])  # Select features of node 0, 10 and 15
out = x[node_idx]

Does that answer your question?

[oliverluo1993]

Thanks!

But in this way, it only works with batch size = 1.

Is there a way to work with mini-batching of pytorch geometric?

[rusty1s]

If you name your tensor to include the word index, mini-batching should increase indices out-of-the-box:

data.node_index = ...

[oliverluo1993]

Thanks! I will try that!