Can I put the edges' weights into the edge_attr?

❓ Questions & Help

Hi, first of all, thank you so much for this wonderful tool for PyTorch. I am a newbie in this library so I have two very strange questions.

---

1. edge_attr

I notice that the Data class has an attribution named edge_attr which holds the edges’ features in the graph. And currently, I need to construct a weighted directed graph Data instance for GCNConv. So here is my question: Can I put the weight into the edge_attr for creating the Data instance? In my view, I think it is just okay, but seems so strange. Because in this time the shape of the edge_attr tensor is actually [num_edges, 1].

So, is it okay to do it?

---

2. Dynamical batch

I also want to confirm one thing. Here FAQ2, your introduction seems to show that I can dynamically construct a graph DataLoader in this way:

from torch_geometric.data import Data, DataLoader

data_list = [Data(...), ..., Data(...)]
loader = DataLoader(data_list, batch_size=32)

Is that all I need to do? Don’t I need to do some stuff like padding for the batches?

Finally, thank you so much for your work and hope you to respond to my questions.

[dabao0903]

You can save your edge weights into edge_attr. This is no problem. Alternatively, you can save them to any attribute, e.g., data.edge_weight = edge_weight. The DataLoader correctly merges all attributes into a single graph. So this is indeed all you need to do.

Hello, author. Can you explain the principle of “Weighted GCN layer”, thanks!

[rusty1s]

Please have a look at the GATConv layer in torch_geometric.nn.