the impact of multiple edges

[xyqbest]

✋ Background：
I run a GNN on a graph with more than 350 nodes. Each graph has 64 historical frames. The length of edge_index in each frame is different (approximately 8 000 ~ 20 000 edges ). The output is node value in the next frame.
My GNN can deal with each frame one by one under batch_size=1. It seems that there are 64 batches for GNN model.
If my GNN deal with batch_size=40, then the input will torch.cat the k-th frame over all 40 batches. As a result, the edge_index will very long, almost 512 000 ~ 1 280 000 edges. They are lots of repeated edges. It seems that there are 40*64 batches for GNN model.

❓ My question

1. Should I torch.unique them before I put them into GNN? I mean removing all repeated edge_index.
2. What is the difference between multiple edges and simplified edges? I also run a test, code is in below. I found the output is different. It means that I should not remove multiple edges?

import os
import random

import numpy as np
import torch
from torch_geometric.nn import GCNConv


def seed_everything(seed=666):
    # for reproducibility
    random.seed(seed)
    os.environ['PYTHONHASHSEED'] = str(seed)
    np.random.seed(seed)
    torch.manual_seed(seed)
    torch.cuda.manual_seed(seed)
    torch.backends.cudnn.deterministic = True


if __name__ == '__main__':
    seed_everything(seed=666)
    x = torch.rand(10, 7)

    tmp = torch.randint(0, 10, (30, 2))
    # tmp = torch.cat([tmp, tmp[:8]]) # comment this line in the fist run. uncomment this line in the second run to create multiple edges.

    edge_index = torch.LongTensor(tmp.T)
    tmp_layer = GCNConv(7, 4)

    out = tmp_layer(x, edge_index)
    print(out)

Thank you 😄

[rusty1s]

How do you encode the different edges across multiple frames? I don't think there is any opportunity to call torch.unique() here since values in edge_index are typically not duplicated (or at least they shouldn't be): Each frame should be saved in its own subgraph. If you are running OOM when processing 64 frames across 40 graphs (batch_size=40), why not simply reduce the batch size?

If you have duplicated edges, that means that you will receive multiple messages coming from the same neighbor. As such, you give a "higher weight" to this kind of neighbor.

[xyqbest]

If you have duplicated edges, that means that you will receive multiple messages coming from the same neighbor. As such, you give a "higher weight" to this kind of neighbor.

This is what I want.

By the way,
When I use the loader with batch size. The data will torch.cat edge_index . For example,
[Data(..., edge_index=[2,20],...),Data(..., edge_index=[2,22],...),Data(..., edge_index=[2,25],...)]
with the batch_size = 3 , the loader will load data Data(..., edge_index=[2,67],...) and put it into the model. There are some duplicated edges among edge_index=[2,67].

Does it impact the output of model? make some neighbors "higher weight"?
Does it lead to the different results compared to batch_size=1?

[rusty1s]

For mini-batching, there will not be any overlap between edges across different examples. This is described here. Importantly, batching will not result in different results compared to batch_size=1.