Data object in Pytorch geometric

[abhijeetdhakane]

Is it possible to store two graphs in the PyTorch geometrics Data object?
e.g.

Data(x=[B, 1100], edge_index=[B, 2, 4400], edge_attr=[B, 4400, 35], y=[B, 2])
where B=2 or B=1

In the official documentation, it has mentioned that the object should be as follows
Data(x=[1100], edge_index=[2, 4400], edge_attr=[4400, 35], y=[2]) as per documentation Data object is limited to hold/store single graph object.

A graph is used to model pairwise relations (edges) between objects (nodes). A single graph in PyG is described by an instance of torch_geometric.data.Data, which holds the following attributes by default:

(from https://pytorch-geometric.readthedocs.io/en/latest/notes/introduction.html#data-handling-of-graphs)
Can someone clarify, I was trying to create a custom Dataloader.

[Padarn]

You can use a DataLoader to batch graphs together, for example.

Graph =[graph1, graph2, graph3, ... , graph500]
loader = DataLoader(dataset = Graph , batch_size=10, shuffle=True)

When you sample from this loader you will get a single graph that has an extra attribute batch indicating which graph each node feature corresponds to.

Essentially this stores multiple graphs as a single graph that is not connected.

[rusty1s]

You can also look up the concept of PairData.

[abhijeetdhakane]

I used a torch.utils.data import Dataset and stored two graphs separately with torch_geometric.data import Data object like return data_a, data_b

If I use torch_geometric.data import DataLoader with batch_size=1 I'm getting the right shapes of Data attributes.
e.g. for data_a

x -> torch.Size([1100])
edge_index -> torch.Size([2, 4400])
edge_attr -> torch.Size([4400, 20])
y -> torch.Size([100])

Now, if I changed batch_size=10 I'll be getting

data_a
x -> torch.Size([11000])
edge_index -> torch.Size([2, 44000])
edge_attr -> torch.Size([44000, 20])
y -> torch.Size([1000])

I should get something like

data_a
x -> torch.Size([10,1100])
edge_index -> torch.Size([10, 2, 4400])
edge_attr -> torch.Size([10, 4400, 20])
y -> torch.Size([10,100])

Any advice?

[Padarn]

This is not how the batching works in the torch_geometric Dataloader, instead there is a new attribute of the batch index that in your case will look like

index = [0,0,0,....,1,1,1,...., 9,9,9]

where the value indicates which node that batch is from.

If you really want to do it this way you could define your own collate_fn something like

def collate_fn(datas: List[Data]):
    return Data(edge_index=torch.cat([data.edge_index for data in data]))

loader = torch.utils.data.DataLoader([graph1, graph2,...], collate_fn=collate_fn)

but torch_geometric doesn't have much support for working with edge_index of the shape you're suggesting.

[rusty1s]

Also take a look at the mini-batching tutorial.

[abhijeetdhakane]

Thanks @rusty1s and @Padarn

I used __inc__ to increment one of the data attribute which isn't taken care by pytorch_dataloader by default. Other data attributes x, edge_index, edge_attr taken care by pytorch_dataloader and get incremented as increased batch_size.

However, my batch_size value isn't accurate. i.e. batch attribute.

For
train_loader = DataLoader(Data_gen, batch_size=2, shuffle=True)

output should be
Batch(x=[2200], edge_index=[2, 8800], edge_attr=[8800, 20], y=[200], batch=[2], ptr=[3])

instead I am getting
Batch(x=[2200], edge_index=[2, 8800], edge_attr=[8800, 20], y=[200], batch=[2200], ptr=[3])

while num_nodes = 1100, looks like it just multiplied by batch * num_nodes

[abhijeetdhakane]

Also checked batch attribute which has tensor([0, 0, 0, ..., 1, 1, 1]),

As @Padarn mentioned values pointing out the nodes belongs to particular batch, but still I need to input batch_size value i.e. 2 here, to another function.

[rusty1s]

This looks correct to me. The batch attribute is a vector of shape [num_nodes] and maps each node to its example in the mini-batch. You can get the batch_size by running data.batch_size.