What is the format of dataset of multiple graphs?

[lingxuan9]

Hi,

I am working on a graph level regression issue by following this Colab, but got block by the first step, prepare my own dataset.

The way I use: get a graph by using data = Data(x=x, edge_index=edge_index), and set dataset is a list of data. Is it correct or not? How can I build my own dataset?

Thanks in advance.

[EdisonLeeeee]

Yes. Here is a simple example to build a dataset of multiple graphs:

from torch_geometric.data import InMemoryDataset

class MyDataset(InMemoryDataset):
    def __init__(self, data_list):
        super().__init__('my_root_path')
        self.data, self.slices = self.collate(data_list)

where data_list is a list of Data objects.

[lingxuan9]

Thanks! I solved it. Forgive me I'm quite new for PyG.

[EdisonLeeeee]

Glad to see that :)

[lingxuan9]

Not want to open more issues but I indeed has some questions. Not sure if other people can see it.

I create a dataset with 1000 graphs, for each graph, it has 654 nodes, 16038 edges and a predict label: score. But when I call data.has_isolated_nodes(), I got a error of Segmentation fault: 11. This error is about
"This error means that the process tried to access an area of memory it did not alocated previously. This could be due to a bug in the software, a hacker trying to force a Denial of Service (DoS) on your web server, incorrect shared memory settings (SHM), or a faulty memory card, among other possible causes."

Seems there is some inappropriate data format? Can anyone meet this error before? Thanks!

[rusty1s]

What happens if you ran data.validate() on your data object. Can you confirm that edge_index only refers to valid input nodes?

[lingxuan9]

I got "True" when run data.validate(), yes I think edge_index is valid, it is the edge_index, I have 654 nodes in this graph.

[lingxuan9]

Is it possible because a poor server?

[rusty1s]

Mh, I am not sure. Would be good to know which op causes the segfault.

[lingxuan9]

Hi, there is a update, on a stronger server I didn't get this error.