An error was encountered while creating the dataset

[yaoysyao]

I am currently doing text classification, using graph classification method. Now I need to construct data set conforming to PYG format according to data set. I refer to the TUDataset class to construct data set, because this data set is also used for graph classification. And validate(raise_on_error=True) method to prove my data is correct, I refer to the source code found that the TUDataset method calls the split(data, batch) method.
I would like to ask you the following questions:

1. What does this method do?
2. I also called this method, but found that edge_index was negative, resulting in the following error :RuntimeError: index -14102 is out of bounds for dimension 0 with size 8817, The problem occurs in data.edge_index -= node_slice[batch[row].unsqueeze(0)
3. The diagram constructed by TUDataset class is how to assemble nodes and graphs. I see that there are nodes in the data set belonging to a txt file of which graph, but how to judge the belonging of each node in the code? How do you determine which graph each node belongs to?

[rusty1s]

Hi, I am not sure if io/tu.py is absolutely necessary for you to look into as it contains parsing logic to convert the raw graph format defined in https://chrsmrrs.github.io/datasets/ to PyG format. Regarding your questions:
1/2: The graph format of TUDataset stores all graphs in a single file, so split takes care of splitting this dataset into individual components. As such, each edge_index part is decreased to start from index zero (this is the data.edge_index -= node_slice[batch[row].unsqueeze(0) line).
3: There exists a DS_graph_indicator.txt text file which contains this information.

[yaoysyao]

Thank you very much for your answer, I am currently working on text categorizing. I have looked at the examples in the source code, but I don't seem to find any examples of text categorizing, such as TextGcn and TextlevelGnn. However, due to the failure to correctly construct data format conforming to PYG, I encountered some problems. Is there any plan to add relevant code in the example, which will be very helpful to me

[rusty1s]

Can you do me a favor and create an issue for the models you would like to see in PyG? That would be very helpful.