Creating custom dataset from list of networkx graphs

[AdarshMJ]

Hi! So I am working on converting a list of networkx graphs to PyG graphs to train a graph classification framework. I observed that when I convert PyG ---> NetworkX ---> PyG again the labels are not there. If I use

`from torch_geometric.data import Data
from torch_geometric.loader import DataLoader

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

I get an error while training AttributeError: 'list' object has no attribute 'num_node_features'

How do I convert the dataset appropriately? Any insights are welcome. Thanks again!

[rusty1s]

You might need to specify the node_attrs and edge_attrs arguments in the to_networkx call.

[AdarshMJ]

Hi! So I am trying to run graph classification on the QM9 dataset, the linear layer uses -

self.lin = Linear(hidden_channels, dataset.num_classes)
The batch size is set to 64. The hidden_channels = 64 and num_classes=19.

While calculating the loss, I get an error -
RuntimeError: The size of tensor a (64) must match the size of tensor b (19) at non-singleton dimension 1

Any idea how to resolve this?

[rusty1s]

Are you sure you apply self.lin during the forward pass of the model?

[AdarshMJ]

Hi Yes. Im following this PyG Graph classification tutorial - https://colab.research.google.com/drive/1I8a0DfQ3fI7Njc62__mVXUlcAleUclnb?usp=sharing

I have only changed the dataset to QM9 from the default Proteins.

[rusty1s]

I see. Please note that QM9 is a regression dataset, and so you will need to make use of torch.MSELoss rather than CrossEntropyLoss for training.

[AdarshMJ]

Okay that makes sense. I will make the change. Thank you!