Pytorch_geometric: GEDDataset num_classes request leads to AttributeError

[TNO-Knowledge-Based-Systems]

🐛 Bug

When loaded a GEDDataset and querying the num_classes as descriped in the tutorial (dataset.num_classes) an AttributeError: 'NoneType' object has no attribute 'dim'-error is thrown.

To Reproduce

Steps to reproduce the behavior:

1. Load the IMDBMulti dataset with GEDDataset
2. Ask the num_classes.

Code for reproducing error:

from torch_geometric.datasets import GEDDataset
dataset = GEDDataset(root='/datasets/IMDBMulti', name="IMDBMulti")
print(dataset.num_classes)

Expected behavior

I would have expected a return of the number of classes: 3.

Environment

• OS: Ubuntu 18.04 in Docker
• Python version: 3.6.5
• PyTorch version: 1.4.0
• CUDA/cuDNN version: 10.1.243 / 7.6.5.32
• GCC version: 7.5.0
• Any other relevant information: Basis of our Dockerfile is from your provided dockerfile.

Additional context

I noticed that in the examples, the ground truth is saved in the key 'y'.

data = dataset[0]
data.keys()

would return something like ['edge_index', 'y']

The GEDDataset returns
['edge_index', 'i'].

Also tested with the ALKANE dataset, same problem.

[rusty1s]

The GEDDataset do not have any node labels (y=None), so a call to num_classes is expected to crash. Instead, the GEDDataset contains the GED distance between all pairs of graphs in the dataset.

[TNO-Knowledge-Based-Systems]

Thank you for your reply.
What is stored in 'i', when it is not the node label?

The IMDB-Binary, which is availble in TUDataset, seems to have node labels. In that dataset it is possible to query the num_classes. Why is this different for the IMDB-Multi?

[rusty1s]

The i refers to the index of the graph in the dataset, so that you can access the GED between two pairs of graphs via dataset.ged[data1.i, data2.i]. I think we should mention this in the documentation.

[andac-demir]

Is there a way we can infer the graph labels from ged measures?

[rusty1s]

There are no graph labels available for this dataset. This dataset should be only used in case you are working on accelerating GED computation with machine learning.

[tobelegit]

SimGNN, paper referred in dataset , seems to have labels for Aids dataset.

Aids Dataset : Each node is labeled
with one of 29 types, as illustrated in Fig. 4a.

Linux Data : We randomly select 1000
graphs of equal or less than 10 nodes each. The nodes are unla-
beled.

IMDB Data : To test the scal-
ability and efficiency of our proposed approach, we use the full
dataset without any selection. The nodes are unlabeled.

How to access aids dataset labels?

[rusty1s]

You can access them via data.x.