Pytorch_geometric: Classes in the QM9 dataset

[elissaross]

❓ Questions & Help

What happened to the QM9 dataset? Previously there were multiple regression targets. Now it seems that y is a 1-D tensor, and when I query num_classes it reports "30.5564...".
The qm9_nn_conv example is also no longer functional as a result of these changes. Any insights appreciated!

[rusty1s]

I can not reproduce this. Which PyG version are you using? It is working in current master branch:

Batch(batch=[230], edge_attr=[400, 4], edge_index=[2, 400], pos=[230, 3], x=[230, 13], y=[32, 16])

[elissaross]

Curious. I am on 1.3.2

[elissaross]

Would you be able to give me the code you used to generate the above? Thank you!

[rusty1s]

Sure

from torch_geometric.datasets import QM9
from torch_geometric.data import DataLoader

dataset = QM9('/tmp/QM9')
print(dataset)
print(dataset[0])
loader = DataLoader(dataset, batch_size=32)

for data in loader:
    print(data)
    break

[elissaross]

Thank you -- that is very helpful. I was loading a slightly different dataset (?) as follows:
dataset = QM9('/tmp/molecules')
which was working fine for a while, but has somehow changed. I will use the dataset you are using above. This fixes my issue -- thanks again.

[rusty1s]

Great!

[muhanzhang]

I am facing the same issue when loading the preprocessed QM9 from 'http://www.roemisch-drei.de/qm9.zip'. When I first load it and print dataset.data, I got

Data(edge_attr=[4823498, 5], edge_index=[2, 4823498], name=[129433], pos=[2333625, 3], x=[2333625, 13], y=[129433])

showing y has one single dimension. Then I completely delete the dataset cache and re-download QM9 (also from http://www.roemisch-drei.de/qm9.zip), I print dataset.data again, and surprisingly got

Data(edge_attr=[4823498, 5], edge_index=[2, 4823498], name=[129433], pos=[2333625, 3], x=[2333625, 13], y=[129433, 19])

showing y has 19 tasks. Can you please have a check? Thanks.

[muhanzhang]

Another issue is, the dataset shows edge_attr should have 5 dimensions. But when I print dataset[0], I got

Data(edge_attr=[8, 6], edge_index=[2, 8], name=[1], pos=[5, 3], x=[5, 13], y=[1, 19])

showing it has 6 dimensions. This makes 1-qm9.py in k-gnn fail to work.

I am using python 3.8.1, pytorch 1.4.0 and pyg 1.4.2

[rusty1s]

There was a small bug in QM9 where the data was not correctly loaded when using transforms. It should have been fixed in PyG 1.4.3.

[rusty1s]

Oh sorry, it is only fixed on master.