QM9EdgeDataset labels are wrong

🐛 Bug

With QM9EdgeDataset, it seems that the prediction labels are broken. This may be because of the preprocessing, or because of a bad source for QM9.

@hengruizhang98 @mufeili

To Reproduce

from dgl.data import QM9EdgeDataset as DGLQM9Edge
from dgl.data import QM9Dataset as DGLQM9
import matplotlib.pyplot as plt

keys = ['mu', 'alpha', 'homo', 'lumo', 'gap', 'r2', 'zpve']
f, axs = plt.subplots(2, len(keys), figsize=(20, 5))
for i, task in enumerate(keys):
    ds_dgl = DGLQM9Edge([task])
    ds_dgl2 = DGLQM9([task])

    targets_dgl = ds_dgl.targets[:,i]
    targets_dgl2 =ds_dgl2.label[:,0]
    
    axs[0][i].hist(targets_dgl2, bins=50)
    axs[1][i].hist(targets_dgl, bins=50)

f.tight_layout()
plt.show()

The first row is the histograms of labels from QM9Dataset, and the second row is the ones from QM9EdgeDataset.

Expected behavior

Labels should be the same for all QM9 datasets.

Environment

• DGL Version (e.g., 1.0): commit 1f4c0b7
• Backend Library & Version (e.g., PyTorch 0.4.1, MXNet/Gluon 1.3): PyTorch 1.9.0
• OS (e.g., Linux): Linux
• How you installed DGL (conda, pip, source): source
• Build command you used (if compiling from source): cmake -DUSE_CUDA=ON -DUSE_FP16=ON … && make -j8
• Python version: 3.8.8

Additional context

In the docs it says that the preprocessing is done here https://gist.github.com/hengruizhang98/a2da30213b2356fff18b25385c9d3cd2 so there must be something wrong there.

[hengruizhang98]

Yes, It will take about 1 min to load graphs from QM9v2. While using QM9Edge is much faster as the graphs are constructed when called. You can choose the way you prefer.

[milesial]

I see that QM9V2 is directly loading DGL graphs with load_graphs, and QM9Edge is creating them on the fly. Maybe one is faster than the other.