Test on HGB

[shirinmous]

Hi.
I am working with HGBDataset.
There is no validation set for ACM dataset and the sum of train and test sets is 1966. The labels of test set are -1.
How can we test our algorithm using this dataset?

[rusty1s]

The test sets of the HGBDataset are hidden and one would need to upload the test predictions of your algorithm to the leaderboard, see here.

[EdisonLeeeee]

@rusty1s My bad. It seems that I made a wrong PR. Really sorry :(.

[shirinmous]

I downloaded ACM from HGB website. There are three files for labels
label.dat with 907 labels
label.dat.test with 1059 labels
label.dat.test_full with 2118 labels
So there are labels for all 3025 nodes.
I am confused. They provide labels but when we read the dataset by HGBDataset, we can't access the labels.

[shirinmous]

In HGBDataset code labels are loaded from label.dat.test but not saved in hetero graph.

I think the code has changed after that PR ,
'for y in test_ys:
n_id, n_type = mapping_dict[int(y[0])], n_types[int(y[2])]
if data[n_type].y.dim() > 1: # multi-label
for v in y[3].split(','):
data[n_type].y[n_id, int(v)] = 1
else:
data[n_type].y[n_id] = int(y[3])
data[n_type].test_mask[n_id] = True'
but when I install torch_geometric the old code is instlled. I don't know why?
for y in test_ys: n_id, n_type = mapping_dict[int(y[0])], n_types[int(y[2])] data[n_type].test_mask[n_id] = True

[rusty1s]

You might need to install PyG from master to have access to the PR.
@EdisonLeeeee Shall we revert the PR to avoid confusion?

[EdisonLeeeee]

@rusty1s Sure.