Division by zero error for node classification

[AdarshMJ]

Hi @rusty1s

I am trying to run a simple GCN on Cornell dataset(WebKB). My model looks like this -

GCN(
  (conv1): GCNConv(1703, 32)
  (conv2): GCNConv(32, 5)
)

The dataset looks like this originally -

Data(x=[183, 1703], edge_index=[2, 298], y=[183], train_mask=[183, 10], val_mask=[183, 10], test_mask=[183, 10])

I do some modifications by converting it to NetworkX graph and then convert it back to PyG graph, to ensure I get the masks back I apply the random node split transform -

Newdatapyg = from_networkx(gcornell)
from torch_geometric.transforms import RandomNodeSplit
transform = RandomNodeSplit(split='random')
Newdatapyg  = transform(Newdatapyg)
print(Newdatapyg)

I get -

Data(x=[183, 1703], edge_index=[2, 561], y=[183], train_mask=[183], val_mask=[183], test_mask=[183])
When I train this, my accuracy looks like this -

But when I try to test it I get the following error -

Turns out in this line - test_acc = int(test_correct.sum()) / int(Newdatapyg.test_mask.sum()) the denominator is 0. Is it the problem of the model? Is it bad that its not able to work on test set? Or am I doing something wrong? Thanks for your time :)

[Padarn]

Are you sure that there are any nodes in the test set? The test_mask could be full of zeros, making the denominator zero.

[AdarshMJ]

Is it possible because of the way the test_mask looks? If i print -

Data(x=[183, 1703], edge_index=[2, 298], y=[183], train_mask=[183], val_mask=[183], test_mask=[183])

print(Newdatapyg.test_mask)

I get -
tensor([False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False, False, False, False, False, False, False, False,
        False, False, False])

print(Newdatapyg.test_mask.sum())
I get - tensor(0)

How to train a model if all the test_mask consists of all False?

[Padarn]

Yes then this is the cause.

You can still train the model you'll just need to remove the test_acc (which is meaningless as you have no test cases). It looks like this is from code you wrote so it should be simple to remove?

[AdarshMJ]

Yes. Thank you :) I do have val_mask which are not all False. So I can use that instead of test_mask.