Node classification on WebKB datasets

[AdarshMJ]

Hi! @rusty1s I am trying to run node classification on WebKB datasets - Texas, Wisconsin and Cornell (https://pytorch-geometric.readthedocs.io/en/latest/modules/datasets.html#torch_geometric.datasets.WebKB), but I keep getting this error.
Any idea how to resolve this?

Thank you in advance!

[EdisonLeeeee]

Hi @AdarshMJ

There are ten different splits in WebKB so the train_mask (or test_mask) got the shape [183, 10]. You need to specify the split_idx for a given mask:

# `split_idx` could be 0, 1, ..., 9
mask = data.train_mask[:, split_idx]
loss =  criterion(out[mask], data.y[mask])

[AdarshMJ]

Hi @EdisonLeeeee Thank you for your response. I have got it to working, but the accuracy is very low, heres the code I am working with -

The model is a simple GCN.

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

def train():
      model.train()
      optimizer.zero_grad()  # Clear gradients.
      out = model(data.x, data.edge_index)  # Perform a single forward pass.
      loss = criterion(out[data.train_mask[:, 0]], data.y[data.train_mask[:, 0]])  # Compute the loss solely based on the training nodes.
      loss.backward()  # Derive gradients.
      optimizer.step()  # Update parameters based on gradients.
      return loss

def test():
      model.eval()
      out = model(data.x, data.edge_index)
      pred = out.argmax(dim=1)  # Use the class with highest probability.
      correct = pred[data.val_mask[:, 0]] == data.y[data.val_mask[:, 0]]  # Check against ground-truth labels.
      test = int(correct.sum()) / int(data.val_mask.sum())  # Derive ratio of correct predictions.
      return test


for epoch in range(1, 101):
    loss = train()
    #print(f'Epoch: {epoch:03d}, Loss: {loss:.4f}')
test = test()
print(f'Test Accuracy: {test:.4f}')

The accuracy I get is very low - Test Accuracy: 0.0610

Do you think I should play around with the hyperparameters or is there something wrong with my code? Any insights would be helpful! Thank you.

[EdisonLeeeee]

In test, the computation of accuracy should be

test = int(correct.sum()) / int(data.val_mask[:, 0].sum())

and you'll get your accuracy of 0.61

[AdarshMJ]

Thank you!