RandomNodeSplit "NotImplementedError: " Error

[archids]

Hi:

I am trying out a HeteroGraph for Node Classification using HeteroGNN based on pytorch_geometric examples with pretty much default settings. Everything seems to be ok without errors except when it comes to splitting the custom dataset using RandomNodeSplit.

transform = random_node_split.RandomNodeSplit('train_rest')
train_mask, val_mask, test_mask = transform(dataset)

File f:\condos\lib\site-packages\torch_geometric\data\hetero_data.py:777, in HeteroData.__iter__(self)
    776 def __iter__(self):
--> 777     raise NotImplementedError

My Heterodata looks like this.

HeteroData(
  Room={
    x=[1446, 2],
    y=[1446]
  },
  Wall={ x=[2647, 18] },
  Level={ x=[6, 7] },
  Door={ x=[1189, 26] },
  Window={ x=[316, 7] },
  Stair={ x=[33, 6] },
  (Room, BoundedBy, Wall)={ edge_index=[2, 7717] },
  (Room, LooksOut, Window)={ edge_index=[2, 317] },
  (Door, Connects, Room)={ edge_index=[2, 2366] },
  (Room, ConnectedBy, Door)={ edge_index=[2, 2366] },
  (Room, LocatedOn, Level)={ edge_index=[2, 1446] },
  (Door, HostedOn, Wall)={ edge_index=[2, 1189] },
  (Window, HostedOn, Wall)={ edge_index=[2, 316] },
  (Level, AccessedBy, Stair)={ edge_index=[2, 78] }
)

And the codes for HeteroGNN model:

class HeteroGNN(torch.nn.Module):
    def __init__(self, metadata, hidden_channels, out_channels, num_layers):
        super().__init__()

        self.convs = torch.nn.ModuleList()
        for _ in range(num_layers):
            conv = HeteroConv({
                edge_type: SAGEConv((-1, -1), hidden_channels)
                for edge_type in metadata[1]
            })
            self.convs.append(conv)

        self.lin = Linear(hidden_channels, out_channels)

    def forward(self, x_dict, edge_index_dict):
        for conv in self.convs:
            x_dict = conv(x_dict, edge_index_dict)
            x_dict = {key: F.leaky_relu(x) for key, x in x_dict.items()}
        return self.lin(x_dict['Room'])

model = HeteroGNN(dataset.metadata(), hidden_channels=64, out_channels = num_room_labels,
                  num_layers=2)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
data, model = dataset.to(device), model.to(device)

with torch.no_grad():  # Initialize lazy modules.
    out = model(data.x_dict, data.edge_index_dict)

optimizer = torch.optim.Adam(model.parameters(), lr=0.005, weight_decay=0.001)

def train():
    model.train()
    optimizer.zero_grad()
    out = model(data.x_dict, data.edge_index_dict)
    mask = data['Room'].train_mask
    loss = F.cross_entropy(out[mask], data['Room'].y[mask])
    loss.backward()
    optimizer.step()
    return float(loss)

@torch.no_grad()
def test():
    model.eval()
    pred = model(data.x_dict, data.edge_index_dict).argmax(dim=-1)

    accs = []
    for split in ['train_mask', 'val_mask', 'test_mask']:
        mask = data['Room'][split]
        acc = (pred[mask] == data['Room'].y[mask]).sum() / mask.sum()
        accs.append(float(acc))
    return accs

for epoch in range(1, 101):
    loss = train()
    train_acc, val_acc, test_acc = test()
    print(f'Epoch: {epoch:03d}, Loss: {loss:.4f}, Train: {train_acc:.4f}, '
          f'Val: {val_acc:.4f}, Test: {test_acc:.4f}')

Any help is highly appreciated. Thanks!

[EdisonLeeeee]

Hi, RandomNodeSplit returns the transformed graph (a HeteroData instance), so it should be

transform = random_node_split.RandomNodeSplit('train_rest')
data = transform(data)

HeteroData is not iterable.

[EdisonLeeeee]

No worries. You can specify RandomNodeSplit(num_val=200, num_test=500) for the smaller size of validation and test set, where 200 and 500 can be specified as you like, just make sure that val+test<1446.

[archids]

I just saw your follow up response.
i get this error now.

File f:\condos\lib\site-packages\torch\nn\functional.py:3014, in cross_entropy(input, target, weight, size_average, ignore_index, reduce, reduction, label_smoothing)
   3012 if size_average is not None or reduce is not None:
   3013     reduction = _Reduction.legacy_get_string(size_average, reduce)
-> 3014 return torch._C._nn.cross_entropy_loss(input, target, weight, _Reduction.get_enum(reduction), ignore_index, label_smoothing)

IndexError: Target 43 is out of bounds.

[EdisonLeeeee]

Have a check on data['room'].y.max(). It is larger than your out-channel specified as num_room_labels. It should be ensured that num_room_labels=data['room'].y.max().item() + 1

[archids]

You are awesome! That works. For some reason i thought that the num_room_label should be equal to Unique Labels for the Room.

Epoch: 295, Loss: 0.4030, Train: 0.8807, Val: 0.8350, Test: 0.8720
Epoch: 296, Loss: 0.4020, Train: 0.8753, Val: 0.8300, Test: 0.8660
Epoch: 297, Loss: 0.3972, Train: 0.8780, Val: 0.8400, Test: 0.8640
Epoch: 298, Loss: 0.3903, Train: 0.8780, Val: 0.8350, Test: 0.8620
Epoch: 299, Loss: 0.3813, Train: 0.8874, Val: 0.8350, Test: 0.8660
Epoch: 300, Loss: 0.3771, Train: 0.8861, Val: 0.8300, Test: 0.8620

I am quite happy with it for the first try. Thanks again!!!

[EdisonLeeeee]

Great : )