IndexError：尺寸超出范围（预计在 [-1, 0] 范围内，但得到 -2）与 torch_geometry

Question

. Answers to this question are eligible for a +50 reputation bounty. spencer is looking for an .

I have some trouble with IndexError in my GCN code. The error occurred on the GCNConv layer. I cannot understand why self.node_dim is included in the size function which is shown in the error code.

How can I solve this problem? data.x is node_feature in which the number of feature for each node is just one and the number of node is 2058.

Thanks in advance.

'''

class GCNEncoder(torch.nn.Module):

    def __init__(self, in_channels, out_channels):
        super(GCNEncoder, self).__init__()
        self.conv1 = GCNConv(in_channels, out_channels, cached=True)
        self.conv2 = GCNConv(out_channels, out_channels, cached=True)

    def forward(self, x, edge_index):
        x = self.conv1(x, edge_index).relu()
        return self.conv2(x, edge_index)

class OurGAE(torch.nn.Module):

    def __init__(self, num_features, out_chennels, node):
        super(OurGAE, self).__init__()
        self.encoder = GCNEncoder(num_features, out_channels)
        
        self.linear = nn.Linear(out_channels, len(node))
    
    def encode(self, node_feature, edge_lst):
        z = self.encoder(node_feature, edge_lst)
        return z
    
    def predict(self, z):        
        predicted_node_features = self.linear(z)
        return predicted_node_features
    
    
############ train model #######

dataset = OurDataset(one_node_feature, super_edge_lst)

out_channels = 16
num_features = dataset.num_features

model = OurGAE(num_features, out_channels, node)

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = model.to(device)

optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
epochs = 100

criterion = nn.MSELoss()

#model.train()

for epoch in range(1, epochs + 1):

    loss_ep = 0
    for data in dataset:    

        optimizer.zero_grad()
        
        x = data.x.to(device)
        edge_lst = data.edge_index.to(device)
        
        z = model.encode(x, edge_lst)
        
        super_z = z[-1]
        
        predicted_x = model.predict(super_z)
        
        loss = criterion(x.reshape(-1), predicted_x)
        
        loss.backward()
        optimizer.step()
        
        loss_ep += loss.cpu().detach().data.numpy()

    if epoch % 1 == 0:
        print(f"epoch: {epoch}, loss : {loss_ep}") '''

---------------------------------------------------------------------------
IndexError                                Traceback (most recent call last)
<ipython-input-103-9dd243b81c8e> in <module>
     71         edge_lst = data.edge_index.to(device)
     72 
---> 73         z = model.encode(x, edge_lst)
     74 
     75         super_z = z[-1]

<ipython-input-103-9dd243b81c8e> in encode(self, node_feature, edge_lst)
     36 
     37     def encode(self, node_feature, edge_lst):
---> 38         z = self.encoder(node_feature, edge_lst)
     39         return z
     40 

~/anaconda3/lib/python3.8/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
    725             result = self._slow_forward(*input, **kwargs)
    726         else:
--> 727             result = self.forward(*input, **kwargs)
    728         for hook in itertools.chain(
    729                 _global_forward_hooks.values(),

<ipython-input-103-9dd243b81c8e> in forward(self, x, edge_index)
     15 
     16     def forward(self, x, edge_index):
---> 17         x = self.conv1(x, edge_index).relu()
     18         return self.conv2(x, edge_index)
     19 

~/anaconda3/lib/python3.8/site-packages/torch/nn/modules/module.py in _call_impl(self, *input, **kwargs)
    725             result = self._slow_forward(*input, **kwargs)
    726         else:
--> 727             result = self.forward(*input, **kwargs)
    728         for hook in itertools.chain(
    729                 _global_forward_hooks.values(),

~/anaconda3/lib/python3.8/site-packages/torch_geometric/nn/conv/gcn_conv.py in forward(self, x, edge_index, edge_weight)
    159                 if cache is None:
    160                     edge_index, edge_weight = gcn_norm(  # yapf: disable
--> 161                         edge_index, edge_weight, x.size(self.node_dim),
    162                         self.improved, self.add_self_loops)
    163                     if self.cached:

IndexError: Dimension out of range (expected to be in range of [-1, 0], but got -2)

Answer 1

TLDR: reshape your data-points to [num_nodes, num_node_features] => [2058, 1] before constructing the dataset.

It looks like the OurDataset you created doesn't fit the assumptions of the graph dataset. I just assume this is because you have only one feature at each node, and you most probably didn't reshape the data before passing it to the Dataset.

Taken from the documentation : Data Handling of Graphs

• data.x: Node feature matrix with shape [num_nodes, num_node_features]

from torch_geometric.data import Data

edge_index = torch.tensor([[0, 1, 1, 2], [1, 0, 2, 1]], dtype=torch.long)

x_wrong_dims = torch.tensor([-1, 0, 1], dtype=torch.float)
data_wrong_dims = Data(x=x_wrong_dims, edge_index=edge_index)
data_wrong_dims.x.size()
# torch.Size([3])
data_wrong_dims.x.size(-2)
# IndexError: Dimension out of range (expected to be in range of [-1, 0], but got -2)

x_correct_dims = torch.tensor([-1, 0, 1], dtype=torch.float).reshape([-1, 1])
data_correct_dims = Data(x=x_correct_dims, edge_index=edge_index)
data_correct_dims.x.size()
# torch.Size([3, 1])
data_correct_dims.x.size(-2)
# 3

As to the question why gcn_norm uses the x.size(self.node_dim) I would assume that the function has to infer along which dimension to aggregate for the normalization and node_dim is -2 by default which means it aggregates over the second to last dimensions which in the case of the correctly shaped dataset is the num_nodes dimension.