Forwarding A Custom Batch Through GCNConv Network

[askrix]

Dear community,

I am a beginner in geometric learning and attempting to solve the following task at hand. Having transductive learning setting, I wish to forward a custom batch of nodes from Cora dataset through GCNConv network of two layers, which I took over from the example.

The complete/default Cora training dataset consists of 

• x – node features matrices
• edge_index – graph adjacency matrix
• edge_attr - edge weights

Each of these tensors is of shape

• torch.Size([2708, 1433])
• torch.Size([2, 346624])
• torch.Size([346624])

In my code, I have a batch of 100 randomly depicted nodes and therefore the shapes of the three tensors are different

• torch.Size([100, 1433])
• torch.Size([2, 27073])
• torch.Size([27073])

Running my code, I got the following error message

Traceback (most recent call last):
  File "dataset_generation.py", line 324, in <module>
    outputs = gnn(batch, custom_forwarding=True)
  File "/home/user/Documents/test/dev_venv/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "dataset_generation.py", line 239, in forward
    x = F.relu(self.conv1(x, edge_index, edge_weight))
  File "/home/user/Documents/test/dev_venv/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "/home/user/Documents/test/dev_venv/lib/python3.8/site-packages/torch_geometric/nn/conv/gcn_conv.py", line 185, in forward
    out = self.propagate(edge_index, x=x, edge_weight=edge_weight,
  File "/home/user/Documents/test/dev_venv/lib/python3.8/site-packages/torch_geometric/nn/conv/message_passing.py", line 310, in propagate
    coll_dict = self.__collect__(self.__user_args__, edge_index,
  File "/home/user/Documents/test/dev_venv/lib/python3.8/site-packages/torch_geometric/nn/conv/message_passing.py", line 202, in __collect__
    data = self.__lift__(data, edge_index,
  File "/home/user/Documents/test/dev_venv/lib/python3.8/site-packages/torch_geometric/nn/conv/message_passing.py", line 172, in __lift__
    return src.index_select(self.node_dim, index)
IndexError: index out of range in self

Adding the print statement before the line of the error's occurrence in message_passing.py, I got the shapes of the tensors involved

print(src.shape, index.shape, self.node_dim)
>>> torch.Size([100, 16]) torch.Size([27073]) -2

I concluded that the problem lies in the dimensionality mismatch at the message-processing part. The mismatch is caused by the size of adjacency matrix compared with the section of feature matrices under consideration. However, I do not understand how to solve the issue myself and ask therefore for your help.

I have not found similar questions in the Q&A section of the forum, no one having had asked this question before, and certainly, someone will eventually ask such or a similar question.

Many thanks in advance and have yourselves an epic day!

[rusty1s]

It looks like there is a mismatch between indices in edge_index and the number of nodes in your dataset (denoted by x.size(0)). In general, PyG requires that

assert edge_index.min() >= 0
assert edge_index.max() < x.size(0)

That is, all indices in edge_index point to valid nodes. Can you ensure that the assert statements pass for you?

[askrix]

Good afternoon @rusty1s,

As you suggested, I added both assertion statement right at the beginning of the forward method and got the following error message

Traceback (most recent call last):
  File "dataset_generation.py", line 326, in <module>
    outputs = gnn(batch, custom_forwarding=True)
  File "/home/user/test/dev_venv/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "dataset_generation.py", line 238, in forward
    assert edge_index.min() > 0
AssertionError

I am uncertain if I understand the cause of the error as of it now. To make the discussion easier, please find in the following the code I am using to construct the custom dataset:

class PairwiseCora(Dataset):
    """
    pairwise node version of Cora dataset based on the query list
    torch.utils.data.Dataset version of PyTorch Geometric'c edition of Cora
    I.e., torch.geometric.data.Data => torch.utils.data.Dataset
    
    @param dataset:
    @param examples:
    :return: torch.utils.data.Dataset
    """
    def __init__(self, dataset, examples):
        self.data = list()
        self.dataset = dataset
        #the file with node indices for each query musst exist!
        assert os.path.exists(examples)
        #read the file and save the indices per query as tuple 
        with open(examples) as f:
            for line in f:
                line = line.strip().split(' ')
                self.data.append(tuple([int(i) for i in line]))

    def __len__(self):
        return len(self.data)
    
    def get_node_related_features(self, target_node_id:int):
        #obtain from the graph individual node's characteristics.
        #i.e., feature matrix, adjacency matrix, edge features.
        #1. get node's feature matrix
        node_feature_matrix = self.dataset.x[target_node_id]
        #construct a tensor on node's graph connectivity.
        edge_id_list = []
        edge_index_list = []
        for edge_id in range(len(self.dataset.edge_index[0])):
            if self.dataset.edge_index[0][edge_id] == target_node_id or self.dataset.edge_index[1][edge_id] == target_node_id:
                edge_id_list.append(edge_id)
                edge_index_list.append(torch.tensor([self.dataset.edge_index[0][edge_id],self.dataset.edge_index[1][edge_id]]))
        a = torch.tensor([i[0] for i in edge_index_list])
        b = torch.tensor([i[1] for i in edge_index_list])
        #2. the stack of a and b expresses node's graph connectivity, i.e. adjacency matrix.
        node_graph_connectivity = torch.vstack((a,b)).view(2,-1)
        #3. obtrain node's edge feature matrix.
        edge_feature_matrix = data.edge_attr[torch.tensor(edge_id_list)]
        #return all three.
        return node_feature_matrix, node_graph_connectivity, edge_feature_matrix

    def __getitem__(self, index):
        #get indices of nodes involved in the query under the given 'index'.
        n1, n2 = self.data[index]
        #get n1's and n2's characteristics.
        nfm1, ngc1, efm1 = self.get_node_related_features(n1)
        n1 = torch.Tensor([n1]).to(torch.int32)
        nfm2, ngc2, efm2 = self.get_node_related_features(n2)
        n2 = torch.Tensor([n2]).to(torch.int32)        
        return {'n1': [n1, nfm1, ngc1, efm1], 'n2': [n2, nfm2, ngc2, efm2]}

After having read your answer, I thought that to resolve the initial error, I have to pass  the feature matrices of all nodes which are connected to the node(s) under consideration. Now, it appears to me that it is not the only issue.

Thanks and easy start to the new week!

[askrix]

Hey @rusty1s,

Could you confirm my understanding of the meaning of both the assertion statements you formulated in your initial answer?

In my understanding, the assertion of

assert edge_index.min() > 0

ensures that the minimum of edge index's tensor points to a positive integer, depicting a node in the graph. It is important to ensure since all nodes in a graph have positive integers as indices.

However, I do not understand the true meaning of the second assertion
The second assertion of

assert edge_index.max() < x.size(0)

ensures that there is the smaller number of the node feature matrices is given as the maximal index of the node from the adjacency matrix?
Why exactly the maximal index of the connected nodes suppose to be smaller than the feature matrix of the node under consideration?

Thanks again!

[rusty1s]

Yes, this is correct. Let me know if the fixed assert with edge_index.min() >= 0 now runs through for you.

[rusty1s]

Sorry, I mean assert edge_index.min() >= 0 :)

[askrix]

Hello @rusty1s,

Thanks for the confirmation. Unfortunately, even after the correction I still got an error, but this time due to the second assertion of

assert edge_index.max() < x.size(0)

Since I am certain that which constraint, this assertion expresses, I have come up with the trick to form a zero-tensor for the node feature matrices of the same size as the original dataset. Thus, filling only the relevant rows of tensor, my forward pass works again.

Please have a look at the code attached

class PairwiseCora(Dataset):
    """
    pairwise node version of Cora dataset based on the query list
    torch.utils.data.Dataset version of PyTorch Geometric'c edition of Cora
    I.e., torch.geometric.data.Data => torch.utils.data.Dataset
    
    @param dataset:
    @param examples:
    :return: torch.utils.data.Dataset
    """
    def __init__(self, dataset, examples):
        self.data = list()
        self.dataset = dataset
        #the file with node indices for each query musst exist!
        assert os.path.exists(examples)
        #read the file and save the indices per query as tuple 
        with open(examples) as f:
            for line in f:
                line = line.strip().split(' ')
                self.data.append(tuple([int(i) for i in line]))

    def __len__(self):
        return len(self.data)
    
    def get_node_related_features(self, target_node_id:int):
        #obtain from the graph individual node's characteristics.
        #I.e., feature matrix, connectivity, edge feature matrix.
        #1. get node's feature matrix
        tmp = torch.zeros(self.dataset.x.shape, dtype=self.dataset.x.dtype)
        tmp[target_node_id] = self.dataset.x[target_node_id]
        #construct a tensor on node's graph connectivity
        edge_id_list = []
        edge_index_list = []
        for edge_id in range(len(self.dataset.edge_index[0])):
            if self.dataset.edge_index[0][edge_id] == target_node_id or self.dataset.edge_index[1][edge_id] == target_node_id:
                edge_id_list.append(edge_id)
                edge_index_list.append(torch.tensor([self.dataset.edge_index[0][edge_id],self.dataset.edge_index[1][edge_id]]))
        a = torch.tensor([i[0] for i in edge_index_list])
        b = torch.tensor([i[1] for i in edge_index_list])
        #3. the stack von a and b is node's graph connectivity
        node_graph_connectivity = torch.vstack((a,b))  # .view(2,-1)
        #2. obtrain node's edge feature matrix
        edge_feature_matrix = data.edge_attr[torch.tensor(edge_id_list)]
        #1. add the feature matrices of the nodes connected to 'target_node_id'
        tmp[a] = self.dataset.x[a]
        node_feature_matrix = tmp
        #return all three
        return node_feature_matrix, node_graph_connectivity, edge_feature_matrix

    def __getitem__(self, index):
        #get indices of nodes involved in the query under the given 'index'
        n1, n2 = self.data[index]
        #get n1's and n2's characteristics
        nfm1, ngc1, efm1 = self.get_node_related_features(n1)
        nfm2, ngc2, efm2 = self.get_node_related_features(n2)
        n1 = torch.Tensor([n1]).to(torch.int32)
        n2 = torch.Tensor([n2]).to(torch.int32)        
        return {'n1': [n1, nfm1, ngc1, efm1], 'n2': [n2, nfm2, ngc2, efm2]}

[askrix]

However, building the current version of the custom dataset takes a lot of RAM. I will try just to stack the relevant node feature matrices instead.

[askrix]

If I attempt to stack only the relevant node feature matrices, I get the same error due to the second assertion:

Traceback (most recent call last):
  File "dataset_generation.py", line 312, in <module>
    outputs = gnn(batch, custom_forwarding=True)
  File "/home/user/test/dev_venv/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1102, in _call_impl
    return forward_call(*input, **kwargs)
  File "dataset_generation.py", line 244, in forward
    assert edge_index.max() < x[0].size(0)
AssertionError

This takes the cake. I need your help, @rusty1s, to find a less memory consuming solution.

Thanks in advance!

[rusty1s]

Can you clarify what you mean? What is x[0] here?