Heterogeneous graph model application design problem? GCN, GAT

[Zhang-1218]

Heterogeneous graph model application design problem? GCN, GAT

Hello, I am learning graph network recently, thank you for providing the pyg open source library

My demo dataset is a heterogeneous graph data, referring to the dataset designed by the official website, which contains two types of nodes and two types of edges
Node:
(1) paper
(2) auther
Edge（edge is undirected）:
(1) paper-cite-paper
(2) auther-write-paper

The data is as follows：

HeteroData(
  paper={
    x=[156, 7],
    y=[156],
    num_nodes=156,
    train_mask=[156],
    val_mask=[156],
    test_mask=[156]
  },
  author={
    x=[152, 7],
    y=[152],
    num_nodes=152
  },
  (author, write, paper)={
    edge_index=[2, 365],
    edge_attr=[365, 3]
  },
  (paper, cite, paper)={
    edge_index=[2, 2675],
    edge_attr=[2675, 6]
  },
  (paper, rev_write, author)={
    edge_index=[2, 365],
    edge_attr=[365, 3]
  }
)

Both author and paper nodes have features.
Both paper-cite-paper and auther-write-paper also have features
The purpose of the experiment is to hope to predict the category of the paper

1、I'm currently running into a problem:

  File "GCNProject/model/baseHeteModel.py", line 64, in basehetegat_train
    out = hgtmodel(data.x_dict, data.edge_index_dict)
  File "/root/anaconda3/envs/lib/python3.8/site-packages/torch/fx/graph_module.py", line 308, in wrapped_call
    return cls_call(self, *args, **kwargs)
  File "/root/anaconda3/envs/lib/python3.8/site-packages/torch/fx/graph_module.py", line 308, in wrapped_call
    return cls_call(self, *args, **kwargs)
  File "/root/anaconda3/envs/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "<eval_with_key_1>", line 8, in forward
  File "/root/anaconda3/envs/lib/python3.8/site-packages/torch/nn/modules/module.py", line 889, in _call_impl
    result = self.forward(*input, **kwargs)
  File "/root/anaconda3/envs/lib/python3.8/site-packages/torch_geometric/nn/conv/gat_conv.py", line 218, in forward
    out = self.propagate(edge_index, x=x, alpha=alpha, edge_attr=edge_attr,
  File "/root/anaconda3/envs/lib/python3.8/site-packages/torch_geometric/nn/conv/message_passing.py", line 262, in propagate
    size = self.__check_input__(edge_index, size)
  File "/root/anaconda3/envs/lib/python3.8/site-packages/torch_geometric/nn/conv/message_passing.py", line 149, in __check_input__
    raise ValueError(
ValueError: `MessagePassing.propagate` only supports `torch.LongTensor` of shape `[2, num_messages]` or `torch_sparse.SparseTensor` for argument `edge_index`.

I refer to the official website documentation and design the following model:

class HeteroGNN(torch.nn.Module):
    def __init__(self, hidden_channels, out_channels):
        super().__init__()
        self.conv1 = GATConv((-1, -1), hidden_channels, add_self_loops=False)
        self.lin1 = Linear(-1, hidden_channels)
        self.conv2 = GATConv((-1, -1), out_channels, add_self_loops=False)
        self.lin2 = Linear(-1, out_channels)

    def forward(self, x, edge_index):
        x = self.conv1(x, edge_index) + self.lin1(x)
        x = x.relu()
        x = self.conv2(x, edge_index) + self.lin2(x)
        return x


model = HeteroGNN(hidden_channels=4, out_channels=dataset.num_classes)
hgtmodel = to_hetero(model, data.metadata(), aggr='sum')

hgtmodel.train()
optimizer.zero_grad()
out = hgtmodel(data.x_dict, data.edge_index_dict)

I hope to be able to introduce both node and edge features into the calculation, but the documentation on the official website does not show how to use edge features for learning on heterogeneous graphs converted from homogeneous graphs.（GAT、GCN）
https://pytorch-geometric.readthedocs.io/en/latest/notes/heterogeneous.html

2、I saw on the official website that the heterogeneous convolution wrapper can be used to learn different models for different edges, so for my demo, is it possible to set paper-cite-paper to use gcn, auther-write-paper to use GAT, paper-rev_write-auther using GraphSAGE?
https://pytorch-geometric.readthedocs.io/en/latest/notes/heterogeneous.html#using-the-heterogenous-convolution-wrapper

If possible, my question is still how to pass edge features to convolutional models corresponding to different edges?

[rusty1s]

Can you upload the data (e.g., via torch.save(data, 'data.pt')? I'm happy to look into the issue.

You should be able to use any GNN operator that supports edge features as well, e.g.:

conv = GATConv((-1, -1), hidden_channels, edge_dim=-1)

and call it via

conv(x, edge_index, edge_attr)

Heterogeneous models can then be called via

hgtmodel(data.x_dict, data.edge_index_dict, data.edge_attr_dict)

2. Yes, that is possible. The procedure to add edge features is the same (but you need to use GNN operators that can operate on them).

[Zhang-1218]

Emmm
it doesn't work, GCNConv can't find edge_dim

 GCNConv(-1, hidden_channels, edge_dim=-1),
  File "/root/anaconda3/envs/lib/python3.8/site-packages/torch_geometric/nn/conv/gcn_conv.py", line 126, in __init__
    super(GCNConv, self).__init__(**kwargs)
TypeError: __init__() got an unexpected keyword argument 'edge_dim'

[rusty1s]

GCNConv does not have an edge_dim attribute. I'm sorry: I read GATConv instead of GCNConv. I fixed the example in my code snippet above.

[Zhang-1218]

Emm
I have tried modifying it, but the problem still exists.

conv = HeteroConv({
    ('paper', 'cites', 'paper'): GCNonv(-1, hidden_channels),
    ('author', 'writes', 'paper'): SAGEConv((-1, -1), hidden_channels),
    ('paper', 'rev_writes', 'author'): GATConv((-1, -1), hidden_channels, edge_dim=...),
}, aggr='sum')

ERROR:

GATConv((-1, -1), hidden_channels, edge_dim=...),
  File "/root/anaconda3/envs/lib/python3.8/site-packages/torch_geometric/nn/conv/gat_conv.py", line 147, in __init__
    self.lin_edge = Linear(edge_dim, heads * out_channels, bias=False,
  File "/root/anaconda3/envs/hang/lib/python3.8/site-packages/torch_geometric/nn/dense/linear.py", line 58, in __init__
    if in_channels > 0:
TypeError: '>' not supported between instances of 'ellipsis' and 'int'

And in this case, there is no way for GCN to learn the weights of the edges.?

[rusty1s]

If you have edge features for ('paper', 'cites', 'paper') relation, why not make use of GNN layer that can support it? Note that you will need to replace the edge_dim=... attribute of GATConv with the actual value of edge_dim, in your case edge_dim=3.

[Zhang-1218]

oh, I'm so sorry, I forgot to reply to you
My problem has been solved, thank you very much for your answer