to_hetero() tries to activate None

[fierval]

🐛 Describe the bug

Converting the following simple model with to_hetero():

    def __init__(self, config: ELConfig):
        super().__init__()

        self.config = config
        self.conv_layers = 2

        self.conv1 = TransformerConv(
            256,
            256,
            heads=4,
            dropout=0.6,
            edge_dim=4
        )

        self.conv2 = TransformerConv(
            256 * 4,
            256,
            heads=1,
            dropout=0.6,
            edge_dim=4
        )

    def forward(self, x, edge_index, edge_attr):

        x = self.conv1(x, edge_index, edge_attr).relu()
        x = self.conv2(x, edge_index, edge_attr).relu()

        return x

With metadata: (['category', 'vertex'], [('category', 'assign', 'vertex'), ('vertex', 'groups', 'vertex'), ('vertex', 'continues', 'vertex'), ('vertex', 'keyof', 'vertex')])

Yields the following:

def forward(self, x, edge_index, edge_attr):
    x__category = x.get('category')
    x__vertex = x.get('vertex');  x = None
    edge_index__category__assign__vertex = edge_index.get(('category', 'assign', 'vertex'))
    edge_index__vertex__groups__vertex = edge_index.get(('vertex', 'groups', 'vertex'))
    edge_index__vertex__continues__vertex = edge_index.get(('vertex', 'continues', 'vertex'))
    edge_index__vertex__keyof__vertex = edge_index.get(('vertex', 'keyof', 'vertex'));  edge_index = None
    edge_attr__category__assign__vertex = edge_attr.get(('category', 'assign', 'vertex'))
    edge_attr__vertex__groups__vertex = edge_attr.get(('vertex', 'groups', 'vertex'))
    edge_attr__vertex__continues__vertex = edge_attr.get(('vertex', 'continues', 'vertex'))
    edge_attr__vertex__keyof__vertex = edge_attr.get(('vertex', 'keyof', 'vertex'));  edge_attr = None
    conv1__vertex1 = self.conv1.category__assign__vertex((x__category, x__vertex), edge_index__category__assign__vertex, edge_attr__category__assign__vertex);  x__category = None
    conv1__vertex2 = self.conv1.vertex__groups__vertex((x__vertex, x__vertex), edge_index__vertex__groups__vertex, edge_attr__vertex__groups__vertex)
    conv1__vertex3 = self.conv1.vertex__continues__vertex((x__vertex, x__vertex), edge_index__vertex__continues__vertex, edge_attr__vertex__continues__vertex)
    conv1__vertex4 = self.conv1.vertex__keyof__vertex((x__vertex, x__vertex), edge_index__vertex__keyof__vertex, edge_attr__vertex__keyof__vertex);  x__vertex = None
    conv1__vertex5 = torch.add(conv1__vertex1, conv1__vertex2);  conv1__vertex1 = conv1__vertex2 = None
    conv1__vertex6 = torch.add(conv1__vertex3, conv1__vertex4);  conv1__vertex3 = conv1__vertex4 = None
    conv1__vertex = torch.add(conv1__vertex5, conv1__vertex6);  conv1__vertex5 = conv1__vertex6 = None
    relu__category = None.relu()
    relu__vertex = conv1__vertex.relu();  conv1__vertex = None
    conv2__vertex1 = self.conv2.category__assign__vertex((relu__category, relu__vertex), edge_index__category__assign__vertex, edge_attr__category__assign__vertex);  relu__category = edge_index__category__assign__vertex = edge_attr__category__assign__vertex = None
    conv2__vertex2 = self.conv2.vertex__groups__vertex((relu__vertex, relu__vertex), edge_index__vertex__groups__vertex, edge_attr__vertex__groups__vertex);  edge_index__vertex__groups__vertex = edge_attr__vertex__groups__vertex = None
    conv2__vertex3 = self.conv2.vertex__continues__vertex((relu__vertex, relu__vertex), edge_index__vertex__continues__vertex, edge_attr__vertex__continues__vertex);  edge_index__vertex__continues__vertex = edge_attr__vertex__continues__vertex = None
    conv2__vertex4 = self.conv2.vertex__keyof__vertex((relu__vertex, relu__vertex), edge_index__vertex__keyof__vertex, edge_attr__vertex__keyof__vertex);  relu__vertex = edge_index__vertex__keyof__vertex = edge_attr__vertex__keyof__vertex = None
    conv2__vertex5 = torch.add(conv2__vertex1, conv2__vertex2);  conv2__vertex1 = conv2__vertex2 = None
    conv2__vertex6 = torch.add(conv2__vertex3, conv2__vertex4);  conv2__vertex3 = conv2__vertex4 = None
    conv2__vertex = torch.add(conv2__vertex5, conv2__vertex6);  conv2__vertex5 = conv2__vertex6 = None
    relu_1__category = None.relu()
    relu_1__vertex = conv2__vertex.relu();  conv2__vertex = None
    return {'category': relu_1__category, 'vertex': relu_1__vertex}

Then relu_category = None.relu() causes it to crash.

Environment

• PyG version: 2.03
• PyTorch version: 1.10
• OS: Win/Linux
• Python version: 3.8
• CUDA/cuDNN version: 11.3 / 8.1
• How you installed PyTorch and PyG (conda, pip, source): pip
• Any other relevant information (e.g., version of torch-scatter): all

[rusty1s]

The issue is that there is no edge type pointing to category. Upgrading to torch-geometric==2.0.4 should at least warn you about this.

[fierval]

Figured as much but thought code like None.relu() was worth reporting.

[rusty1s]

Yes, definitely :) I think the current workaround of warning the user is okay, but I agree that in your case it should definitely crash prior to model execution. I need to look into torch.fx to see if there is some way to check for this.

[stayones]

So every node in the heterogeneous graph should have an edge pointing to it? How should we resolve this problem? I tried to change the graph to Undirected but it didn't work. Should we add reverse edge type to our data by ourselves?

[rusty1s]

Yes, otherwise certain node types will not get properly updated during message passing. The ToUndirected transform should take care of that. Let me know if you encounter any issues wit that.