AttributeError: 'HeteroDataBatch' has no attribute 'y'

[AlizadehF]

Hello,
my code is>
model.train()
for batch in train_loader:
out = model(batch.x_dict, batch.edge_index_dict, batch)

    loss = criterion(out, batch.y) 
    loss.backward()  
    optimizer.step()  
    optimizer.zero_grad()  

error is>
AttributeError: 'HeteroDataBatch' has no attribute 'y'

and when used batch.y_dict, the error is >
TypeError: cross_entropy_loss(): argument 'target' (position 2) must be Tensor, not dict

Can you make suggestions for this error?

[wsad1]

Could you share the result of print(batch).
I am guessing y is a node level tensor, in that case you have to pass batch.y_dict['node_type'] to criterion, where node_type is the node type you are making a prediction on.

[AlizadehF]

print (batch):

HeteroDataBatch( y=[763144], userId={ x=[763144, 1], batch=[763144], ptr=[2] }, processId={ x=[763144, 1], batch=[763144], ptr=[2] }, eventId={ x=[763144, 1], batch=[763144], ptr=[2] }, (userId, processName, processId)={ edge_index=[2, 763144], edge_attr=[1] }, (userId, eventName, eventId)={ edge_index=[2, 763144], edge_attr=[1] }, (processId, rev_processName, userId)={ edge_index=[2, 763144] }, (eventId, rev_eventName, userId)={ edge_index=[2, 763144] }, (userId, rev_processName, processId)={ edge_index=[2, 763144] }, (userId, rev_eventName, eventId)={ edge_index=[2, 763144] }, (processId, rev_processName, userId)={ edge_index=[2, 763144] }, (eventId, rev_eventName, userId)={ edge_index=[2, 763144] }, (userId, rev_processName, processId)={ edge_index=[2, 763144] }, (userId, rev_eventName, eventId)={ edge_index=[2, 763144] }, (processId, rev_processName, userId)={ edge_index=[2, 763144] }, (eventId, rev_eventName, userId)={ edge_index=[2, 763144] }

[AlizadehF]

Thanks I have hetero edges with hetero label And each category of edge have different y size. In the out I have tensor with size 162, 162, 55,55 but in the batch.y I have size 162, 55, 162, 55

[AlizadehF]

It means I have different size of label for each subgraph or edge type I have an edge classification Task. i have different node type, different edge type and for each edge, 22 features and label i have. for example, ``` data = HeteroData() data['n1'].x = torch.tensor(df['f1']).reshape([-1, 1]) data['n1'].x = data['n1'].x.type(torch.FloatTensor) data['n2'].x = torch.tensor(df['f2']).reshape([-1, 1]) data['n2'].x = data['n2'].x.type(torch.FloatTensor) data['n1','type1', 'n2'].edge_index = e1 data['n1','type1', 'n2'].edge_attr = node_features data['n3'].x = torch.tensor(df['f3']).reshape([-1, 1]) data['n3'].x = data['n3'].x.float() data['n4'].x = torch.tensor(df['f3']).reshape([-1, 1]) data['n4'].x = data['n4'].x.float() data['n3','type2', 'n4'].edge_index = e3 data['n3','type2', 'n4'].edge_attr = node_features data['n1','type1', 'n2'].y = y1 data['n3','type2', 'n4'].y = y3 ``` my label vector size is different per edge type

[AlizadehF]

`
data = T.ToUndirected()(data)
class GNN(torch.nn.Module):
def init(self, hidden_channels, out_channels):
super().init()
self.conv1 = SAGEConv((-1,-1), hidden_channels)
self.conv2 = SAGEConv((-1,-1), out_channels)

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

model = GNN(hidden_channels=64, out_channels=1)
model = to_hetero(model, data.metadata(), aggr='sum')
`

out = model(data.x_dict, data.edge_index_dict) loss = criterion(out, data.y_dict) loss.backward(retain_graph=True) optimizer.step() optimizer.zero_grad()

error:
TypeError: cross_entropy_loss(): argument 'input' (position 1) must be Tensor, not dict

i guess my model is for node classification but my goal is edge classification, for this aim, the out is for nodes and the data.y_dict is for edges.
i dont know, how to make a GNN mode for edge classification Heterogeneous Graph.

[rusty1s]

Please note that GNNs compute node-level representations. The general procedure to let them return edge-level representations (e.g., used for edge classification), is to combine the source and destination node features and pass them to a classifier, e.g.:

class MyModel():
    def __init__(self., ...):
        self.gnn = to_hetero(MyGNN())
        self.classifier = Linear(2 * hidden_channels, num_classes)
    
    def forward(self, x_dict, edge_index_dict):
        x_dict = self.gnn(x_dict, edge_index_dict)
        x_src = x_dict[src_node_type]
        x_dst = x_dict[dst_node_type]
        return self.classifier(torch.cat([x_src, x_dst], dim=-1))

[AlizadehF]

thanks for your response,,
i define this code x_src = x_dict[src_node_type], x_dst = x_dict[dst_node_type] for all node types?
and , this code is for edge classification??

[rusty1s]

You can define them for all edge types you are interested in for classification

[AlizadehF]

I have used this code for init
def __init__(self, hidden_channels, out_channels): super().__init__() self.conv1 = SAGEConv((-1,-1), hidden_channels) self.conv2 = SAGEConv((-1,-1), hidden_channels) self.pool = aggr.MeanAggregation() self.lin = Linear(hidden_channels, 11)

your init def is different , may you explain self.classifier = Linear(2 * hidden_channels, num_classes) ?

[rusty1s]

So in my example, you would have a GNN (just as you did), but then have a parent module that wraps the heterogeneous GNN (via to_hetero()) and a final edge-level predictor.3