TypeError: 'int' object is not callable in PYG HeteroConv

[amalislam675]

I apply the HetroConv via this below mentioned code, but every time I find this error, 'int' object is not callable. Can you help me to fix this error?
Exact error message is :
Traceback (most recent call last): File "/hetegeneousgraph/code/src/method/model1/main.py", line 248, in <module> train(model, optimizer, train_data) File "/hetegeneousgraph/code/src/method/model1/main.py", line 58, in train trn_loss = train_epoch() File "/hetegeneousgraph/code/src/method/model1/main.py", line 40, in train_epoch score = model(data.x_dict, data.edge_index_dict, data.edge_weight_dict) File "/.conda/envs/foo/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl return forward_call(*input, **kwargs) File "/hetegeneousgraph/code/src/method/model1/model1.py", line 28, in forward h = self.conv1(x_dict, edge_index_dict, edge_weight_dict) File "/.conda/envs/foo/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl return forward_call(*input, **kwargs) File "/.conda/envs/foo/lib/python3.10/site-packages/torch_geometric/nn/conv/hetero_conv.py", line 136, in forward out = conv(x_dict[src], edge_index, *args, **kwargs) File "/.conda/envs/foo/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl return forward_call(*input, **kwargs) File "/.conda/envs/foo/lib/python3.10/site-packages/torch_geometric/nn/conv/gcn_conv.py", line 211, in forward edge_index, edge_weight, x.size(self.node_dim), TypeError: 'int' object is not callable

The code which I used is this one, which is shown below:
`import torch
from torch_geometric.nn import GCNConv, HeteroConv

class Model(torch.nn.Module):
def init(self, out_channels, hidden_channels=512):
super().init()

def edge_weight_fn(edge_index, node_features):
  # Compute the edge weights
  edge_weights = torch.ones(edge_index.size(1), device=edge_index.device)
  return edge_weights

self.conv1 = HeteroConv({
    ('ntype1', 'etype1', 'ntype1'): GCNConv(8704, hidden_channels, edge_weight=edge_weight_fn),
    ('ntype2', 'etype2', 'ntype2'): GCNConv(8704, hidden_channels, edge_weight=edge_weight_fn),
    ('ntype3', 'etype3', 'ntype3'): GCNConv(8704, hidden_channels, edge_weight=edge_weight_fn)
  }, aggr='sum')
self.conv2 = HeteroConv({
    ('ntype1', 'etype1', 'ntype1'): GCNConv(hidden_channels, hidden_channels, edge_weight=edge_weight_fn),
    ('ntype2', 'etype2', 'ntype2'): GCNConv(hidden_channels, hidden_channels, edge_weight=edge_weight_fn),
    ('ntype3', 'etype3', 'ntype3'): GCNConv(hidden_channels, hidden_channels, edge_weight=edge_weight_fn)
  }, aggr='sum')

self.lin = Linear(hidden_channels*3, out_channels)

def forward(self, x_dict, edge_index_dict, edge_weight_dict):

h = self.conv1(x_dict, edge_index_dict, edge_weight_dict)
h = {key: x.relu() for key, x in h.items()}
h = self.conv2(h, edge_index_dict, edge_weight_dict)
h = {key: x.relu() for key, x in h.items()}

# Concatenate node embeddings from different node types
concatenated_h = torch.cat([h[key] for key in h.keys()], dim=1)

# Apply the linear layer to the concatenated embeddings
out = self.lin(concatenated_h)

return out`

My data Object is of this shape
HeteroData( ntype1={ x=[8704, 8704] }, ntype2={ x=[8704, 8704] }, ntype3={ x=[8704, 8704] }, (ntype1, etype1, ntype1)={ edge_index=[2, 1511590], edge_weight=[1511590] }, (ntype2, etype2, ntype2)={ edge_index=[2, 1645254], edge_weight=[1645254] }, (ntype3, etype3, ntype3)={ edge_index=[2, 322990] }, (ntype3, etype3, ntype3)={ edge_weight=[322990] } )
main function
model = Model(out_channels=5) out = model(data.x_dict, data.edge_index_dict, data.edge_weight_dict)

[rusty1s]

Mh, this looks correct to me (except for the edge_weight argument in the definition of GCNConv). Which version of PyG are you using? This example here runs fine for me:

import torch

from torch_geometric.data import HeteroData
from torch_geometric.nn import GCNConv, HeteroConv

conv = HeteroConv(
    {
        ('ntype1', 'etype1', 'ntype1'): GCNConv(-1, 32),
        ('ntype2', 'etype2', 'ntype2'): GCNConv(-1, 32),
        ('ntype3', 'etype3', 'ntype3'): GCNConv(-1, 32),
    }, aggr='sum')

data = HeteroData()
data['ntype1'].x = torch.randn(10, 16)
data['ntype2'].x = torch.randn(10, 16)
data['ntype3'].x = torch.randn(10, 16)
data['ntype1', 'etype1', 'ntype1'].edge_index = torch.randint(0, 10, (2, 20))
data['ntype1', 'etype1', 'ntype1'].edge_weight = torch.rand(20)
data['ntype2', 'etype2', 'ntype2'].edge_index = torch.randint(0, 10, (2, 20))
data['ntype2', 'etype2', 'ntype2'].edge_index = torch.randint(0, 10, (2, 20))
data['ntype3', 'etype3', 'ntype3'].edge_weight = torch.rand(20)
data['ntype3', 'etype3', 'ntype3'].edge_weight = torch.rand(20)

out_dict = conv(data.x_dict, data.edge_index_dict, data.edge_weight_dict)

[amalislam675]

@rusty1s , I checked all input types, all inputs type is of pytorch tensor. I donot understand why this error come?

[rusty1s]

Do you have a small example to reproduce? I can try to find the issue :)

[amalislam675]

@rusty1s , I run this code on my data then error occurs, If I run it on random data then it works fine. In the previous comment, I shared, how I convert my adjacency matrices to PYG edge_index, edge_weight format. I will try to share the example.

[rusty1s]

Is it possible for you to upload a subset of your data? Alternatively, you can just share the HeteroData object with me via torch.save.

[amalislam675]

@rusty1s , I will try to share the data.

[vTechSavyy]

In case it helps, I was also facing the same error
packages/torch_geometric/nn/conv/gcn_conv.py", line 211, in forward edge_index, edge_weight, x.size(self.node_dim), TypeError: 'int' object is not callable

In my case, when I was creating the hetero-data I was mistakenly setting the node feature tensors x as numpy arrays instead of torch tensors.
The HeteroData class seemed to accept that, but then in the call at gcn_conv.py", line 211, the size() of torch::Tensor is called.
And in numpy arrays the member size is an int indicating the total number of elements in the array. Hence, the error 'int' object is not callable.
For my case the fix was simple, I just had to change the numpy arrays to torch Tensors :-) (torch.from_numpy did the job)