Error in Heterogeneous graph by applying the PYG model

[amalislam675]

I try to apply the GCN2Conv model within HeteroGNN, I found this below mentioned error, can you tell how to fix? With the same code, If I apply GCNConv from PYG, it works fine. Here, need to add initial features, that's why, I am giving the x_dict two times as residual feature.
import numpy as np
import torch
from torch_geometric.data import HeteroData

Random data just to show how store values of nodes work

authors = torch.rand((10,8))
papers = torch.rand((10,8))
authors_y = torch.rand(10)

Random data just to show how store values of edges work

write_from = torch.tensor(np.random.choice(10, 15, replace = True))
write_to = torch.tensor(np.random.choice(10, 15, replace=True))
write = torch.concat((write_from, write_to)).reshape(-1,15).long()

Random dat justo to show how store values of edges work

cite_from = torch.tensor(np.random.choice(10, 15, replace=True))
cite_to = torch.tensor(np.random.choice(10, 15, replace=True))
cite = torch.concat((cite_from, cite_to)).reshape(-1,15).long()

edge_weight1 = torch.rand(15)
edge_weight1 = edge_weight1.unsqueeze(-1)
edge_weight2 = torch.rand(15)
edge_weight2 = edge_weight2.unsqueeze(-1)

from torch_geometric.data import HeteroData

data = HeteroData()
data['author'].x = authors
data['paper'].x = papers

data['author', 'writes', 'author'].edge_index=write
data['paper', 'cited_by', 'paper'].edge_index=cite

data['author', 'writes', 'author'].edge_weight= edge_weight1
data['paper', 'cited_by', 'paper'].edge_weight= edge_weight1

print(data)

import torch
import torch_geometric.transforms as T
from torch_geometric.nn import HeteroConv, GCN2Conv, Linear
class HeteroGNN(torch.nn.Module):
def init(self, hidden_channels, out_channels):
super().init()

    self.lin1=torch.nn.Linear(8, hidden_channels)
    self.conv1 = HeteroConv({
            ('author', 'writes', 'author'): GCN2Conv(hidden_channels, alpha=0.5, theta=0.5, layer=1, cached=True),
            ('paper', 'cited_by', 'paper'): GCN2Conv(hidden_channels, alpha=0.5, theta=0.5, layer=1, cached=True)
        }, aggr='sum')
    print(self.conv1)
    self.conv2 = HeteroConv({
            ('author', 'writes', 'author'): GCN2Conv(hidden_channels, alpha=0.5, theta=0.5, layer=2, cached=True),
            ('paper', 'cited_by', 'paper'): GCN2Conv(hidden_channels, alpha=0.5, theta=0.5, layer=2, cached=True)
        }, aggr='sum')

    self.lin = Linear(hidden_channels*2, out_channels)

def forward(self, x_dict, edge_index_dict, edge_weight_dict):
    x_dict = {key: self.lin1(x) for key, x in x_dict.items()}
    h = self.conv1(x_dict, x_dict, edge_index_dict, edge_weight_dict)
    h = {key: x.relu() for key, x in h.items()}
    
    
    h = self.conv2(h, x_dict, edge_index_dict, edge_weight_dict)
    h = {key: x.relu() for key, x in h.items()}
    
    out = h["author"]
    return out

Error
Traceback (most recent call last):
File "/123/PYG/example1_1.py", line 87, in
out = model(data.x_dict, data.edge_index_dict, data.edge_weight_dict)
File "/123/PYG/.conda/envs/foo/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "/123/PYG/example1_1.py", line 67, in forward
h = self.conv1(x_dict, x_dict, edge_index_dict, edge_weight_dict)
File "/123/pyg/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1130, in _call_impl
return forward_call(*input, **kwargs)
File "/123/pyg/lib/python3.10/site-packages/torch_geometric/nn/conv/hetero_conv.py", line 106, in forward
src, rel, dst = edge_type
ValueError: too many values to unpack (expected 3)

[rusty1s]

You need to call via

h = self.conv1(x_dict, edge_index_dict, edge_weight_dict)

[rusty1s]

Oh, I see. I missed that GCN2Conv expects two feature tensors. Let me confirm in a test :)

[rusty1s]

Will be fixed in #8166.

[amalislam675]

@rusty1s , than you for your consideration. As you mentioned, it will fixed in #8166. Through which way, I will use this updated hetroconv module. I have installed PYG 2.3.1 in my system. Do I need to use this code from source by separately running HetroConv module?

[rusty1s]

At best, install PyG from GitHub/nightly (see README for more details).

[amalislam675]

@rusty1s , thanks, I try to intall.