CGConv

[mxv001]

Hello all,
Is there anywhere an example of how to use the CGConv layer? I have read the paper, but there are some non-trivial differences between author code and py-g implementation. Looking for something like below:

import torch
import torch.nn.functional as F
from torch_geometric.nn import CGConv

class CGCNN(torch.nn.Module):
    def __init__(self, dataset):
        super().__init__()
        self.conv1 = CGConv((channel_variable1, channel_variable2), dim=dataset.num_edge_features)
        self.conv2= CGConv((channel_variable3, channel_variable4), dim=dataset.num_edge_features)

    def forward(self, data):
        x, edge_index, edge_attr = data.x, data.edge_index, data.edge_attr

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

In particular I do not understand

1. the usage of the "channels" variables in creating the layer (see the code) and how they match to graph dimensions
   (for example: does channel_variable2 need to equal channel_variable3? Does channel_variable4 = the overall output of the model?)
2. An example of a single graph that can be put through forward function

For reference, I looked at the questions here: #2859

But the answer is not posted (even though it is solved).

Thank you all very much!

[rusty1s]

1. Yes, the out_channels (the number of features you output for each node), needs to match with the in_channels of the follow-up GNN layer.
2. Your code snippet should work once you fix x = self.conv2(x) to be x = self.conv2(x, edge_index, edge_attr).

Happy to help with any follow-up questions. We sadly do not have an example of every GNN layer we provide in PyG :(

[mxv001]

Thank you for your speedy reply! And thanks for having so many cool layers -- can't always have a tutorial for advanced usage!

I am getting an error when I try to use this -- perhaps it is due to batch dimension? Here is the error:

File ~/anaconda3/envs/GNN/lib/python3.10/site-packages/torch/nn/modules/module.py:1501, in Module._call_impl(self, *args, **kwargs)
   1496 # If we don't have any hooks, we want to skip the rest of the logic in
   1497 # this function, and just call forward.
   1498 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
   1499         or _global_backward_pre_hooks or _global_backward_hooks
   1500         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1501     return forward_call(*args, **kwargs)
   1502 # Do not call functions when jit is used
   1503 full_backward_hooks, non_full_backward_hooks = [], []

Cell In[91], line 28, in CGConvNN.forward(self, data)
     26 x, edge_index, edge_attr = data.x, data.edge_index, data.edge_attr
     27 batch = data.batch
---> 28 x = self.conv1(x, edge_index, edge_attr)
     29 x = self.act1(x)

File ~/anaconda3/envs/GNN/lib/python3.10/site-packages/torch/nn/modules/module.py:1501, in Module._call_impl(self, *args, **kwargs)
   1496 # If we don't have any hooks, we want to skip the rest of the logic in
   1497 # this function, and just call forward.
   1498 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
   1499         or _global_backward_pre_hooks or _global_backward_hooks
   1500         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1501     return forward_call(*args, **kwargs)
   1502 # Do not call functions when jit is used
   1503 full_backward_hooks, non_full_backward_hooks = [], []

File ~/anaconda3/envs/GNN/lib/python3.10/site-packages/torch_geometric/nn/conv/cg_conv.py:88, in CGConv.forward(self, x, edge_index, edge_attr)
     85     x: PairTensor = (x, x)
     87 # propagate_type: (x: PairTensor, edge_attr: OptTensor)
---> 88 out = self.propagate(edge_index, x=x, edge_attr=edge_attr, size=None)
     89 out = out if self.bn is None else self.bn(out)
     90 out = out + x[1]

File ~/anaconda3/envs/GNN/lib/python3.10/site-packages/torch_geometric/nn/conv/message_passing.py:469, in MessagePassing.propagate(self, edge_index, size, **kwargs)
    467     if res is not None:
    468         msg_kwargs = res[0] if isinstance(res, tuple) else res
--> 469 out = self.message(**msg_kwargs)
    470 for hook in self._message_forward_hooks.values():
    471     res = hook(self, (msg_kwargs, ), out)

File ~/anaconda3/envs/GNN/lib/python3.10/site-packages/torch_geometric/nn/conv/cg_conv.py:98, in CGConv.message(self, x_i, x_j, edge_attr)
     96 else:
     97     z = torch.cat([x_i, x_j, edge_attr], dim=-1)
---> 98 return self.lin_f(z).sigmoid() * F.softplus(self.lin_s(z))

File ~/anaconda3/envs/GNN/lib/python3.10/site-packages/torch/nn/modules/module.py:1501, in Module._call_impl(self, *args, **kwargs)
   1496 # If we don't have any hooks, we want to skip the rest of the logic in
   1497 # this function, and just call forward.
   1498 if not (self._backward_hooks or self._backward_pre_hooks or self._forward_hooks or self._forward_pre_hooks
   1499         or _global_backward_pre_hooks or _global_backward_hooks
   1500         or _global_forward_hooks or _global_forward_pre_hooks):
-> 1501     return forward_call(*args, **kwargs)
   1502 # Do not call functions when jit is used
   1503 full_backward_hooks, non_full_backward_hooks = [], []

File ~/anaconda3/envs/GNN/lib/python3.10/site-packages/torch/nn/modules/linear.py:114, in Linear.forward(self, input)
    113 def forward(self, input: Tensor) -> Tensor:
--> 114     return F.linear(input, self.weight, self.bias)

RuntimeError: mat1 and mat2 shapes cannot be multiplied (3546x21 and 76x64)

The model is very similar to the one above:

import torch
import torch.nn.functional as F
from torch_geometric.nn import CGConv

class CGCNN(torch.nn.Module):
    def __init__(self, dataset):
        super().__init__()
        self.conv1 = CGConv((dataset.num_node_features, hidden_channels), dim=dataset.num_edge_features)
        self.act1 = LeakyReLU(hidden_channels,1)
    def forward(self, data):
        x, edge_index, edge_attr = data.x, data.edge_index, data.edge_attr

        x = self.conv1(x, edge_index,edge_attr)
        x = self.act1(x)
        return x

The dataset has:
dataset.num_node_features = 7 ,
dataset.num_edge_features = 4.

My batching has 64 graphs per batch. And the hidden_channels = 64.

Thank you again for your help!

[rusty1s]

The issue is that you are initializing CGConv to expect different input feature dimensionalities for source and destination nodes (i.e., it expects to work in a bipartite graph where destination nodes have a feature dimensionality of 32). Fixable by just calling

self.conv1 = CGConv(dataset.num_node_features,
                    dim=dataset.num_edge_features)

[mxv001]

Thanks this helps!