How to use the size of the input of GNN layers for in_channels?

[CTY]

I am trying to use TransformerConv to train a model which takes in graphs of different sizes. This is to be used on a graph classification task. My code snippet is as follows:

self.conv1 = TransformerConv(in_channels=-1, out_channels=256)

def forward(self, x_onein, x_twoin, lambd):
        x_one, edge_index_one, batch_one = x_onein.x, x_onein.edge_index, x_onein.batch
        x_one = self.conv1(x_one, edge_index_one)
        return x_one

My input data come in batches of various sizes, and each graph has between tens of nodes to 1200 nodes. Because of this, I require the first input layer to be able to take in graphs of different sizes.
When I run the code, I get the error:

RuntimeError: Trying to create tensor with negative dimension -1: [256, -1]

This is strange since I followed the documentation:
in_channels (int or tuple) – Size of each input sample, or -1 to derive the size from the first input(s) to the forward method. A tuple corresponds to the sizes of source and target dimensionalities.

Any advice on this matter would be greatly appreciated.

[ptrblck]

I can’t reproduce the issue using:

x1 = torch.randn(4, 8)
x2 = torch.randn(2, 16)
edge_index = torch.tensor([[0, 1, 2, 3], [0, 0, 1, 1]])  
conv = TransformerConv(in_channels=-1, out_channels=256)
out = conv(x1, edge_index)

so maybe your torch_geometric version is too old to support the lazy initialization of the layers (I don’t know when it was introduced)?

[CTY]

Thank you for the prompt reply.

My Pytorch version is 1.7.1 and my Pytorch geometric version is 1.6.3. I shall try to update them to the latest versions and try again.

[ptrblck]

I’ve executed my code snippet with PyTorch 1.10.1 and Geometric 2.0.3, so let’s see if updating helps.

[CTY]

Thank you for the help. Updating did help and I am able to run the program my desired way now. Lazy inputs seem to be a more recent functionality.