Cannot define a GATv2Conv layer for a Bipartite graph

[michalisfrangos]

Hello,

From the description of GATConv and GATv2Conv they both support bipartite graphs and the layers can be instantiated in the same way. I cannot define a layer for a Bipartite graph with GATv2Conv. I was wondering if this is a bug in GATv2Conv?
Please see additional info below.

Thanks
M

Example with GATConv (all good):

GATConv(in_channels = (10,10),
          out_channels = 32,
          heads = 1, concat=True)

gives a layer GATConv((10, 10), 32, heads=1)

Example with GATv2Conv (error):

GATv2Conv(in_channels = (10,10),
          out_channels = 32,
          heads = 1, concat=True)

gives an error:
`---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
~\AppData\Local\Temp/ipykernel_12336/3079787839.py in
----> 1 GATv2Conv(in_channels = (10,10),
2 out_channels = 32,
3 heads = 1, concat=True)

~\miniconda3\envs\env_gnn\lib\site-packages\torch_geometric\nn\conv\gatv2_conv.py in init(self, in_channels, out_channels, heads, concat, negative_slope, dropout, add_self_loops, bias, share_weights, **kwargs)
80 self.share_weights = share_weights
81
---> 82 self.lin_l = Linear(in_channels, heads * out_channels, bias=bias,
83 weight_initializer='glorot')
84 if share_weights:

~\miniconda3\envs\env_gnn\lib\site-packages\torch_geometric\nn\dense\linear.py in init(self, in_channels, out_channels, bias, weight_initializer, bias_initializer)
47 self.bias_initializer = bias_initializer
48
---> 49 if in_channels > 0:
50 self.weight = Parameter(torch.Tensor(out_channels, in_channels))
51 else:

TypeError: '>' not supported between instances of 'tuple' and 'int'`

[rusty1s]

Which version of PyG are you using (torch_geometric.__version__)? This should be already fixed in version 2.0.3.

[michalisfrangos]

got it! I was using 2.0.1.... all good in version 2.0.3.