TypeError: BasicGNN.forward() got an unexpected keyword argument 'return_attention_weights'

[Batene]

🐛 Describe the bug

In GATv2Conv, return_attention_weights is expected to be passed as a parameter to the forward function.
When using Gatv2Conv indirectly by using GAT with v2=True, I cannot pass return_attention_weights=True when calling the GAT model (= using the forward function). This will throw the error above.
One can pass return_attention_weights=True in the GAT initalization. But then, this will never be used, as the forward function in GATv2Conv does not access self.return_attention_weights (it does not exist)). The forward function requires the parameter return_attention_weights instead of accessing a class attribute.

My solution for now is:

1. add in the __init__ function of Gatv2Conv:

    `self.return_attention_weights = kwargs["return_attention_weights"]`
   

2. in the forward function of Gatv2Conv:

   pytorch_geometric/torch_geometric/nn/conv/gatv2_conv.py

   Line 312 in 38bb5f2

   if isinstance(return_attention_weights, bool):

    `if isinstance(self.return_attention_weights, bool):`
   

I am using pytorch geometric version 2.5.2.
I assume, the same problem occurs when using v2=False (with GatConv).

Versions

Python version: 3.11.7

[rusty1s]

Are you referring to models.GAT? If you need to return attention weights, I suggest to use the GNN layers and build your own model on top.