Getting Error for GAT message function

[qm-intel]

Hi,

I am running a code for GAT class using MessagePassing layer

class GATConv(MessagePassing):
    """
    Implementation of GAT using message passing layers
    Link: https://github.com/njmarko/machine-learning-with-graphs/blob/master/MyAttempts/CS224W_Colab4.ipynb
    """
    def __init__(self, emb_dim, heads=2, negative_slope=0.2, aggr = "add"):
        super(GATConv, self).__init__()

so the message function is as follows:

    def message(self, edge_index, x_i, x_j, edge_attr):
        """

        :param edge_index: shape [2, 4371]
        :param x_i: shape [1439, 4371, 300]
        :param x_j: shape [1439, 4371, 300]
        :param edge_attr: shape [4371, 600]
        :return:
        """
        
        edge_attr = edge_attr.view(-1, self.heads, self.emb_dim)  # [4252, 2, 300]

        x_j += edge_attr

        alpha = (torch.cat([x_i, x_j], dim=-1) * self.att).sum(dim=-1)

        alpha = F.leaky_relu(alpha, self.negative_slope)
        alpha = softmax(alpha, edge_index[0])

        return x_j * alpha.view(-1, self.heads, 1)

I am getting the following error:

  File "/home/user/anaconda3/envs/pyg-meta/lib/python3.10/site-packages/torch_geometric/nn/conv/message_passing.py", line 374, in propagate
    out = self.message(**msg_kwargs)
  File "/home/user/model.py", line 209, in message
    x_j += edge_attr
RuntimeError: The size of tensor a (2949) must match the size of tensor b (2) at non-singleton dimension 1

How can I resolve this?

• My other question is what is the difference between when we define a class based on MessagePassing module and nn.Module?

[rusty1s]

I assume the shapes of x_j and edge_attr need to match, but they are [4252, 2, 300] and [1439, 4371, 300].

My other question is what is the difference between when we define a class based on MessagePassing module and nn.Module?

MessagePassing inherits from torch.nn.Module and provides the capability to perform message passing propagation via propagate, message and aggregate. You can use torch.nn.Module as well, but then you can no longer make use of the MessagePassing interface.