About proxies in forward pass

[gabriom6]

Hi! I'm currently implementing a custom heterograph gnn and I found something that I don't understand which shows up while performing a forward pass with a heterodata object.

Concretely, I'm trying to print the network output inside a training loop. Turns out that instead of printing the output tensor as I would expect in a normal torch network, it only prints Proxy(getattr_1) at the first iteration of the loop and then it stops for the rest of the remaining iterations.

A simplification of the code is as follows:

`
class gnn(torch.nn.Module):
def init(self):
super(gnn, self).init()
self.gat = torch_geometric.nn.conv.GATv2Conv(in_channels=(-1,-1), out_channels=4, heads=2, concat=False, add_self_loops=False)

def forward(self, x, edge_index):
    print('forward')
    x = self.gat(x=x, edge_index=edge_index)
    print('x',x['node1'].x)
    return x


for _ in range(3):
    model(heterograph.x_dict, heterograph.edge_index_dict)

`

The output prints this:

forward
x Proxy(getattr_1)

Whereas I was expecting something more like this:

forward
x node1 attributes tensor...

forward
x node1 attributes tensor...

forward
x node1 attributes tensor...

Can someone explain me what is this proxy thing for and why it does not work as a normal torch forward pass?

Thank you very much for your help :)