Input size change after fed into HGT

[Gori-LV]

Hi, I'm trying out HGT but having a problem of copied input size change. Below is the code I add to the end of this.

import copy

x_dict, edge_index_dict = data.x_dict, data.edge_index_dict
copied_x_dict, copied_edge_index_dict = copy.deepcopy(data.x_dict), copy.deepcopy(data.edge_index_dict)

print(data.x_dict['author'].size())
print(x_dict['author'].size())
print(copied_x_dict['author'].size())

model.eval()
with torch.no_grad():
    _ = model(x_dict, edge_index_dict)
    _ = model(data.x_dict, data.edge_index_dict)
    _ = model(copied_x_dict, copied_edge_index_dict)

print(data.x_dict['author'].size())
print(x_dict['author'].size())
print(copied_x_dict['author'].size())

then got this output

torch.Size([4057, 451])
torch.Size([4057, 451])
torch.Size([4057, 451])
torch.Size([4057, 451])
torch.Size([4057, 64])
torch.Size([4057, 64])

Not sure where is the problem :( can someone help? Thx a lot!

[rusty1s]

I can't seem to reproduce this, see this newly added test. Which PyG version are you using?

[Gori-LV]

I tested the newly added test and got no error either, cuz HGTConv is not the problem. I figure that this is the cause, when x_dict copied from data.x_dict is passed to the function, it is modified. But if data.x_dict is directly passed, it will fine (can you explain why?).

I add x_dict = copy.deepcopy(x_dict) in the forward() function before x_dict is processed, the issue is solved.

[Gori-LV]

I suppose this may cause a waste of memory, is there a better way to do it? :(

[rusty1s]

Ah, I see. You are talking about HGT rather than HGTConv. This is easily fixable by replacing it with:

for node_type, x in x_dict.items():
    x_dict[node_type] = self.lin_dict[node_type](x).relu_()

with

x_dict = {key: self.lin_dict[key](x).relu_() for key, x in x_dict.items()}

[Gori-LV]

That's cool, thanks alot!