Pair of Hetero Graphs

[tehranixyz]

Hi,
I'm having difficulties creating a HeteroData object that stores two heterogeneous graphs.

The batching page provides nice details on how to create a new class based on the Data class to store 2 graphs and also overwrite the inc function to ensure the batching is working properly.
However, in my case, I have pairs of heterogeneous graphs that I want to predict their similarity.
I have around 200K pairs of graphs for each pair I have label 1 or 0 showing that the two graphs are similar or not.
The node types and edge types are the same across the graphs. So all graphs have the same node types and edge types.
Since my heterogeneous graphs are not very large, I intend to use DataLoader for batching.
However, I don't know how to override cat_dim and inc for my HeteroData-based class.
Let's say for the first graph we have:

data['nodes'].x_a = ...
data['node', 'control', 'node'].edge_index_a = ...
data['node', 'data', 'node'].edge_index_a = ...
data['node', 'call', 'node'].edge_index_a = ...
data['node', 'control', 'node'].edge_attr_a = ...
data['node', 'data', 'node'].edge_attr_a = ...
data['node', 'call', 'node'].edge_attr_a = ...

For the second graph, we have

data['nodes'].x_b = ...
data['node', 'control', 'node'].edge_index_b = ...
data['node', 'data', 'node'].edge_index_b = ...
data['node', 'call', 'node'].edge_index_b = ...
data['node', 'control', 'node'].edge_attr_b = ...
data['node', 'data', 'node'].edge_attr_b = ...
data['node', 'call', 'node'].edge_attr_b = ...

and the label is:
data.y = torch.tensor([1])

How can I create a class based on HeteroData and store this information?
Thanks

[rusty1s]

I think you should avoid integrating PairData with HeteroData objects, this is likely to become a nightmare to maintain. Instead, you should have a PairDataset that returns a tuple of HeteroData objects:

class MyDataset(...):
    ...

    def __getitem__(self, idx):
        return self.dataset1[idx], self.dataset2[idx]

You can then use this dataset as part of torch_geometric.loader.DataLoader:

loader = DataLoader(my_dataset, ...)

for batch1, batch2 in loader:
    pass

WDYT?

[tehranixyz]

That worked flawlessly. Thanks for the suggestion!

[sssleder]

I'm glad this got answered, but I'm stubborn and want to do it the hard way. Mostly so that I can understand the collating that is going on with the HeteroData() class but also because I spent so much time working on this functionality before I found this discussion.

I'm stuck at the same issues with modifying HeteroData()'s __inc__() function for batching to properly collate with the following error:

File ~/.conda/envs/from_pygcu/lib/python3.11/site-packages/torch_geometric/data/collate.py:322, in <listcomp>(.0)
    319 def get_incs(key, values: List[Any], data_list: List[BaseData],
    320              stores: List[BaseStorage]) -> Tensor:
    321     repeats = [
--> 322         data.__inc__(key, value, store)
    323         for value, data, store in zip(values, data_list, stores)
    324     ]
    325     if isinstance(repeats[0], Tensor):
    326         repeats = torch.stack(repeats, dim=0)

File ~/.conda/envs/from_pygcu/lib/python3.11/site-packages/torch_geometric/data/hetero_data.py:349, in HeteroData.__inc__(self, key, value, store, *args, **kwargs)
    347     return int(value.max()) + 1
    348 elif isinstance(store, EdgeStorage) and 'index' in key:
--> 349     return torch.tensor(store.size()).view(2, 1)
    350 else:
    351     return 0

RuntimeError: Could not infer dtype of NoneType

For some reason size() of EdgeStorage gives me [None, None], I figure it is a side affect num_nodes not being properly set, but a nudge in the right direction would be helpful. I figure there is some way to associate an EdgeStorage instance with a NodeStorage instance...

As a random aside: Ali if you see this it's been a very long time and I'm glad you figured out how to parse ASTs for hetero GNNs. Also please give Dr. Jannesari my apologies for disappearing after the Euromicro paper.