Issue in Retaining the Origninal Shape of Adjacency Matrix after Conversion from Dense to Sparse

[srikesh-07]

Hello Community,

Thank you for intending to help me.

I am working on a custom bipartite dataset that seems good until I meet this issue. The issue is while reconstructing the edge indices from dense matrix of Data instances by the DataLoader

BipartiteGraph(x_constraints=[180, 1], edge_index=[2, 8130], edge_attr=[8130, 1], x_variables=[255, 9], num_nodes=435)
BipartiteGraph(x_constraints=[410, 1], edge_index=[2, 31510], edge_attr=[31510, 1], x_variables=[510, 9], num_nodes=920)

The Batch of the above instance looks like,

BipartiteGraphBatch(x_constraints=[590, 1], x_constraints_batch=[590], x_constraints_ptr=[3], edge_index=[2, 39640], edge_attr=[39640, 1], x_variables=[765, 9], x_variables_batch=[765], x_variables_ptr=[3], num_nodes=1355, batch=[1355], ptr=[3])

I am trying to convert the sparse adjacency matrix to a dense adjacency matrix and finally reconstruct the sparse adjacency matrix using the recently converted dense adjacency matrix.

I have used the methods to_dense_adj and dense_to_sparse from torch_geometric.utils package. The conversion output seems to be a bit sketchy,

Sparse Shape: [2, 39640]
Dense Shape: [2, 920, 920]
Reconstructing Sparse with Above Dense Matrix: [2, 39566]

We can see some edges are being missed in the newly constructed sparse adjacency matrix. I didn't find the reason. Can anyone help me to fix the issue or give me the reason for the above issue? Thank you for reading and the help from the community will be much appreciated.

Custom Data function code is attached below,

class BipartiteGraph(Data):
    masked_cons_idx: int
    bias_label: FloatTensor
    degree_label: LongTensor
    logits_label: FloatTensor
    connected_vars_idx: LongTensor
    weights_label: FloatTensor

    def __init__(
        self,
        x_constraints: FloatTensor = None,
        edge_index: LongTensor = None,
        edge_attr: FloatTensor = None,
        x_variables: FloatTensor = None,
        **kwargs
    ):
        """
        Bipartite graph data object, each representing a MILP instance.
        """
        super().__init__(**kwargs)
        self.x_constraints = x_constraints
        self.edge_index = edge_index
        self.edge_attr = edge_attr
        self.x_variables = x_variables
        if x_constraints is not None:
            self.num_nodes = len(x_constraints) + len(x_variables)

    def __inc__(self, key, value, store, *args, **kwargs):
        if key == "edge_index":
            return torch.tensor(
                [[self.x_constraints.size(0)], [self.x_variables.size(0)]]
            )
        if key == "masked_cons_idx":
            return self.x_constraints.size(0)
        elif key == "connected_vars_idx":
            return self.x_variables.size(0)
        else:
            return super().__inc__(key, value, *args, **kwargs)

    @property
    def num_variables(self):
        return len(self.x_variables)

    @property
    def num_constraints(self):
        return len(self.x_constraints)

Again Thank You !!

[rusty1s]

I assume you hold some duplicated edges in edge_index, which should explain this. Can you confirm?

[srikesh-07]

Thank you for your suggestion @rusty1s !!!.

I already tried to verify the edge indices of each instance in Dataset using the coalesce method. There are no duplicated edges,

Also, I am facing this reconstruction issue while using the Batch instance and not the single Data instance.

PS: Also tried with the RemoveDuplicateEdges, but no changes !!

[rusty1s]

Thanks for checking. Does edge_index hold the constraint of

assert edge_index[0].max() < x_constraints.size(0)
assert edge_index[1].max() < x_variables.size(0)