Building a zero-diagonal adjacency matrix sample-wise from bipartite node's edge_index

[ndrwnaguib]

Hello,

I have a problem concatenatingedge_index and edge_attr across batches, I have a custom impl that works only after I have loaded the batch. Here is the function I use:

def _pad_adj(self, edge_index, edge_attr, num_left_feat, num_right_feat, fill=-9e15):
    row, col = edge_index
    edge_attr = edge_attr.reshape(-1)
    row_features = sp.coo_matrix(
        (edge_attr, (row, col)),
        (num_right_feat, num_left_feat),
    )

    Axb = row_features.todense().T
    num_nodes = num_left_feat + num_right_feat
    adj = fill * np.ones((num_nodes, num_nodes))
    adj[:num_vars, num_vars:] = Axb
    adj[num_vars:, :num_vars] = Axb.T

    assert (adj[:num_left_feat, :num_left_feat] == fill).all(), "adjacency is ill-constructed"
    assert (adj[num_left_feat:, num_left_feat:] == fill).all(), "adjacency is ill-constructed"

    adj = torch.FloatTensor(adj)
    return adj.to_sparse_coo()

The above implementation produces the intended results only if edge_index and edge_attr were passed after the batch was loaded. But, this is too slow to include in a training step, so I am trying to achieve the same results by applying the same function sample-wise in the get(...) function in the Dataset

    adj = self._pad_adj(
        edge_index=edge_indices,
        edge_attr=edge_features,
        num_left_feat=left_features.shape[0],
        num_right_feat=right_features.shape[0],
        fill=0,
    )

    graph = BipartiteNodeData(
        edge_index=adj.indices(),
        edge_attr=adj.values(),
    )

However, results are not the same,

Here are the indices() and edge_index from both implementations

>>> adj.indices()
tensor([[    0,     0,     0,  ..., 31247, 31247, 31247],
        [15516, 15517, 15592,  ..., 14718, 15307, 15417]])
>>> edge_index
tensor([[    0,     0,     0,  ..., 31247, 31247, 31247],
        [ 2118,  2119,  2194,  ..., 28883, 29472, 29582]])

Although it may seem that edge_index[0] values are equal to those of adj.indices()[0], they're actually not.

Note:

_pad_adj basically adds zeros to the diagonal and puts the features on the right top quadrant of the matrix and its transpose to the left bottom quadrant.

Any thoughts? Thanks.

[rusty1s]

Hey, I answered you on Slack as well, but just for completeness:

Batch-wise adjacency matrices are stacked diagonally, so converting your batched adjacency matrix to a dense one and then applying modifications to adj[:num_vars, num_vars:] will not work as intended.

I am also wondering why you do

adj = fill * np.ones((num_nodes, num_nodes))

since this will result in a sparse matrix with 0% sparsity.