Torch-geometric

When I set batch size to 512, an error is reported:
RuntimeError: nonzero is not supported for tensors with more than INT_MAX elements, file a support request

But when I set batch size to 256, it will run, why?

Report error code at :

def dense_to_sparse(tensor):
    r"""Converts a dense adjacency matrix to a sparse adjacency matrix defined
    by edge indices and edge attributes.

    Args:
        tensor (Tensor): The dense adjacency matrix.
     :rtype: (:class:`LongTensor`, :class:`Tensor`)
    """
    assert tensor.dim() == 2
    index = tensor.nonzero(as_tuple=False).t().contiguous()
    value = tensor[index[0], index[1]]
    return index, value

Thank you very much for your reply

1 Like

Could you create an issue on GitHub and explain your use case there?
Based on the error message it seems there is an internal limitation for nonzero using INT_MAX elements.

I have met the same error:
File “/home/lib/python3.6/site-packages/torch_geometric/utils/sparse.py”, line 18, in dense_to_sparse
index = tensor.nonzero().t().contiguous()
RuntimeError: nonzero is not supported for tensors with more than INT_MAX elements, file a support request

When I set batch size to 20, it will run, but when I set batch size to 40 or 30, the error will be reported.

So I wonder if you have solved this problem?