[FEA] pytorch geometric interface

[jannessm]

I am currently using Pytorch geometric for deep learning on graphs. It supports way more convolutions than DGL. So I was wondering if there is a plan to support an interface of converting graphs to a cugraph graph and vice versa.

Regards.

[BradReesWork]

@jannessm great question. I need to dive into pytorch format before fully answering the questions. The roadmap is currently Pandas and CuPy this year.

[flandolfi]

Hi @jannessm,

I am working with PyTorch-Geometric (PyG) and for the moment I'm using the following solution (only regards PyG to cuGraph for the moment). Maybe it can be of some help.

from torch.utils.dlpack import from_dlpack, to_dlpack
from torch_geometric.data import Data
from torch_geometric.utils import add_self_loops

import cudf
import cugraph as cx


def to_cugraph(data: Data):
    edge_index, weights = add_self_loops(data.edge_index, data.edge_attr, num_nodes=data.num_nodes)
    df = edge_index.contiguous().t().clone().detach()
    df = cudf.from_dlpack(to_dlpack(df))

    if weights is not None:
        weights = weights.contiguous().view(-1).clone().detach()
        df[2] = cudf.from_dlpack(to_dlpack(weights))

    return cx.from_cudf_edgelist(df, source=0, destination=1,
                                 edge_attr=2 if df.shape[1] == 3 else None,
                                 renumber=False)

Adding self loops is needed to handle the issue #1206 I opened a while ago.

@BradReesWork the conversion cuDF/PyTorch seem to work using DLPack, but we need to take care of some details, like the row-major format and the case in which a tensor is non-contiguous.

Anyway, probably the conversion from and to PyG should be done from the PyG side (paging @rusty1s :)).

Regards,
Francesco

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.

[rusty1s]

This issue can likely be closed. I will work on cugraph conversion on PyG side.
@flandolfi Let me know if you have any plans to contribute your implementation.

[flandolfi]

@flandolfi Let me know if you have any plans to contribute your implementation.

Sure, I will make a PR in a week or so!

[github-actions]

This issue has been labeled inactive-30d due to no recent activity in the past 30 days. Please close this issue if no further response or action is needed. Otherwise, please respond with a comment indicating any updates or changes to the original issue and/or confirm this issue still needs to be addressed. This issue will be labeled inactive-90d if there is no activity in the next 60 days.