Graclus weight

[mapapin]

Hi,
I'm trying to do a cluster by edge attribute using graclus function but it returns

torch.cumsum(deg, 0, out=rowptr[1:])

    return torch.ops.torch_cluster.graclus(rowptr, col, weight)
           ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
RuntimeError: optional_weight.value().dim() == 1INTERNAL ASSERT FAILED at "csrc/cpu/graclus_cpu.cpp":12, please report a bug to PyTorch. Input mismatch

My edge attribute have a shape of [num_edges, 1] and edge_index [2, num_edges] and i call graclus like that :
cluster = graclus(data.edge_index, weight=data.edge_attr, num_nodes=data.x.size(0))

Any suggestions are welcome :)

And then, not related to the error, what am I supposed to have in the pos attribute of a Data object?

Thanks !

[wsad1]

Is the shape of data.edge_attr [num_edges,1]. graclus only supports a 1-d weight argument.

[rusty1s]

METIS does not have a GPU implementation. There are also some tricks involved to convert edge_weight internally since METIS can only handle discrete edge weights. If you have a reproducible example, I am happy to look into this.

[mapapin]

Oh ok, that's why my model gets stuck in graclus ? My edge_attr are equivalent to euclidean distance between nodes.

I'm a newbie, maybe this isn't the way to do it but I wanted to make graclus create a cluster based on the closest nodes in space for pooling purpose.
I have my Data object : DataBatch(x=[30622, 1], edge_index=[2, 258078], edge_attr=[258078, 1], tag=[64], ids=[30622], num_nodes=30622, batch=[30622], ptr=[65] and i send data to graclus like that : cluster = graclus(data.edge_index, weight=data.edge_attr.view(-1), num_nodes=data.x.size(0)) with data previously sent to my convolution.

All my Data objects are sent to GPU before i start training.

I hope that's enough to see what i'm trying to do...

Thanks :)

[rusty1s]

This shouldn't be a problem. We auto-convert to CPU inside METIS, see here.

[mapapin]

It's weird, when I send my data to the GPU it doesn't work, whereas when I send it to the CPU and it works.

Is it possible that the conversion takes a very long time even if i send a very small amount of data (10 samples) ?

NB: If I remove graclus, my code works with GPU, it's seems really stuck in the function.

[rusty1s]

Can you do some debugging on your end, e.g., where exactly it gets stuck in the metis.py function in torch-sparse?