Which version of torch_geometric is needed? about rshn HOT 13 CLOSED

The 'RelationConv' is implemented in the fold ./torch_geometric/nn/conv, and 'Entities' can be imported from torch_geometric/dataset. Not from the original package.

from rshn.

Yeah, I see the code. But I am still confused about how to modify the code to make it run?

from rshn.

You can specify the import from the local package.

from rshn.

Yes, I tried to import from local. But failed. I also tried to google "how to import from local" but no available knowledge. Forgive my ignorance. [sad]

from rshn.

Maybe you could try PyCharm to deploy the code, which is friendly to you to make it run.

from rshn.

Ah... It is a weird requirement... So is there no way to run it just in terminal? (What if I run code on a Linux server?)

from rshn.

Of course the code is running on Linux server, through configuring an interpreter using SSH.

from rshn.

Yes, I use PyCharm now and it seems better. But there is still one problem: Could you specify the version of torch_sparse because now the code raises this error:

    out = spmm(edge_index, alpha, x.size(0), x)
TypeError: spmm() missing 1 required positional argument: 'matrix'

I think it's due to the version of torch_sparse mismatching.

If you could provide the version of the following four packages, it's better:

pip install torch-scatter
pip install torch-sparse
pip install torch-cluster
pip install torch-spline-conv

from rshn.

torch-scatter == 1.3.0
torch-sparse == 0.4.0
torch-cluster == 1.4.2
torch-spline-conv == 1.0.4

from rshn.

Apologize for disturbing you again. Which version of CUDA are you using? I use the default cuda9.0 for pytorch 1.1. But I fail to install torch-sparse==0.4.0.

from rshn.

Hi I made some modification and a pull request. Now the code can run with the latest version of torch-scatter and torch-sparse. (with torch 1.4.0 and cuda 10.1)

Maybe it's not necessary to merge my request, just provide for others' convenience.

from rshn.

BTW, two lines of code can be added to the head of RSHN.py get away with the necessity of PyCharm:

import sys
sys.path.insert(0, '../')

from rshn.

Thanks for the improvement, and I will consider the modification with the latest version.

from rshn.