Pytorch_geometric: torch-cluster has no CUDA version, as mentioned in the installation instructions of PyG

[douglasrizzo]

📚 Installation

The installation instructions mention that, in order to install torch-cluster, we need to ask for torch-cluster==latest+${CUDA}, but torch-cluster has no such version (e.g. with cu92, cu100 or cu101). In torch-cluster's repo, the installation instructions are simply to run pip install torch-cluster, without an explicit CUDA version.

In short, I believe the command to install torch-cluster in the pytorch-geometric README may be wrong.

Environment

• OS: Manjaro Linux 19.0.2 Kyria
• Python version: 3.7.6
• PyTorch version: 1.4.0
• CUDA/cuDNN version: 10.1/7.6.5
• GCC version: 8.3.0
• How did you try to install PyTorch Geometric and its extensions (pip, source): pip
• Any other relevant information:

Checklist

• [X] I followed the installation guide.
• [ ] I cannot find my error message in the FAQ.
• [ ] I set up CUDA correctly and can compile CUDA code via nvcc.
• [ ] I have cloned the repository and tried a manual installation from source.
• [ ] I do have multiple CUDA versions on my machine.
• [ ] I checked if the official extension example runs on my machine.
• [ ] The offical extension example runs on my machine.

[rusty1s]

Yeah, I am aware of this. Try https://s3.eu-central-1.amazonaws.com/pytorch-geometric.com/whl/torch-1.4.0.html in the meantime.

[mwussow]

It seems to me as though this version of torch-cluster is not running as intended.
I successfully installed it with pip install torch-cluster==latest+cu100 -f https://s3.eu-central-1.amazonaws.com/pytorch-geometric.com/whl/torch-1.4.0.html and ran mnist_graclus.py from the examples folder without running into an error. However, it seems like it ran into an infinite loop, since the first epoch did not complete within 60min of training on a Nvidia Quadro P5000.

EDIT: I analyzed it further and indeed the graclus computation does not complete (but does not throw an error either)

[rusty1s]

Ok, let me look into this.

[rusty1s]

Fixed in v1.5.2.