search

rusty1s / pyg_autoscale

Implementation of "GNNAutoScale: Scalable and Expressive Graph Neural Networks via Historical Embeddings" in PyTorch
http://arxiv.org/abs/2106.05609
MIT License
159 stars 27 forks source link

Results on OGBN-Papers100M #31

Open RX28666 opened 7 months ago

RX28666 commented 7 months ago

Hello,

I am wondering if there are any results of GAS on OGBN-Papers100M. (Or results on some datasets larger than ogbn-products)

BTW, since the dataset is so big, preprocessing steps such as partitioning with METIS are unrealistic to implement as usual. I am also wondering if there are any code scripts I can refer to that can help solve this issue.

Thanks.

rusty1s commented 7 months ago

We didn't test on ogbn-papers100M at this point in time. Given enough RAM, this should definitely work though. I partially agree that METIS becomes infeasible on larger datasets (although all large-scale graph learning variants rely on it to scale), but you can also fallback to random partitioning.

RX28666 commented 7 months ago

Thanks for your advice, I will try to run GAS on ogbn-papers100m.

RX28666 commented 7 months ago

Hello Matthias,

I also faced a issue when I try to install the package, recently I updated my CUDA to 12.1:

CUDA Version: 12.1
PyTorch Version: 2.2.1+cu121

I tried to install the package in both ways provided, they all returned:

RuntimeError:
      The detected CUDA version (11.6) mismatches the version that was used to compile
      PyTorch (12.1). Please make sure to use the same CUDA versions.

Is this because the current package doesn't support cuda 12.1? Thanks.

rusty1s commented 7 months ago

You need to re-install this package if you also update your CUDA version.

RX28666 commented 7 months ago

Hello,

I deleted the original one by:

pip uninstall pyg_autoscale
WARNING: Skipping pyg_autoscale as it is not installed.

then re-install using pip install git+https://github.com/rusty1s/pyg_autoscale.git and also tried python setup.py install

they both returned the same bug.

The detected CUDA version (11.6) mismatches the version that was used to compile
      PyTorch (12.1). Please make sure to use the same CUDA versions.

Is there anything I was missing? Any help would be appreciated.

rusty1s commented 7 months ago

Ok, got it. This is IMO expected. You are using your local CUDA version (11.6) to compile this package, while you have PyTorch installed with CUDA version 12.1. What you can do