Universal Function Approximation on Graphs

This repository is the official PyTorch implementation of the experiments in the following paper:

Rickard Brüel-Gabrielsson. Universal Function Approximation on Graphs (NeurIPS 2020)

arXiv

The code is based on the code of How Powerful are Graph Neural Networks?

Compiling C++ Extensions

If you haven't already, clone the repository

You are now ready to compile extensions. PyTorch tutorial on extensions here

Important: in environment, it seems like using the pytorch conda channel is important

source activate environment
conda install pytorch torchvision -c pytorch

Compilation uses python's setuptools module.

To complile (from home directory):

source activate environment
python setup.py install --record files.txt

You should now have the package available in your environment. You can run the above command any time you modify the source code, and the package on your path should update.

MacOS Information: If PyTorch was compiled using clang++, you may run into issues if pip defaults to g++. You can make pip use clang++ by setting the CXX environment variable. The CPPFLAGS environment variable also needs to be set to look at libc++ to avoid compatibility issues with the PyTorch headers. The MACOSX_DEPLOYMENT_TARGET environment variable may also need to be set (set the target to be whatever your OS version is).

export CXX=/usr/bin/clang++
export CPPFLAGS="-stdlib=libc++"
export MACOSX_DEPLOYMENT_TARGET=$(sw_vers -productVersion)
python setup.py install --record files.txt