README.md

SpeqNets

Code for the ICML 2022 paper "SpeqNets: Sparsity-aware Permutation-equivariant Graph Networks".

Requirements

• Python 3.9
• eigen3
• numpy
• pandas
• scipy
• sklearn
• torch 1.10.x
• torch-geometric 2.0.x
• pybind11
• libsvm

All results in the paper and the appendix can be reproduced by the following the steps below.

Reproducing the kernel experiments (Table 1, 5)

You first need to build the Python package:

• cd k_s_wl_cpp/implementation/tud_benchmark/kernel_baselines
• You might need to adjust the path to pybind in kernel_baselines.cpp, then run
  • MaxOS: c++ -O3 -shared -std=c++11 -undefined dynamic_lookup python3 -m pybind11 --includes kernel_baselines.cpp src/*cpp -o ../kernel_baselinespython3-config --extension-suffix
  • Linux: c++ -O3 -shared -std=c++11 -fPIC python3 -m pybind11 --includes kernel_baselines.cpp src/*cpp -o ../kernel_baselinespython3-config --extension-suffix
• cd ..
• Run python main_kernel.py and python main_gnn.py

Reproducing the neural higher-order results (Table 2(a), 6)

You first need to build the Python package:

• cd neural_graph/preprocessing

• You might need to adjust the path to pybind in preprocessing.cpp, then run

  • MaxOS: c++ -O3 -shared -std=c++11 -undefined dynamic_lookup python3 -m pybind11 --includes preprocessing.cpp src/*cpp -o ../preprocessingpython3-config --extension-suffix
  • Linux: c++ -O3 -shared -std=c++11 -fPIC python3 -m pybind11 --includes preprocessing.cpp src/*cpp -o ../preprocessingpython3-config --extension-suffix

• Run the Python scripts

  • For example: python main_2_2_alchemy_10K.py, python main_1_1_QM9.py, ...

Reproducing the node classification experiments (Tables 2(b))

• cd neural_node
• Run python gnn_1.py, python simple_node_2_1.py, ...