Propagate & Disill (P&D)

Implementation of the paper "Propagate & Distill: Towards Effective Graph Learners Using Propagation-Embracing MLPs" (LoG 2023) (link).

Tested in the following versions:

• Pytorch: 1.12.0
• Pytorch Geometric: 2.2.0
• Deep Graph Library: 0.9.1
• Numpy: 1.21.2

Main files:

• train_teacher.py: Train teacher GNNs
• train_student.py: Train student MLPs
  • Set --prop_iteration to a non-zero value for P&D
  • Include --fix_train flag for P&D_fix

Other notes:

• The majority of the codebase is from the implementation of GLNN (Zhang et al., "Graph-less Neural Networks: Teaching Old MLPs New Tricks via Distillation", ICLR 2022)!
• The data files need to be downloaded from the GLNN repo (link).