evansoong517/messagepassing_for_gnns

Experiments with Message Passing GNNs in C++ and PyTorch.

Message Passing GNNs C++

My experiments with Graph Neural Nets at the scratch level using C++. Graph Convolutional Network (GCN) is one of the most popular GNN architectures and is extremely powerful. Of the popular graph representation learning methods which can be found at https://github.com/dsgiitr/graph_nets, this repo aims to implement GCNs in C++.

GCN C++ Forward Pass

This is a C++ implementation using Eigen for the forward pass of Graph Convolutional Neural Networks . The model doesn't involve any training loops and backpropagation. Pytorch is used to train the GCN model in Python and save the weights learnt after convergence. These saved weights are then imported and used in the C++ implementation of the model for the forward pass on the same dataset.

The GCN architecture and PyTorch implementation are explained in this blog are followed. The network is a 2 layer gcn model.

Training Visualization

The predicted outputs from the Python forward pass and C++ forward pass are saved in resources/saved/. As expected, they are almost identical and the similarity between both can be seen quantitatively as well visually with scatter plots in Compare_Predictions.ipynb.

Usage

Step 1

# Clone the repository
git clone https://github.com/AnirudhDagar/MessagePassing_for_GNNs.git

# Download the Karate Club Dataset
bash ./get_karate_club.sh

Step 2

# Train the model and save the weights.
python train.py

Step 3

# Use the CMakeLists.txt to build and run the project for C++ implementation.

# In the source directory
mkdir _build

# Change dir into _build
cd _build

# Build the project
cmake ..
make

OR

# Use a compiler directly to compile the executables.
g++ main.cpp -I eigen -std=c++17

Run at Once

# Run everything at once.
bash run.sh

Requirements

# C++
eigen

# Python
numpy==1.18.1
torch==1.0.0
ConfigArgParse==0.13.0
matplotlib==3.0.1
imageio==2.4.1
celluloid==0.2.0

Contributing

Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.