Spectral DGL

This repository was numerical basis of the masters thesis "On the Transferability of Spectral Graph Neural Networks". To find the final Master's thesis: https://infoscience.epfl.ch/record/278148?&ln=fr

Lib

Regroupment of functions used in all of the project

lib/
├── dataset.py
├── graphs.py
├── laplacian.py
├── model.py
├── ninja.py
└── utils.py

Resources

Shared resources between different part of the project

• models regroup several trained ChebNets that are all supposedly achieving 98% accuracy over their testing set

├── imgs
├── models

Chebnet on MNIST

Notebook implementation of the Chebyconv using the DGL framework. The net is a simple LeNet 5 with three convolution layer and two linear layers. The dataset is the good-old MNIST

├──ChebGCNs.ipynb
├── README.md
├── Eigonvalues_Viz.ipynb
├── Gui.py
├── Training_Chebnet_on_MNIST.ipynb
├── filtervisualisation.ipynb
├── laplacianprediction.ipynb
├── transform_training.ipynb
└── visualisations.ipynb

Stochastic learning

Notebook for stochastic learning. For now only two notebooks to train and test the model are implemented.

├── Training
├── Testing