Interferometric Graph Transform

This repository contains the code corresponding to the ICML-2020 publication: https://arxiv.org/abs/2006.05722

If you have some troubles to reproduce some results, some questions regarding the code or even find a bug, please open directly an issue in this repository. If your questions concern some technical elements of the paper, please send directly an email to the author, firstname.name[AT]lip6.fr

If you wish to cite this paper, please use:

  title={Interferometric Graph Transform: a Deep Unsupervised Graph Representation},
  author={Oyallon, Edouard},
  booktitle={37th International Conference on Machine Learning (ICML 2020)},
  year={2020}
}

Reproducing the numerical results

In order to reproduce the experiments of the paper Interferometric Graph Transform: a Deep Unsupervised Graph Representation, make sure you have pytorch and python 3 installed, as well as MATLAB (only for the Haar Scattering experiments) and please follow the next instructions.

Community detection experiments

Please replace (or add) the files of https://github.com/alelab-upenn/graph-scattering-transforms with the ones in the folder "community-detection". Then run the code as described in this GitHub repository, by adding the corresponding dataset, as described.

Images experiments

Please run the following instructions on a GPU:

CIFAR10/MNIST, graph

$DATASET=cifar/mnist, $PATH_TO_MODEL refers to the path in which the first script stores the weight of the model.

python build_representation.py --K 40 --J 3 --lr_schedule '{0:1e0,500:1e-1,1000:1e-2,1500:1e-3}' --epochs 5  --dataset $DATASET
python classif.py --J 3 --feature interferometric --C 1e-3 --mode 1 --classifier 'svm'  --dataset $DATASET --path $PATH_TO_MODEL

CIFAR10, scattering

python classif.py --J 3 --feature scattering --C 1e-3 --mode 1 --classifier 'svm'  --dataset cifar

Graph experiments

NTU/SBU

Move the dataset (available https://github.com/mazariahmed06/graph_skeletons) in the same folder as the main file, then $DATASET=NTU_xview_values/NTU_xsub_values/SBU, $PATH_TO_MODEL refers to the path in which the first script stores the weight of the model, $ORDER is the required order and $K the respective number of filters, which should end by 0 (e.g., $ORDER=1 leads to $K=30,0 for instance).

Finally, $LR_SCHEDULE must be the learning rate schedule reported in the paper (as above). Add --no_averaging to remove the averaging from the experiments, yet this will substantially increase the computation time because the representation is larger.

python build_representation_graph.py --K $K --order $ORDER --lr_schedule $LR_SCHEDULE --epochs 5 --data_name $DATASET
python classif_graph.py --mode 'svm' --order 1 --data_name $DATASET --file $PATH_TO_MODEL  --K 10,5,0 --C 1e-1 

Haar scattering

First download HaarScat from https://www.di.ens.fr/data/software/ and then switch and run the scripts in the folder demo that we provided.