CNN_Graph.pytorch

This repository is an unofficial implement of the paper "Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering" in NIPS 2016 with PyTorch. It supports to train and evaluate the network on the 20news and mnist.

Requirements

• Python 3.6
• PyTorch 1.1

Train and Evaluation

With following scripts, you can train and evaluate the graph-based CNN with corresponding network architectures on mnist dataset.

python train.py --data mnist --filter fourier --gc_layers 1
python train.py --data mnist --filter fourier --gc_layers 2
python train.py --data mnist --filter chebyshev --gc_layers 1
python train.py --data mnist --filter chebyshev --gc_layers 2

where gc_layers=1 corresponds to the network architecture GC32, and gc_layers=2 to GC32-P4-GC64-P4-FC512.

To train and evaluate on 20news dataset, you need to run following script in order to preprocess the documents and generate required intermediate data.

python scripts/20news_preprocess.py

Then you can train the network with different negtwork architectures on 20news dataset.

python train.py --data 20news --filter fourier --gc_layers 1
python train.py --data 20news --filter chebyshev --gc_layers 1

Note that the codes under the folder lib are completely borrowed from original codebase CNN_Graph, where the redundant functions have been removed. This part is responsible for graph building and coarsening.

Performance

MNIST

filters	gc_layer=1	gc_layer=2
Fourier	0.9747	0.9788
Chebyshev	0.9816	0.9818

20NEWS

filters	gc_layer=1	gc_layer=2
Fourier	0.5504	-
Chebyshev	0.5554	-

Acknowledgements

Thanks the official CNN_Graph implemented with TensorFlow and awesome PyTorch team.