Skip to content
PyTorch implementation of spectral graph ConvNets, NIPS’16
Branch: master
Clone or download
Latest commit 1919c6f Oct 15, 2017
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
lib first commit Oct 1, 2017
pic clean cache Oct 1, 2017
results_html minor_modif Oct 15, 2017
.gitignore make dev folder Oct 1, 2017
01_standard_convnet_lenet5_mnist_pytorch.ipynb first commit Oct 1, 2017
02_graph_convnet_lenet5_mnist_pytorch.ipynb minor_modif Oct 15, 2017
LICENSE Initial commit Oct 1, 2017
README.md fix_date Oct 15, 2017
check_install.py image size 6 Oct 1, 2017
requirements.txt first commit Oct 1, 2017

README.md

Graph ConvNets in PyTorch

October 15, 2017

Xavier Bresson

http://www.ntu.edu.sg/home/xbresson
https://github.com/xbresson
https://twitter.com/xbresson

Description

Prototype implementation in PyTorch of the NIPS'16 paper:
Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering
M Defferrard, X Bresson, P Vandergheynst
Advances in Neural Information Processing Systems, 3844-3852, 2016
ArXiv preprint: arXiv:1606.09375

Code objective

The code provides a simple example of graph ConvNets for the MNIST classification task.
The graph is a 8-nearest neighbor graph of a 2D grid.
The signals on graph are the MNIST images vectorized as $28^2 \times 1$ vectors.

Installation

git clone https://github.com/xbresson/graph_convnets_pytorch.git
cd graph_convnets_pytorch
pip install -r requirements.txt # installation for python 3.6.2
python check_install.py
jupyter notebook # run the 2 notebooks

Results

GPU Quadro M4000

  • Standard ConvNets: 01_standard_convnet_lenet5_mnist_pytorch.ipynb, accuracy= 99.31, speed= 6.9 sec/epoch.
  • Graph ConvNets: 02_graph_convnet_lenet5_mnist_pytorch.ipynb, accuracy= 99.19, speed= 100.8 sec/epoch

Note

PyTorch has not yet implemented function torch.mm(sparse, dense) for variables: https://github.com/pytorch/pytorch/issues/2389. It will be certainly implemented but in the meantime, I defined a new autograd function for sparse variables, called "my_sparse_mm", by subclassing torch.autograd.function and implementing the forward and backward passes.

class my_sparse_mm(torch.autograd.Function):
    """
    Implementation of a new autograd function for sparse variables, 
    called "my_sparse_mm", by subclassing torch.autograd.Function 
    and implementing the forward and backward passes.
    """
    
    def forward(self, W, x):  # W is SPARSE
        self.save_for_backward(W, x)
        y = torch.mm(W, x)
        return y
    
    def backward(self, grad_output):
        W, x = self.saved_tensors 
        grad_input = grad_output.clone()
        grad_input_dL_dW = torch.mm(grad_input, x.t()) 
        grad_input_dL_dx = torch.mm(W.t(), grad_input )
        return grad_input_dL_dW, grad_input_dL_dx

When to use this algorithm?

Any problem that can be cast as analyzing a set of signals on a fixed graph, and you want to use ConvNets for this analysis.




You can’t perform that action at this time.