search

mdeff / cnn_graph

Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering
https://arxiv.org/abs/1606.09375
MIT License
1.34k stars 390 forks source link

Can graph_conv_cheby method be applied to an arbitrary graph ? #26

Closed pinkfloyd06 closed 6 years ago

pinkfloyd06 commented 6 years ago

Hello,

Let me first thank you for your work.

l would like to ask you whether your graph_conv_cheby() can be applied to an arbitrary graph such as an irregular graph or just on a grid graph ?

Thank you

mdeff commented 6 years ago

Yes absolutely. You only need an adjacency matrix which describes your graph.

pinkfloyd06 commented 6 years ago

@mdeff Thank you for your answer. Hoever l get confused after looking at your experiment on MNIST https://github.com/mdeff/cnn_graph/blob/master/nips2016/mnist.ipynb

A = grid_graph(28, corners=False) and the number of edges per vertex is at least 8.

What if l don't have a grid. l deal with irregular graphs where each graph has N nodes such that :

each node_i has k_i vertices.

-So the number of vertices varies from node to another either in the same graph or on different graphs

mdeff commented 6 years ago

If the number of nodes per graph is variable, then your main issue is that you'll have to "summarize" the information in a fixed-length vector at some point. See the discussion in #5. As far as the graph convolution is concerned, you can have a different adjacency matrix for each data point.