ChebyGCN 0.0.3

Implements graph convolution keras layers based on Michaël Defferrard, Xavier Bresson, Pierre Vandergheynst, Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering, Neural Information Processing Systems (NIPS), 2016.

ChebyGCN

pip install ChebyGCN

Notice, for training and testing data, permutations of the data must be done in a certain way to align with pooling of the graph lapacian. Further, every level of graph corsening is a pool of size two, thus if you want to pool by 2 and then 4, you need log_2(2 * 4)= 3 levels. You will also need to index your Lapancians as seen below.

from ChebyGCN import layers, coarsening
A = scipy.sparse.csr.csr_matrix(A) #load adjanecy matrix 
graphs, perm = coarsening.coarsen(A, levels=3, self_connections=True) #produce graph coarsenings 
X_train = coarsening.perm_data(X_train, perm)
X_test = coarsening.perm_data(X_test, perm)
L = [coarsening.laplacian(A, normalized=True) for A in graphs]

x_input = Input(shape=(X_train.shape[1],))
x = Reshape((X_train.shape[1],1))(x_input)
x = layers.GraphConvolution( 8, 2, 20, L[0])(x)
x = layers.GraphConvolution( 8, 4, 10, L[2])(x)
x = Flatten()(x)
x = Dense(66, activation='softmax')(x)

This code is 96% based on https://github.com/mdeff/cnn_graph Michaël Defferrard, Xavier Bresson, Pierre Vandergheynst, Convolutional Neural Networks on Graphs with Fast Localized Spectral Filtering, Neural Information Processing Systems (NIPS), 2016.