graphConv.py

Implementation of two types of graph convolution layers, and application of them on TCGA RNA-seq data and a graph defined over genes by BioGrid PPI.

Code

Graph convolution without self connection:

def convolutionGraph(inputs,
                     num_outputs,
                     glap,
                     activation_fn=nn.relu,
                     weights_initializler=initializers.xavier_initializer(),
                     biases_initializer=init_ops.zeros_initializer(),
                     reuse=None,
                     scope=None)

convlolutionGraph() implements a graph convolution layer defined by Kipf et al.

inputs is a 2d tensor that goes into the layer.
num_outputs specifies the number of channels wanted on the output tensor.
glap is an instance of tf.SparseTensor that defines a graph laplacian matrix DAD.

Graph convolution with self connection:

def convolutionGraph_sc(inputs,
                     num_outputs,
                     glap,
                     activation_fn=nn.relu,
                     weights_initializler=initializers.xavier_initializer(),
                     biases_initializer=init_ops.zeros_initializer(),
                     reuse=None,
                     scope=None):

convlolutionGraph_sc() implements a graph convolution layer defined by Kipf et al, except that self-connection of nodes are allowed.

inputs is a 2d tensor that goes into the layer.
num_outputs specifies the number of channels wanted on the output tensor.
glap is an instance of tf.SparseTensor that defines a graph laplacian matrix DAD.

inits.py:

This file contains 4 common initialization methods for network weights, i.e., uniform(), glorot(), zeros(), and ones(). We currently do not use it.

utils.py:

This file contains

Two versions of DataFeeder instances.
Helper functions for data processing. See ipynb/DataProcessing.ipynb for how to use them.
A class definition for intx.

Test Data Files

We have ppi matrix extracted from BioGrid in March 2018. This defines a graph. We also have TCGA RNA-seq data for approx 9000 samples. This is the data that are getting convolved over the graph. You can download datafiles from here.

Models

See ipynb/runningGraphConv.py for implemention and training of the models. It is should be straight forward.

For any question, e-mail me at hiranumn at cs dot washington dot edu.

hiranumn/GraphConv