layer_diffusion_conv: DiffusionConv

In rdinnager/rspektral: What the Package Does (One Line, Title Case)

Description

\loadmathjax

Applies Graph Diffusion Convolution as descibed by Li et al. (2016)

Mode: single, disjoint, mixed, batch.

This layer expects a dense adjacency matrix.

Given a number of diffusion steps \mjeqnK and a row normalized adjacency matrix \mjeqn\hat A , this layer calculates the q'th channel as: \mjdeqn \mathbfH _ ~:,~q = \sigma\left( \sum_f=1^F \left( \sum_k=0^K-1\theta_k \hat A^k \right) X_~:,~f \right)

Input

• Node features of shape ([batch], N, F);

• Normalized adjacency or attention coef. matrix \mjeqn\hat A of shape ([batch], N, N); Use DiffusionConvolution.preprocess to normalize.

Output

• Node features with the same shape as the input, but with the last dimension changed to channels.

Usage

layer_diffusion_conv(
  object,
  channels,
  num_diffusion_steps = 6,
  kernel_initializer = "glorot_uniform",
  kernel_regularizer = NULL,
  kernel_constraint = NULL,
  activation = "tanh",
  ...
)

Arguments

channels	number of output channels
num_diffusion_steps	How many diffusion steps to consider. \(K\) in paper.
kernel_initializer	initializer for the weights
kernel_regularizer	regularization applied to the weights
kernel_constraint	constraint applied to the weights
activation	activation function \(\sigma\) (\(\tanh\) by default)

rdinnager/rspektral documentation built on June 12, 2021, 1:26 a.m.