layers.py

import math
import torch
from torch.nn.parameter import Parameter
from torch.nn.modules.module import Module


class SparseMM(torch.autograd.Function):
    """
    Sparse x dense matrix multiplication with autograd support.

    Implementation by Soumith Chintala:
    https://discuss.pytorch.org/t/does-pytorch-support-autograd-on-sparse-matrix/6156/7
    """

    @staticmethod
    def forward(ctx, matrix1, matrix2):
        ctx.save_for_backward(matrix1, matrix2)
        return torch.mm(matrix1, matrix2)

    @staticmethod
    def backward(ctx, grad_output):
        matrix1, matrix2 = ctx.saved_tensors
        grad_matrix1 = grad_matrix2 = None

        if ctx.needs_input_grad[0]:
            grad_matrix1 = torch.mm(grad_output, matrix2.t())

        if ctx.needs_input_grad[1]:
            grad_matrix2 = torch.mm(matrix1.t(), grad_output)

        return grad_matrix1, grad_matrix2


class GraphConvolution(Module):
    """
    Simple GCN layer, similar to https://arxiv.org/abs/1609.02907
    """

    def __init__(self, in_features, out_features, bias=False):
        super(GraphConvolution, self).__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.weight = Parameter(torch.Tensor(in_features, out_features))
        self.sparse_mm = SparseMM.apply
        if bias:
            self.bias = Parameter(torch.Tensor(out_features))
        else:
            self.register_parameter('bias', None)
        self.reset_parameters()

    def reset_parameters(self):
        init_range = math.sqrt(6.0 / (self.in_features + self.out_features))
        self.weight.data.uniform_(-init_range, init_range)
        if self.bias is not None:
            self.bias.data.uniform_(-init_range, init_range)

    def forward(self, x, adj):
        support = self.sparse_mm(x, self.weight)
        # support = torch.mm(x, self.weight)
        output = self.sparse_mm(adj, support)
        if self.bias is not None:
            return output + self.bias
        else:
            return output


    def __repr__(self):
        return self.__class__.__name__ + ' (' \
               + str(self.in_features) + ' -> ' \
               + str(self.out_features) + ')'