layers.py

import torch
import torch.nn.functional as F
from torch.nn.modules.module import Module
from torch.nn.parameter import Parameter
import torch.nn as nn



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

    def __init__(self, in_features, out_features, dropout=0., act=torch.relu):
        super(GraphConvolution, self).__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.dropout = dropout
        self.act = act
        self.weight = Parameter(torch.FloatTensor(in_features, out_features))
        self.reset_parameters()

    def reset_parameters(self):
        torch.nn.init.xavier_uniform_(self.weight)

    def forward(self, input, adj):
        input = F.dropout(input, self.dropout, self.training)
        support = torch.mm(input, self.weight)
        output = torch.spmm(adj, support)
        output = self.act(output)
        return output

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

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

    def __init__(self, in_features, out_features, dropout=0., act=torch.relu):
        super(GraphConvolutionSparse, self).__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.dropout = dropout
        self.act = act
        self.issparse = True
        self.weight = Parameter(torch.FloatTensor(in_features, out_features))
        self.reset_parameters()

    def reset_parameters(self):
        torch.nn.init.xavier_uniform_(self.weight)

    def forward(self, input, adj):
        input = F.dropout(input, self.dropout, self.training)
        support = torch.spmm(input, self.weight)
        output = torch.spmm(adj, support)
        output = self.act(output)
        return output

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

class Linear(Module):
    """
    to embedding feature
    """

    def __init__(self,in_features,out_features,dropout=0.,act=torch.relu,bias=True,sparse_inputs=False,**kwargs):
        super(Linear,self).__init__(**kwargs)
        self.in_features = in_features
        self.out_features = out_features
        self.dropout = dropout
        self.act = act
        self.bias = bias
        self.sparse_inputs = sparse_inputs

        self.weight= Parameter(torch.FloatTensor(in_features,out_features))

        if self.bias:
            self.weight_bias  = Parameter(torch.FloatTensor(1,out_features))

        self.reset_parameters()

    def reset_parameters(self):
        torch.nn.init.xavier_uniform_(self.weight)
        if self.bias:
            torch.nn.init.xavier_uniform_(self.weight_bias)

    def forward(self,input,adj=None):
        if self.sparse_inputs:
            output = torch.spmm(input,self.weight)
        else:
            output = torch.mm(input,self.weight)

        if self.bias:
            output += self.weight_bias #Find the bug, self.bias should be self.weight_bias
            # output += self.bias #Find the bug, self.bias should be self.weight_bias

        return self.act(output)


class InnerProductDecoder(torch.nn.Module):
    """Decoder for using inner product for prediction."""

    def __init__(self, dropout, act=torch.sigmoid):
        super(InnerProductDecoder, self).__init__()
        self.dropout = dropout
        self.act = act

    def forward(self, z):
        z = F.dropout(z, self.dropout, training=self.training)
        adj = self.act(torch.mm(z, z.t()))
        return adj

class InnerDecoder(torch.nn.Module):
    """Decoder for using inner product for prediction."""

    def __init__(self,dropout=0., act=torch.sigmoid,**kwargs):
        super(InnerDecoder, self).__init__(**kwargs)
        self.dropout = dropout
        self.act = act

    def forward(self,inputs):
        z_u, z_a = inputs
        z_u = F.dropout(z_u, self.dropout, training=self.training)
        z_a = F.dropout(z_a, self.dropout,training = self.training)
        adj = self.act(torch.mm(z_u, z_u.t())) # predicted adj matrix
        features = self.act(torch.mm(z_u,z_a.t())) #predicted feature matrix
        return adj,features

class GraphAttentionLayer(nn.Module):
    """
    Simple GAT layer, similar to https://arxiv.org/abs/1710.10903
    """
    def __init__(self, in_features, out_features, dropout, alpha, concat=True):
        super(GraphAttentionLayer, self).__init__()
        self.dropout = dropout
        self.in_features = in_features
        self.out_features = out_features
        self.alpha = alpha
        self.concat = concat

        self.W = nn.Parameter(torch.empty(size=(in_features, out_features)))
        nn.init.xavier_uniform_(self.W.data, gain=1.414)
        self.a = nn.Parameter(torch.empty(size=(2*out_features, 1)))
        nn.init.xavier_uniform_(self.a.data, gain=1.414)

        self.leakyrelu = nn.LeakyReLU(self.alpha)

    def forward(self, h, adj):
        Wh = torch.mm(h, self.W) # h.shape: (N, in_features), Wh.shape: (N, out_features)
        a_input = self._prepare_attentional_mechanism_input(Wh)
        e = self.leakyrelu(torch.matmul(a_input, self.a).squeeze(2))

        zero_vec = -9e15*torch.ones_like(e)
        attention = torch.where(adj > 0, e, zero_vec)
        attention = F.softmax(attention, dim=1)
        attention = F.dropout(attention, self.dropout, training=self.training)
        h_prime = torch.matmul(attention, Wh)

        if self.concat:
            return F.elu(h_prime)
        else:
            return h_prime

    def _prepare_attentional_mechanism_input(self, Wh):
        N = Wh.size()[0] # number of nodes

        # Below, two matrices are created that contain embeddings in their rows in different orders.
        # (e stands for embedding)
        # These are the rows of the first matrix (Wh_repeated_in_chunks):
        # e1, e1, ..., e1,            e2, e2, ..., e2,            ..., eN, eN, ..., eN
        # '-------------' -> N times  '-------------' -> N times       '-------------' -> N times
        #
        # These are the rows of the second matrix (Wh_repeated_alternating):
        # e1, e2, ..., eN, e1, e2, ..., eN, ..., e1, e2, ..., eN
        # '----------------------------------------------------' -> N times
        #

        Wh_repeated_in_chunks = Wh.repeat_interleave(N, dim=0)
        Wh_repeated_alternating = Wh.repeat(N, 1)
        # Wh_repeated_in_chunks.shape == Wh_repeated_alternating.shape == (N * N, out_features)

        # The all_combination_matrix, created below, will look like this (|| denotes concatenation):
        # e1 || e1
        # e1 || e2
        # e1 || e3
        # ...
        # e1 || eN
        # e2 || e1
        # e2 || e2
        # e2 || e3
        # ...
        # e2 || eN
        # ...
        # eN || e1
        # eN || e2
        # eN || e3
        # ...
        # eN || eN

        all_combinations_matrix = torch.cat([Wh_repeated_in_chunks, Wh_repeated_alternating], dim=1)
        # all_combinations_matrix.shape == (N * N, 2 * out_features)

        return all_combinations_matrix.view(N, N, 2 * self.out_features)

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


class SpecialSpmmFunction(torch.autograd.Function):
    """Special function for only sparse region backpropataion layer."""
    @staticmethod
    def forward(ctx, indices, values, shape, b):
        assert indices.requires_grad == False
        a = torch.sparse_coo_tensor(indices, values, shape)
        ctx.save_for_backward(a, b)
        ctx.N = shape[0]
        return torch.matmul(a, b)

    @staticmethod
    def backward(ctx, grad_output):
        a, b = ctx.saved_tensors
        grad_values = grad_b = None
        if ctx.needs_input_grad[1]:
            grad_a_dense = grad_output.matmul(b.t())
            edge_idx = a._indices()[0, :] * ctx.N + a._indices()[1, :]
            grad_values = grad_a_dense.view(-1)[edge_idx]
        if ctx.needs_input_grad[3]:
            grad_b = a.t().matmul(grad_output)
        return None, grad_values, None, grad_b


class SpecialSpmm(nn.Module):
    def forward(self, indices, values, shape, b):
        return SpecialSpmmFunction.apply(indices, values, shape, b)


class SpGraphAttentionLayer(nn.Module):
    """
    Sparse version GAT layer, similar to https://arxiv.org/abs/1710.10903
    """

    def __init__(self, in_features, out_features, dropout, alpha, concat=False):
        super(SpGraphAttentionLayer, self).__init__()
        self.in_features = in_features
        self.out_features = out_features
        self.alpha = alpha
        self.concat = concat

        self.W = nn.Parameter(torch.zeros(size=(in_features, out_features)))
        nn.init.xavier_normal_(self.W.data, gain=1.414)

        self.a = nn.Parameter(torch.zeros(size=(1, 2*out_features)))
        nn.init.xavier_normal_(self.a.data, gain=1.414)

        self.dropout = nn.Dropout(dropout)
        self.leakyrelu = nn.LeakyReLU(self.alpha)
        self.special_spmm = SpecialSpmm()

    def forward(self, input, adj):
        dv = 'cuda' if input.is_cuda else 'cpu'

        N = input.size()[0]
        edge = adj.nonzero().t()

        h = torch.mm(input, self.W)
        # h: N x out
        assert not torch.isnan(h).any()

        # Self-attention on the nodes - Shared attention mechanism
        edge_h = torch.cat((h[edge[0, :], :], h[edge[1, :], :]), dim=1).t()
        # edge: 2*D x E

        edge_e = torch.exp(-self.leakyrelu(self.a.mm(edge_h).squeeze()))
        assert not torch.isnan(edge_e).any()
        # edge_e: E

        e_rowsum = self.special_spmm(edge, edge_e, torch.Size([N, N]), torch.ones(size=(N,1), device=dv))
        # e_rowsum: N x 1

        edge_e = self.dropout(edge_e)
        # edge_e: E

        h_prime = self.special_spmm(edge, edge_e, torch.Size([N, N]), h)
        assert not torch.isnan(h_prime).any()
        # h_prime: N x out

        h_prime = h_prime.div(e_rowsum)
        # h_prime: N x out
        assert not torch.isnan(h_prime).any()

        if self.concat:
            # if this layer is not last layer,
            return F.elu(h_prime)
        else:
            # if this layer is last layer,
            return h_prime

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

class SpGAT(nn.Module):
    def __init__(self, nfeat, nhid, nclass, dropout, alpha = 0.2, nheads = 3):
        """Sparse version of GAT."""
        super(SpGAT, self).__init__()
        self.dropout = dropout

        self.attentions = [SpGraphAttentionLayer(nfeat,
                                                 nhid,
                                                 dropout=dropout,
                                                 alpha=alpha,
                                                 concat=True) for _ in range(nheads)]
        for i, attention in enumerate(self.attentions):
            self.add_module('attention_{}'.format(i), attention)

        self.out_att = SpGraphAttentionLayer(nhid * nheads,
                                             nclass,
                                             dropout=dropout,
                                             alpha=alpha,
                                             concat=False)

    def forward(self, x, adj):
        x = F.dropout(x, self.dropout, training=self.training)
        x = torch.cat([att(x, adj) for att in self.attentions], dim=1)
        # x = F.dropout(x, self.dropout, training=self.training)
        x = self.out_att(x, adj)
        # x = F.elu(self.out_att(x, adj))
        # return F.log_softmax(x, dim=1)
        return x

class GAT(nn.Module):
    def __init__(self, nfeat, nhid, nclass, dropout, alpha=0.2, nheads = 3):
        """Dense version of GAT."""
        super(GAT, self).__init__()
        self.dropout = dropout

        self.attentions = [GraphAttentionLayer(nfeat, nhid, dropout=dropout, alpha=alpha, concat=True) for _ in range(nheads)]
        for i, attention in enumerate(self.attentions):
            self.add_module('attention_{}'.format(i), attention)

        self.out_att = GraphAttentionLayer(nhid * nheads, nclass, dropout=dropout, alpha=alpha, concat=False)

    def forward(self, x, adj):
        x = F.dropout(x, self.dropout, training=self.training)
        x = torch.cat([att(x, adj) for att in self.attentions], dim=1)
        x = F.dropout(x, self.dropout, training=self.training)
        # x = F.elu(self.out_att(x, adj))
        return x
        # return F.log_softmax(x, dim=1)