Source code for torch_geometric.nn.dense.dense_graph_conv

import torch
from torch.nn import Linear


[docs]class DenseGraphConv(torch.nn.Module): r"""See :class:`torch_geometric.nn.conv.GraphConv`. """ def __init__(self, in_channels, out_channels, aggr='add', bias=True): assert aggr in ['add', 'mean', 'max'] super(DenseGraphConv, self).__init__() self.in_channels = in_channels self.out_channels = out_channels self.aggr = aggr self.lin_rel = Linear(in_channels, out_channels, bias=False) self.lin_root = Linear(in_channels, out_channels, bias=bias) self.reset_parameters()
[docs] def reset_parameters(self): self.lin_rel.reset_parameters() self.lin_root.reset_parameters()
[docs] def forward(self, x, adj, mask=None): r""" Args: x (Tensor): Node feature tensor :math:`\mathbf{X} \in \mathbb{R}^{B \times N \times F}`, with batch-size :math:`B`, (maximum) number of nodes :math:`N` for each graph, and feature dimension :math:`F`. adj (Tensor): Adjacency tensor :math:`\mathbf{A} \in \mathbb{R}^{B \times N \times N}`. The adjacency tensor is broadcastable in the batch dimension, resulting in a shared adjacency matrix for the complete batch. mask (BoolTensor, optional): Mask matrix :math:`\mathbf{M} \in {\{ 0, 1 \}}^{B \times N}` indicating the valid nodes for each graph. (default: :obj:`None`) """ x = x.unsqueeze(0) if x.dim() == 2 else x adj = adj.unsqueeze(0) if adj.dim() == 2 else adj B, N, C = x.size() if self.aggr in ['add', 'mean']: out = torch.matmul(adj, x) if self.aggr == 'mean': out = out / adj.sum(dim=-1, keepdim=True).clamp_(min=1) elif self.aggr == 'max': out = x.unsqueeze(-2).repeat(1, 1, N, 1) adj = adj.unsqueeze(-1).expand(B, N, N, C) out[adj == 0] = float('-inf') out = out.max(dim=-3)[0] out[out == float('-inf')] = 0. out = self.lin_rel(out) out += self.lin_root(x) if mask is not None: out = out * mask.view(-1, N, 1).to(x.dtype) return out
def __repr__(self): return '{}({}, {})'.format(self.__class__.__name__, self.in_channels, self.out_channels)