import torch
import torch.nn.functional as F
from torch.nn import Parameter
from torch_geometric.nn.conv import MessagePassing
from torch_geometric.utils import add_remaining_self_loops
from ..inits import uniform
[docs]class SAGEConv(MessagePassing):
r"""The GraphSAGE operator from the `"Inductive Representation Learning on
Large Graphs" <https://arxiv.org/abs/1706.02216>`_ paper
.. math::
\mathbf{\hat{x}}_i &= \mathbf{\Theta} \cdot
\mathrm{mean}_{j \in \mathcal{N(i) \cup \{ i \}}}(\mathbf{x}_j)
\mathbf{x}^{\prime}_i &= \frac{\mathbf{\hat{x}}_i}
{\| \mathbf{\hat{x}}_i \|_2}.
Args:
in_channels (int): Size of each input sample.
out_channels (int): Size of each output sample.
normalize (bool, optional): If set to :obj:`True`, output features
will be :math:`\ell_2`-normalized. (default: :obj:`False`)
concat (bool, optional): If set to :obj:`True`, will concatenate
current node features with aggregated ones. (default: :obj:`False`)
bias (bool, optional): If set to :obj:`False`, the layer will not learn
an additive bias. (default: :obj:`True`)
**kwargs (optional): Additional arguments of
:class:`torch_geometric.nn.conv.MessagePassing`.
"""
def __init__(self, in_channels, out_channels, normalize=False,
concat=False, bias=True, **kwargs):
super(SAGEConv, self).__init__(aggr='mean', **kwargs)
self.in_channels = in_channels
self.out_channels = out_channels
self.normalize = normalize
self.concat = concat
in_channels = 2 * in_channels if concat else in_channels
self.weight = Parameter(torch.Tensor(in_channels, out_channels))
if bias:
self.bias = Parameter(torch.Tensor(out_channels))
else:
self.register_parameter('bias', None)
self.reset_parameters()
[docs] def reset_parameters(self):
uniform(self.weight.size(0), self.weight)
uniform(self.weight.size(0), self.bias)
[docs] def forward(self, x, edge_index, edge_weight=None, size=None,
res_n_id=None):
"""
Args:
res_n_id (Tensor, optional): Residual node indices coming from
:obj:`DataFlow` generated by :obj:`NeighborSampler` are used to
select central node features in :obj:`x`.
Required if operating in a bipartite graph and :obj:`concat` is
:obj:`True`. (default: :obj:`None`)
"""
if not self.concat and torch.is_tensor(x):
edge_index, edge_weight = add_remaining_self_loops(
edge_index, edge_weight, 1, x.size(self.node_dim))
return self.propagate(edge_index, size=size, x=x,
edge_weight=edge_weight, res_n_id=res_n_id)
def message(self, x_j, edge_weight):
return x_j if edge_weight is None else edge_weight.view(-1, 1) * x_j
def update(self, aggr_out, x, res_n_id):
if self.concat and torch.is_tensor(x):
aggr_out = torch.cat([x, aggr_out], dim=-1)
elif self.concat and (isinstance(x, tuple) or isinstance(x, list)):
assert res_n_id is not None
aggr_out = torch.cat([x[0][res_n_id], aggr_out], dim=-1)
aggr_out = torch.matmul(aggr_out, self.weight)
if self.bias is not None:
aggr_out = aggr_out + self.bias
if self.normalize:
aggr_out = F.normalize(aggr_out, p=2, dim=-1)
return aggr_out
def __repr__(self):
return '{}({}, {})'.format(self.__class__.__name__, self.in_channels,
self.out_channels)