wsgat.py

from typing import Optional, Tuple, Union

import torch
import torch.nn.functional as F
from torch import Tensor
from torch.nn import Parameter, ParameterList
from torch_geometric.nn.conv import MessagePassing
from torch_geometric.nn.dense.linear import Linear
from torch_geometric.nn.inits import glorot, zeros
from torch_geometric.typing import (
    Adj,
    OptPairTensor,
    OptTensor,
    Size,
    SparseTensor,
    torch_sparse,
)
from torch_geometric.typing import NoneType  # noqa
from torch_geometric.utils import (
    add_self_loops,
    is_torch_sparse_tensor,
    remove_self_loops,
    softmax,
)
from torch_geometric.utils.sparse import set_sparse_value


class wsGATConv(MessagePassing):
    r"""The wsGAT graph attentional operator from the `"wsGAT: Weighted and Signed Graph Attention Networks for Link Prediction"
    <https://doi.org/10.1007/978-3-030-93409-5_31>`_ paper.

    ```
    Grassia, M., Mangioni, G. (2022). wsGAT: Weighted and Signed Graph Attention Networks for Link Prediction.
    In: Benito, R.M., Cherifi, C., Cherifi, H., Moro, E., Rocha, L.M., Sales-Pardo, M. (eds) Complex Networks & Their Applications X.
    COMPLEX NETWORKS 2021. Studies in Computational Intelligence, vol 1072. Springer, Cham. https://doi.org/10.1007/978-3-030-93409-5_31
    ```

    """

    def __init__(
            self,
            in_channels: Union[int, Tuple[int, int]],
            out_channels: int,
            heads: int = 1,
            attention_layers: int = 1,
            concat: bool = True,
            dropout: float = 0.0,
            add_self_loops: bool = True,
            edge_dim: Optional[int] = None,
            fill_value: Union[float, Tensor, str] = 'mean',
            bias: bool = True,
            **kwargs,
    ):
        kwargs.setdefault('aggr', 'add')
        super().__init__(node_dim=0, **kwargs)

        self.in_channels = in_channels
        self.out_channels = out_channels
        self.heads = heads
        self.concat = concat
        self.dropout = dropout
        self.add_self_loops = add_self_loops
        self.edge_dim = edge_dim
        self.fill_value = fill_value

        # In case we are operating in bipartite graphs, we apply separate
        # transformations 'lin_src' and 'lin_dst' to source and target nodes:
        if isinstance(in_channels, int):
            self.lin_src = Linear(in_channels, heads * out_channels,
                                  bias=False, weight_initializer='glorot')
            self.lin_dst = self.lin_src
        else:
            self.lin_src = Linear(in_channels[0], heads * out_channels, False,
                                  weight_initializer='glorot')
            self.lin_dst = Linear(in_channels[1], heads * out_channels, False,
                                  weight_initializer='glorot')

        # The learnable parameters to compute attention coefficients:
        self.attention_layers = ParameterList()
        for i in range(attention_layers - 1):
            self.attention_layers.append(
                Parameter(torch.empty(heads, 3 * out_channels, 3 * out_channels))
            )

        self.attention_layers.append(
            Parameter(torch.empty(heads, 1, 3 * out_channels))
        )

        if edge_dim is not None:
            self.lin_edge = Linear(edge_dim, heads * out_channels, bias=False,
                                   weight_initializer='glorot')
            self.att_edge = Parameter(torch.empty(1, heads, out_channels))
        else:
            self.lin_edge = None
            self.register_parameter('att_edge', None)

        if bias and concat:
            self.bias = Parameter(torch.empty(heads * out_channels))
        elif bias and not concat:
            self.bias = Parameter(torch.empty(out_channels))
        else:
            self.register_parameter('bias', None)

        self.reset_parameters()

    def reset_parameters(self):
        super().reset_parameters()
        self.lin_src.reset_parameters()
        self.lin_dst.reset_parameters()
        if self.lin_edge is not None:
            self.lin_edge.reset_parameters()

        for layer in self.attention_layers:
            glorot(layer)

        glorot(self.att_edge)
        zeros(self.bias)

    def forward(self, x: Union[Tensor, OptPairTensor], edge_index: Adj,
                edge_attr: OptTensor = None, size: Size = None,
                return_attention_weights=None):
        # type: (Union[Tensor, OptPairTensor], Tensor, OptTensor, Size, NoneType) -> Tensor  # noqa
        # type: (Union[Tensor, OptPairTensor], SparseTensor, OptTensor, Size, NoneType) -> Tensor  # noqa
        # type: (Union[Tensor, OptPairTensor], Tensor, OptTensor, Size, bool) -> Tuple[Tensor, Tuple[Tensor, Tensor]]  # noqa
        # type: (Union[Tensor, OptPairTensor], SparseTensor, OptTensor, Size, bool) -> Tuple[Tensor, SparseTensor]  # noqa
        r"""Runs the forward pass of the module.

        Args:
            return_attention_weights (bool, optional): If set to :obj:`True`,
                will additionally return the tuple
                :obj:`(edge_index, attention_weights)`, holding the computed
                attention weights for each edge. (default: :obj:`None`)
        """
        # NOTE: attention weights will be returned whenever
        # `return_attention_weights` is set to a value, regardless of its
        # actual value (might be `True` or `False`). This is a current somewhat
        # hacky workaround to allow for TorchScript support via the
        # `torch.jit._overload` decorator, as we can only change the output
        # arguments conditioned on type (`None` or `bool`), not based on its
        # actual value.

        H, C = self.heads, self.out_channels

        # We first transform the input node features. If a tuple is passed, we
        # transform source and target node features via separate weights:
        if isinstance(x, Tensor):
            assert x.dim() == 2, "Static graphs not supported in 'wsGATConv'"
            x_src = x_dst = self.lin_src(x).view(-1, H, C)
        else:  # Tuple of source and target node features:
            x_src, x_dst = x
            assert x_src.dim() == 2, "Static graphs not supported in 'wsGATConv'"
            x_src = self.lin_src(x_src).view(-1, H, C)
            if x_dst is not None:
                x_dst = self.lin_dst(x_dst).view(-1, H, C)

        if self.add_self_loops:
            if isinstance(edge_index, Tensor):
                # We only want to add self-loops for nodes that appear both as
                # source and target nodes:
                num_nodes = x_src.size(0)
                if x_dst is not None:
                    num_nodes = min(num_nodes, x_dst.size(0))
                num_nodes = min(size) if size is not None else num_nodes
                edge_index, edge_attr = remove_self_loops(
                    edge_index, edge_attr)
                edge_index, edge_attr = add_self_loops(
                    edge_index, edge_attr, fill_value=self.fill_value,
                    num_nodes=num_nodes)
            elif isinstance(edge_index, SparseTensor):
                if self.edge_dim is None:
                    edge_index = torch_sparse.set_diag(edge_index)
                else:
                    raise NotImplementedError(
                        "The usage of 'edge_attr' and 'add_self_loops' "
                        "simultaneously is currently not yet supported for "
                        "'edge_index' in a 'SparseTensor' form")

        x = (x_src, x_dst)

        # # Next, we compute node-level attention coefficients, both for source
        # # and target nodes (if present):

        # edge_updater_type: (alpha: OptPairTensor, edge_attr: OptTensor)
        alpha = self.edge_updater(edge_index,
                                  x=x,
                                  # alpha=alpha,
                                  edge_attr=edge_attr,
                                  )

        # propagate_type: (x: OptPairTensor, alpha: Tensor)
        out = self.propagate(edge_index, x=x, alpha=alpha, size=size)

        if self.concat:
            out = out.view(-1, self.heads * self.out_channels)
        else:
            out = out.mean(dim=1)

        if self.bias is not None:
            out = out + self.bias

        if isinstance(return_attention_weights, bool):
            if isinstance(edge_index, Tensor):
                if is_torch_sparse_tensor(edge_index):
                    # TODO TorchScript requires to return a tuple
                    adj = set_sparse_value(edge_index, alpha)
                    return out, (adj, alpha)
                else:
                    return out, (edge_index, alpha)
            elif isinstance(edge_index, SparseTensor):
                return out, edge_index.set_value(alpha, layout='coo')
        else:
            return out

    def edge_update(self, x_i: Tensor, x_j: Tensor,
                    edge_attr: OptTensor, index: Tensor, ptr: OptTensor,
                    size_i: Optional[int]) -> Tensor:

        if index.numel() == 0:
            raise NotImplementedError("Empty edge index not supported yet")

        if edge_attr is not None and self.lin_edge is not None:
            if edge_attr.dim() == 1:
                edge_attr = edge_attr.view(-1, 1)
            edge_attr = self.lin_edge(edge_attr)

            edge_attr = edge_attr.view(-1, self.heads, self.out_channels)

            alpha = torch.cat([x_i, x_j, edge_attr], dim=2)
        else:
            alpha = torch.cat([x_i, x_j], dim=2)

        for i, layer in enumerate(self.attention_layers):
            alpha = torch.matmul(layer, alpha.unsqueeze(-1)).squeeze(-1)

            alpha = alpha.tanh()

        alpha = alpha.squeeze()
        alpha = alpha.sign() * softmax(alpha.abs(), index, ptr, size_i)

        alpha = F.dropout(alpha, p=self.dropout, training=self.training)

        return alpha

    def message(self, x_j: Tensor, alpha: Tensor) -> Tensor:
        return alpha.unsqueeze(-1) * x_j

    def __repr__(self) -> str:
        return (f'{self.__class__.__name__}({self.in_channels}, '
                f'{self.out_channels}, heads={self.heads}, '
                f'attention_layers={len(self.attention_layers)}')