Source code for torch_geometric.nn.conv.point_transformer_conv

from typing import Callable, Optional, Tuple, Union

from torch import Tensor

from torch_geometric.nn.conv import MessagePassing
from torch_geometric.nn.dense.linear import Linear
from torch_geometric.nn.inits import reset
from torch_geometric.typing import (
    Adj,
    OptTensor,
    PairTensor,
    SparseTensor,
    torch_sparse,
)
from torch_geometric.utils import add_self_loops, remove_self_loops, softmax


class PointTransformerConv(MessagePassing):
    r"""The Point Transformer layer from the `"Point Transformer"
    <https://arxiv.org/abs/2012.09164>`_ paper.

    .. math::
        \mathbf{x}^{\prime}_i =  \sum_{j \in
        \mathcal{N}(i) \cup \{ i \}} \alpha_{i,j} \left(\mathbf{W}_3
        \mathbf{x}_j + \delta_{ij} \right),

    where the attention coefficients :math:`\alpha_{i,j}` and
    positional embedding :math:`\delta_{ij}` are computed as

    .. math::
        \alpha_{i,j}= \textrm{softmax} \left( \gamma_\mathbf{\Theta}
        (\mathbf{W}_1 \mathbf{x}_i - \mathbf{W}_2 \mathbf{x}_j +
        \delta_{i,j}) \right)

    and

    .. math::
        \delta_{i,j}= h_{\mathbf{\Theta}}(\mathbf{p}_i - \mathbf{p}_j),

    with :math:`\gamma_\mathbf{\Theta}` and :math:`h_\mathbf{\Theta}`
    denoting neural networks, *i.e.* MLPs, and
    :math:`\mathbf{P} \in \mathbb{R}^{N \times D}` defines the position of
    each point.

    Args:
        in_channels (int or tuple): Size of each input sample, or :obj:`-1` to
            derive the size from the first input(s) to the forward method.
            A tuple corresponds to the sizes of source and target
            dimensionalities.
        out_channels (int): Size of each output sample.
        pos_nn (torch.nn.Module, optional): A neural network
            :math:`h_\mathbf{\Theta}` which maps relative spatial coordinates
            :obj:`pos_j - pos_i` of shape :obj:`[-1, 3]` to shape
            :obj:`[-1, out_channels]`.
            Will default to a :class:`torch.nn.Linear` transformation if not
            further specified. (default: :obj:`None`)
        attn_nn (torch.nn.Module, optional): A neural network
            :math:`\gamma_\mathbf{\Theta}` which maps transformed
            node features of shape :obj:`[-1, out_channels]`
            to shape :obj:`[-1, out_channels]`. (default: :obj:`None`)
        add_self_loops (bool, optional) : If set to :obj:`False`, will not add
            self-loops to the input graph. (default: :obj:`True`)
        **kwargs (optional): Additional arguments of
            :class:`torch_geometric.nn.conv.MessagePassing`.

    Shapes:
        - **input:**
          node features :math:`(|\mathcal{V}|, F_{in})` or
          :math:`((|\mathcal{V_s}|, F_{s}), (|\mathcal{V_t}|, F_{t}))`
          if bipartite,
          positions :math:`(|\mathcal{V}|, 3)` or
          :math:`((|\mathcal{V_s}|, 3), (|\mathcal{V_t}|, 3))` if bipartite,
          edge indices :math:`(2, |\mathcal{E}|)`
        - **output:** node features :math:`(|\mathcal{V}|, F_{out})` or
          :math:`(|\mathcal{V}_t|, F_{out})` if bipartite
    """
    def __init__(self, in_channels: Union[int, Tuple[int, int]],
                 out_channels: int, pos_nn: Optional[Callable] = None,
                 attn_nn: Optional[Callable] = None,
                 add_self_loops: bool = True, **kwargs):
        kwargs.setdefault('aggr', 'add')
        super().__init__(**kwargs)

        self.in_channels = in_channels
        self.out_channels = out_channels
        self.add_self_loops = add_self_loops

        if isinstance(in_channels, int):
            in_channels = (in_channels, in_channels)

        self.pos_nn = pos_nn
        if self.pos_nn is None:
            self.pos_nn = Linear(3, out_channels)

        self.attn_nn = attn_nn
        self.lin = Linear(in_channels[0], out_channels, bias=False)
        self.lin_src = Linear(in_channels[0], out_channels, bias=False)
        self.lin_dst = Linear(in_channels[1], out_channels, bias=False)

        self.reset_parameters()

    def reset_parameters(self):
        super().reset_parameters()
        reset(self.pos_nn)
        if self.attn_nn is not None:
            reset(self.attn_nn)
        self.lin.reset_parameters()
        self.lin_src.reset_parameters()
        self.lin_dst.reset_parameters()

    def forward(
        self,
        x: Union[Tensor, PairTensor],
        pos: Union[Tensor, PairTensor],
        edge_index: Adj,
    ) -> Tensor:

        if isinstance(x, Tensor):
            alpha = (self.lin_src(x), self.lin_dst(x))
            x = (self.lin(x), x)
        else:
            alpha = (self.lin_src(x[0]), self.lin_dst(x[1]))
            x = (self.lin(x[0]), x[1])

        if isinstance(pos, Tensor):
            pos = (pos, pos)

        if self.add_self_loops:
            if isinstance(edge_index, Tensor):
                edge_index, _ = remove_self_loops(edge_index)
                edge_index, _ = add_self_loops(
                    edge_index, num_nodes=min(pos[0].size(0), pos[1].size(0)))
            elif isinstance(edge_index, SparseTensor):
                edge_index = torch_sparse.set_diag(edge_index)

        # propagate_type: (x: PairTensor, pos: PairTensor, alpha: PairTensor)
        out = self.propagate(edge_index, x=x, pos=pos, alpha=alpha)
        return out

    def message(self, x_j: Tensor, pos_i: Tensor, pos_j: Tensor,
                alpha_i: Tensor, alpha_j: Tensor, index: Tensor,
                ptr: OptTensor, size_i: Optional[int]) -> Tensor:

        delta = self.pos_nn(pos_i - pos_j)
        alpha = alpha_i - alpha_j + delta
        if self.attn_nn is not None:
            alpha = self.attn_nn(alpha)
        alpha = softmax(alpha, index, ptr, size_i)
        return alpha * (x_j + delta)

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