transformer.py

# Copyright (c) OpenMMLab. All rights reserved.
from mmcv.cnn.bricks.transformer import MultiheadAttention
from mmengine.registry import MODELS
from torch import nn as nn


@MODELS.register_module()
class GroupFree3DMHA(MultiheadAttention):
    """A wrapper for torch.nn.MultiheadAttention for GroupFree3D.

    This module implements MultiheadAttention with identity connection,
    and positional encoding used in DETR is also passed as input.

    Args:
        embed_dims (int): The embedding dimension.
        num_heads (int): Parallel attention heads. Same as
            `nn.MultiheadAttention`.
        attn_drop (float, optional): A Dropout layer on attn_output_weights.
            Defaults to 0.0.
        proj_drop (float, optional): A Dropout layer. Defaults to 0.0.
        dropout_layer (obj:`ConfigDict`, optional): The dropout_layer used
            when adding the shortcut.
        init_cfg (obj:`mmengine.ConfigDict`, optional): The Config for
            initialization. Default: None.
        batch_first (bool, optional): Key, Query and Value are shape of
            (batch, n, embed_dim)
            or (n, batch, embed_dim). Defaults to False.
    """

    def __init__(self,
                 embed_dims,
                 num_heads,
                 attn_drop=0.,
                 proj_drop=0.,
                 dropout_layer=dict(type='DropOut', drop_prob=0.),
                 init_cfg=None,
                 batch_first=False,
                 **kwargs):
        super().__init__(embed_dims, num_heads, attn_drop, proj_drop,
                         dropout_layer, init_cfg, batch_first, **kwargs)

    def forward(self,
                query,
                key,
                value,
                identity,
                query_pos=None,
                key_pos=None,
                attn_mask=None,
                key_padding_mask=None,
                **kwargs):
        """Forward function for `GroupFree3DMHA`.

        **kwargs allow passing a more general data flow when combining
        with other operations in `transformerlayer`.

        Args:
            query (Tensor): The input query with shape [num_queries, bs,
                embed_dims]. Same in `nn.MultiheadAttention.forward`.
            key (Tensor): The key tensor with shape [num_keys, bs,
                embed_dims]. Same in `nn.MultiheadAttention.forward`.
                If None, the ``query`` will be used.
            value (Tensor): The value tensor with same shape as `key`.
                Same in `nn.MultiheadAttention.forward`.
                If None, the `key` will be used.
            identity (Tensor): This tensor, with the same shape as x,
                will be used for the identity link. If None, `x` will be used.
            query_pos (Tensor, optional): The positional encoding for query,
                with the same shape as `x`. Defaults to None.
                If not None, it will be added to `x` before forward function.
            key_pos (Tensor, optional): The positional encoding for `key`,
                with the same shape as `key`. Defaults to None. If not None,
                it will be added to `key` before forward function. If None,
                and `query_pos` has the same shape as `key`, then `query_pos`
                will be used for `key_pos`. Defaults to None.
            attn_mask (Tensor, optional): ByteTensor mask with shape
                [num_queries, num_keys].
                Same in `nn.MultiheadAttention.forward`. Defaults to None.
            key_padding_mask (Tensor, optional): ByteTensor with shape
                [bs, num_keys]. Same in `nn.MultiheadAttention.forward`.
                Defaults to None.

        Returns:
            Tensor: forwarded results with shape [num_queries, bs, embed_dims].
        """

        if hasattr(self, 'operation_name'):
            if self.operation_name == 'self_attn':
                value = value + query_pos
            elif self.operation_name == 'cross_attn':
                value = value + key_pos
            else:
                raise NotImplementedError(
                    f'{self.__class__.name} '
                    f"can't be used as {self.operation_name}")
        else:
            value = value + query_pos

        return super(GroupFree3DMHA, self).forward(
            query=query,
            key=key,
            value=value,
            identity=identity,
            query_pos=query_pos,
            key_pos=key_pos,
            attn_mask=attn_mask,
            key_padding_mask=key_padding_mask,
            **kwargs)


@MODELS.register_module()
class ConvBNPositionalEncoding(nn.Module):
    """Absolute position embedding with Conv learning.

    Args:
        input_channel (int): input features dim.
        num_pos_feats (int, optional): output position features dim.
            Defaults to 288 to be consistent with seed features dim.
    """

    def __init__(self, input_channel, num_pos_feats=288):
        super().__init__()
        self.position_embedding_head = nn.Sequential(
            nn.Conv1d(input_channel, num_pos_feats, kernel_size=1),
            nn.BatchNorm1d(num_pos_feats), nn.ReLU(inplace=True),
            nn.Conv1d(num_pos_feats, num_pos_feats, kernel_size=1))

    def forward(self, xyz):
        """Forward pass.

        Args:
            xyz (Tensor)： (B, N, 3) the coordinates to embed.

        Returns:
            Tensor: (B, num_pos_feats, N) the embedded position features.
        """
        xyz = xyz.permute(0, 2, 1)
        position_embedding = self.position_embedding_head(xyz)
        return position_embedding