File size: 5,800 Bytes
58f667f
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
import math

import torch
import torch.nn as nn
from mmcv.cnn import uniform_init

from .builder import POSITIONAL_ENCODING


@POSITIONAL_ENCODING.register_module()
class SinePositionalEncoding(nn.Module):
    """Position encoding with sine and cosine functions.

    See `End-to-End Object Detection with Transformers
    <https://arxiv.org/pdf/2005.12872>`_ for details.

    Args:
        num_feats (int): The feature dimension for each position
            along x-axis or y-axis. Note the final returned dimension
            for each position is 2 times of this value.
        temperature (int, optional): The temperature used for scaling
            the position embedding. Default 10000.
        normalize (bool, optional): Whether to normalize the position
            embedding. Default False.
        scale (float, optional): A scale factor that scales the position
            embedding. The scale will be used only when `normalize` is True.
            Default 2*pi.
        eps (float, optional): A value added to the denominator for
            numerical stability. Default 1e-6.
    """

    def __init__(self,
                 num_feats,
                 temperature=10000,
                 normalize=False,
                 scale=2 * math.pi,
                 eps=1e-6):
        super(SinePositionalEncoding, self).__init__()
        if normalize:
            assert isinstance(scale, (float, int)), 'when normalize is set,' \
                'scale should be provided and in float or int type, ' \
                f'found {type(scale)}'
        self.num_feats = num_feats
        self.temperature = temperature
        self.normalize = normalize
        self.scale = scale
        self.eps = eps

    def forward(self, mask):
        """Forward function for `SinePositionalEncoding`.

        Args:
            mask (Tensor): ByteTensor mask. Non-zero values representing
                ignored positions, while zero values means valid positions
                for this image. Shape [bs, h, w].

        Returns:
            pos (Tensor): Returned position embedding with shape
                [bs, num_feats*2, h, w].
        """
        not_mask = ~mask
        y_embed = not_mask.cumsum(1, dtype=torch.float32)
        x_embed = not_mask.cumsum(2, dtype=torch.float32)
        if self.normalize:
            y_embed = y_embed / (y_embed[:, -1:, :] + self.eps) * self.scale
            x_embed = x_embed / (x_embed[:, :, -1:] + self.eps) * self.scale
        dim_t = torch.arange(
            self.num_feats, dtype=torch.float32, device=mask.device)
        dim_t = self.temperature**(2 * (dim_t // 2) / self.num_feats)
        pos_x = x_embed[:, :, :, None] / dim_t
        pos_y = y_embed[:, :, :, None] / dim_t
        pos_x = torch.stack(
            (pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()),
            dim=4).flatten(3)
        pos_y = torch.stack(
            (pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()),
            dim=4).flatten(3)
        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
        return pos

    def __repr__(self):
        """str: a string that describes the module"""
        repr_str = self.__class__.__name__
        repr_str += f'(num_feats={self.num_feats}, '
        repr_str += f'temperature={self.temperature}, '
        repr_str += f'normalize={self.normalize}, '
        repr_str += f'scale={self.scale}, '
        repr_str += f'eps={self.eps})'
        return repr_str


@POSITIONAL_ENCODING.register_module()
class LearnedPositionalEncoding(nn.Module):
    """Position embedding with learnable embedding weights.

    Args:
        num_feats (int): The feature dimension for each position
            along x-axis or y-axis. The final returned dimension for
            each position is 2 times of this value.
        row_num_embed (int, optional): The dictionary size of row embeddings.
            Default 50.
        col_num_embed (int, optional): The dictionary size of col embeddings.
            Default 50.
    """

    def __init__(self, num_feats, row_num_embed=50, col_num_embed=50):
        super(LearnedPositionalEncoding, self).__init__()
        self.row_embed = nn.Embedding(row_num_embed, num_feats)
        self.col_embed = nn.Embedding(col_num_embed, num_feats)
        self.num_feats = num_feats
        self.row_num_embed = row_num_embed
        self.col_num_embed = col_num_embed
        self.init_weights()

    def init_weights(self):
        """Initialize the learnable weights."""
        uniform_init(self.row_embed)
        uniform_init(self.col_embed)

    def forward(self, mask):
        """Forward function for `LearnedPositionalEncoding`.

        Args:
            mask (Tensor): ByteTensor mask. Non-zero values representing
                ignored positions, while zero values means valid positions
                for this image. Shape [bs, h, w].

        Returns:
            pos (Tensor): Returned position embedding with shape
                [bs, num_feats*2, h, w].
        """
        h, w = mask.shape[-2:]
        x = torch.arange(w, device=mask.device)
        y = torch.arange(h, device=mask.device)
        x_embed = self.col_embed(x)
        y_embed = self.row_embed(y)
        pos = torch.cat(
            (x_embed.unsqueeze(0).repeat(h, 1, 1), y_embed.unsqueeze(1).repeat(
                1, w, 1)),
            dim=-1).permute(2, 0,
                            1).unsqueeze(0).repeat(mask.shape[0], 1, 1, 1)
        return pos

    def __repr__(self):
        """str: a string that describes the module"""
        repr_str = self.__class__.__name__
        repr_str += f'(num_feats={self.num_feats}, '
        repr_str += f'row_num_embed={self.row_num_embed}, '
        repr_str += f'col_num_embed={self.col_num_embed})'
        return repr_str