File size: 5,395 Bytes
7344bef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved

# pyre-unsafe

import math
from typing import Optional

import torch
from torch import nn


class PositionEmbeddingSine(nn.Module):
    """
    This is a more standard version of the position embedding, very similar to the one
    used by the Attention is all you need paper, generalized to work on images.
    """

    def __init__(
        self,
        num_pos_feats,
        temperature: int = 10000,
        normalize: bool = True,
        scale: Optional[float] = None,
        precompute_resolution: Optional[int] = None,
    ):
        super().__init__()
        assert num_pos_feats % 2 == 0, "Expecting even model width"
        self.num_pos_feats = num_pos_feats // 2
        self.temperature = temperature
        self.normalize = normalize
        if scale is not None and normalize is False:
            raise ValueError("normalize should be True if scale is passed")
        if scale is None:
            scale = 2 * math.pi
        self.scale = scale

        self.cache = {}
        # Precompute positional encodings under `precompute_resolution` to fill the cache
        # and avoid symbolic shape tracing errors in torch.compile in PyTorch 2.4 nightly.
        if precompute_resolution is not None:
            # We precompute pos enc for all strides used by both DualViTDetNeck and
            # TriViTDetNeck (scale_factors 4.0, 2.0, 1.0, 0.5 applied to backbone
            # output at stride 14 from 1008px input → 72x72).
            precompute_sizes = [
                (int(precompute_resolution // 3.5), int(precompute_resolution // 3.5)),
                (precompute_resolution // 4, precompute_resolution // 4),
                (int(precompute_resolution // 7), int(precompute_resolution // 7)),
                (precompute_resolution // 8, precompute_resolution // 8),
                (int(precompute_resolution // 14), int(precompute_resolution // 14)),
                (precompute_resolution // 16, precompute_resolution // 16),
                (int(precompute_resolution // 28), int(precompute_resolution // 28)),
                (precompute_resolution // 32, precompute_resolution // 32),
            ]
            for size in precompute_sizes:
                tensors = torch.zeros((1, 1) + size, device="cpu")
                self.forward(tensors)
                # further clone and detach it in the cache (just to be safe)
                self.cache[size] = self.cache[size].clone().detach()

    def _encode_xy(self, x, y):
        # The positions are expected to be normalized
        assert len(x) == len(y) and x.ndim == y.ndim == 1
        x_embed = x * self.scale
        y_embed = y * self.scale

        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_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=2
        ).flatten(1)
        pos_y = torch.stack(
            (pos_y[:, 0::2].sin(), pos_y[:, 1::2].cos()), dim=2
        ).flatten(1)
        return pos_x, pos_y

    @torch.no_grad()
    def encode_boxes(self, x, y, w, h):
        pos_x, pos_y = self._encode_xy(x, y)
        pos = torch.cat((pos_y, pos_x, h[:, None], w[:, None]), dim=1)
        return pos

    encode = encode_boxes  # Backwards compatibility

    @torch.no_grad()
    def encode_points(self, x, y, labels):
        (bx, nx), (by, ny), (bl, nl) = x.shape, y.shape, labels.shape
        assert bx == by and nx == ny and bx == bl and nx == nl
        pos_x, pos_y = self._encode_xy(x.flatten(), y.flatten())
        pos_x, pos_y = pos_x.reshape(bx, nx, -1), pos_y.reshape(by, ny, -1)
        pos = torch.cat((pos_y, pos_x, labels[:, :, None]), dim=2)
        return pos

    @torch.no_grad()
    def forward(self, x):
        cache_key = None
        cache_key = (x.shape[-2], x.shape[-1])
        if cache_key in self.cache:
            return self.cache[cache_key].to(device=x.device)[None].repeat(x.shape[0], 1, 1, 1)
        y_embed = (
            torch.arange(1, x.shape[-2] + 1, dtype=torch.float32, device=x.device)
            .view(1, -1, 1)
            .repeat(x.shape[0], 1, x.shape[-1])
        )
        x_embed = (
            torch.arange(1, x.shape[-1] + 1, dtype=torch.float32, device=x.device)
            .view(1, 1, -1)
            .repeat(x.shape[0], x.shape[-2], 1)
        )

        if self.normalize:
            eps = 1e-6
            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale

        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_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)
        if cache_key is not None:
            self.cache[cache_key] = pos[0]
        return pos