File size: 10,859 Bytes
4121bec
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
# Copyright (c) Facebook, Inc. and its affiliates.
import math
from typing import List
import torch
from torch import nn
from torchvision.ops import RoIPool

from detectron2.layers import ROIAlign, ROIAlignRotated, cat, nonzero_tuple
from detectron2.structures import Boxes

"""
To export ROIPooler to torchscript, in this file, variables that should be annotated with
`Union[List[Boxes], List[RotatedBoxes]]` are only annotated with `List[Boxes]`.

TODO: Correct these annotations when torchscript support `Union`.
https://github.com/pytorch/pytorch/issues/41412
"""

__all__ = ["ROIPooler"]


def assign_boxes_to_levels(
    box_lists: List[Boxes],
    min_level: int,
    max_level: int,
    canonical_box_size: int,
    canonical_level: int,
):
    """
    Map each box in `box_lists` to a feature map level index and return the assignment
    vector.

    Args:
        box_lists (list[Boxes] | list[RotatedBoxes]): A list of N Boxes or N RotatedBoxes,
            where N is the number of images in the batch.
        min_level (int): Smallest feature map level index. The input is considered index 0,
            the output of stage 1 is index 1, and so.
        max_level (int): Largest feature map level index.
        canonical_box_size (int): A canonical box size in pixels (sqrt(box area)).
        canonical_level (int): The feature map level index on which a canonically-sized box
            should be placed.

    Returns:
        A tensor of length M, where M is the total number of boxes aggregated over all
            N batch images. The memory layout corresponds to the concatenation of boxes
            from all images. Each element is the feature map index, as an offset from
            `self.min_level`, for the corresponding box (so value i means the box is at
            `self.min_level + i`).
    """
    box_sizes = torch.sqrt(cat([boxes.area() for boxes in box_lists]))
    # Eqn.(1) in FPN paper
    level_assignments = torch.floor(
        canonical_level + torch.log2(box_sizes / canonical_box_size + 1e-8)
    )
    # clamp level to (min, max), in case the box size is too large or too small
    # for the available feature maps
    level_assignments = torch.clamp(level_assignments, min=min_level, max=max_level)
    return level_assignments.to(torch.int64) - min_level


def _fmt_box_list(box_tensor, batch_index: int):
    repeated_index = torch.full_like(
        box_tensor[:, :1], batch_index, dtype=box_tensor.dtype, device=box_tensor.device
    )
    return cat((repeated_index, box_tensor), dim=1)


def convert_boxes_to_pooler_format(box_lists: List[Boxes]):
    """
    Convert all boxes in `box_lists` to the low-level format used by ROI pooling ops
    (see description under Returns).

    Args:
        box_lists (list[Boxes] | list[RotatedBoxes]):
            A list of N Boxes or N RotatedBoxes, where N is the number of images in the batch.

    Returns:
        When input is list[Boxes]:
            A tensor of shape (M, 5), where M is the total number of boxes aggregated over all
            N batch images.
            The 5 columns are (batch index, x0, y0, x1, y1), where batch index
            is the index in [0, N) identifying which batch image the box with corners at
            (x0, y0, x1, y1) comes from.
        When input is list[RotatedBoxes]:
            A tensor of shape (M, 6), where M is the total number of boxes aggregated over all
            N batch images.
            The 6 columns are (batch index, x_ctr, y_ctr, width, height, angle_degrees),
            where batch index is the index in [0, N) identifying which batch image the
            rotated box (x_ctr, y_ctr, width, height, angle_degrees) comes from.
    """
    pooler_fmt_boxes = cat(
        [_fmt_box_list(box_list.tensor, i) for i, box_list in enumerate(box_lists)], dim=0
    )

    return pooler_fmt_boxes


class ROIPooler(nn.Module):
    """
    Region of interest feature map pooler that supports pooling from one or more
    feature maps.
    """

    def __init__(
        self,
        output_size,
        scales,
        sampling_ratio,
        pooler_type,
        canonical_box_size=224,
        canonical_level=4,
    ):
        """
        Args:
            output_size (int, tuple[int] or list[int]): output size of the pooled region,
                e.g., 14 x 14. If tuple or list is given, the length must be 2.
            scales (list[float]): The scale for each low-level pooling op relative to
                the input image. For a feature map with stride s relative to the input
                image, scale is defined as 1/s. The stride must be power of 2.
                When there are multiple scales, they must form a pyramid, i.e. they must be
                a monotically decreasing geometric sequence with a factor of 1/2.
            sampling_ratio (int): The `sampling_ratio` parameter for the ROIAlign op.
            pooler_type (string): Name of the type of pooling operation that should be applied.
                For instance, "ROIPool" or "ROIAlignV2".
            canonical_box_size (int): A canonical box size in pixels (sqrt(box area)). The default
                is heuristically defined as 224 pixels in the FPN paper (based on ImageNet
                pre-training).
            canonical_level (int): The feature map level index from which a canonically-sized box
                should be placed. The default is defined as level 4 (stride=16) in the FPN paper,
                i.e., a box of size 224x224 will be placed on the feature with stride=16.
                The box placement for all boxes will be determined from their sizes w.r.t
                canonical_box_size. For example, a box whose area is 4x that of a canonical box
                should be used to pool features from feature level ``canonical_level+1``.

                Note that the actual input feature maps given to this module may not have
                sufficiently many levels for the input boxes. If the boxes are too large or too
                small for the input feature maps, the closest level will be used.
        """
        super().__init__()

        if isinstance(output_size, int):
            output_size = (output_size, output_size)
        assert len(output_size) == 2
        assert isinstance(output_size[0], int) and isinstance(output_size[1], int)
        self.output_size = output_size

        if pooler_type == "ROIAlign":
            self.level_poolers = nn.ModuleList(
                ROIAlign(
                    output_size, spatial_scale=scale, sampling_ratio=sampling_ratio, aligned=False
                )
                for scale in scales
            )
        elif pooler_type == "ROIAlignV2":
            self.level_poolers = nn.ModuleList(
                ROIAlign(
                    output_size, spatial_scale=scale, sampling_ratio=sampling_ratio, aligned=True
                )
                for scale in scales
            )
        elif pooler_type == "ROIPool":
            self.level_poolers = nn.ModuleList(
                RoIPool(output_size, spatial_scale=scale) for scale in scales
            )
        elif pooler_type == "ROIAlignRotated":
            self.level_poolers = nn.ModuleList(
                ROIAlignRotated(output_size, spatial_scale=scale, sampling_ratio=sampling_ratio)
                for scale in scales
            )
        else:
            raise ValueError("Unknown pooler type: {}".format(pooler_type))

        # Map scale (defined as 1 / stride) to its feature map level under the
        # assumption that stride is a power of 2.
        min_level = -(math.log2(scales[0]))
        max_level = -(math.log2(scales[-1]))
        assert math.isclose(min_level, int(min_level)) and math.isclose(
            max_level, int(max_level)
        ), "Featuremap stride is not power of 2!"
        self.min_level = int(min_level)
        self.max_level = int(max_level)
        assert (
            len(scales) == self.max_level - self.min_level + 1
        ), "[ROIPooler] Sizes of input featuremaps do not form a pyramid!"
        assert 0 <= self.min_level and self.min_level <= self.max_level
        self.canonical_level = canonical_level
        assert canonical_box_size > 0
        self.canonical_box_size = canonical_box_size

    def forward(self, x: List[torch.Tensor], box_lists: List[Boxes]):
        """
        Args:
            x (list[Tensor]): A list of feature maps of NCHW shape, with scales matching those
                used to construct this module.
            box_lists (list[Boxes] | list[RotatedBoxes]):
                A list of N Boxes or N RotatedBoxes, where N is the number of images in the batch.
                The box coordinates are defined on the original image and
                will be scaled by the `scales` argument of :class:`ROIPooler`.

        Returns:
            Tensor:
                A tensor of shape (M, C, output_size, output_size) where M is the total number of
                boxes aggregated over all N batch images and C is the number of channels in `x`.
        """
        num_level_assignments = len(self.level_poolers)

        assert isinstance(x, list) and isinstance(
            box_lists, list
        ), "Arguments to pooler must be lists"
        assert (
            len(x) == num_level_assignments
        ), "unequal value, num_level_assignments={}, but x is list of {} Tensors".format(
            num_level_assignments, len(x)
        )

        assert len(box_lists) == x[0].size(
            0
        ), "unequal value, x[0] batch dim 0 is {}, but box_list has length {}".format(
            x[0].size(0), len(box_lists)
        )
        if len(box_lists) == 0:
            return torch.zeros(
                (0, x[0].shape[1]) + self.output_size, device=x[0].device, dtype=x[0].dtype
            )

        pooler_fmt_boxes = convert_boxes_to_pooler_format(box_lists)

        if num_level_assignments == 1:
            return self.level_poolers[0](x[0], pooler_fmt_boxes)

        level_assignments = assign_boxes_to_levels(
            box_lists, self.min_level, self.max_level, self.canonical_box_size, self.canonical_level
        )

        num_boxes = pooler_fmt_boxes.size(0)
        num_channels = x[0].shape[1]
        output_size = self.output_size[0]

        dtype, device = x[0].dtype, x[0].device
        output = torch.zeros(
            (num_boxes, num_channels, output_size, output_size), dtype=dtype, device=device
        )

        for level, pooler in enumerate(self.level_poolers):
            inds = nonzero_tuple(level_assignments == level)[0]
            pooler_fmt_boxes_level = pooler_fmt_boxes[inds]
            # Use index_put_ instead of advance indexing, to avoid pytorch/issues/49852
            output.index_put_((inds,), pooler(x[level], pooler_fmt_boxes_level))

        return output