File size: 12,467 Bytes
fa0537d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
# --------------------------------------------------------
# InternImage
# Copyright (c) 2025 OpenGVLab
# Licensed under The MIT License [see LICENSE for details]
# --------------------------------------------------------

from __future__ import absolute_import, division, print_function

import warnings

import torch
import torch.nn.functional as F
from torch import nn
from torch.nn.init import constant_, xavier_uniform_

from .dcnv3_func import DCNv3Function, dcnv3_core_pytorch, has_cuda_kernel


class to_channels_first(nn.Module):

    def __init__(self):
        super().__init__()

    def forward(self, x):
        return x.permute(0, 3, 1, 2)


class to_channels_last(nn.Module):

    def __init__(self):
        super().__init__()

    def forward(self, x):
        return x.permute(0, 2, 3, 1)


def build_norm_layer(dim,
                     norm_layer,
                     in_format='channels_last',
                     out_format='channels_last',
                     eps=1e-6):
    layers = []
    if norm_layer == 'BN':
        if in_format == 'channels_last':
            layers.append(to_channels_first())
        layers.append(nn.BatchNorm2d(dim))
        if out_format == 'channels_last':
            layers.append(to_channels_last())
    elif norm_layer == 'LN':
        if in_format == 'channels_first':
            layers.append(to_channels_last())
        layers.append(nn.LayerNorm(dim, eps=eps))
        if out_format == 'channels_first':
            layers.append(to_channels_first())
    else:
        raise NotImplementedError(
            f'build_norm_layer does not support {norm_layer}')
    return nn.Sequential(*layers)


def build_act_layer(act_layer):
    if act_layer == 'ReLU':
        return nn.ReLU(inplace=True)
    elif act_layer == 'SiLU':
        return nn.SiLU(inplace=True)
    elif act_layer == 'GELU':
        return nn.GELU()

    raise NotImplementedError(f'build_act_layer does not support {act_layer}')


def _is_power_of_2(n):
    if (not isinstance(n, int)) or (n < 0):
        raise ValueError(
            'invalid input for _is_power_of_2: {} (type: {})'.format(n, type(n)))

    return (n & (n - 1) == 0) and n != 0


class CenterFeatureScaleModule(nn.Module):
    def forward(self,
                query,
                center_feature_scale_proj_weight,
                center_feature_scale_proj_bias):
        center_feature_scale = F.linear(query,
                                        weight=center_feature_scale_proj_weight,
                                        bias=center_feature_scale_proj_bias).sigmoid()
        return center_feature_scale


class DCNv3_pytorch(nn.Module):
    def __init__(
            self,
            channels=64,
            kernel_size=3,
            dw_kernel_size=None,
            stride=1,
            pad=1,
            dilation=1,
            group=4,
            offset_scale=1.0,
            act_layer='GELU',
            norm_layer='LN',
            center_feature_scale=False,
            remove_center=False,
    ):
        """
        DCNv3 Module
        :param channels
        :param kernel_size
        :param stride
        :param pad
        :param dilation
        :param group
        :param offset_scale
        :param act_layer
        :param norm_layer
        """
        super().__init__()
        if channels % group != 0:
            raise ValueError(
                f'channels must be divisible by group, but got {channels} and {group}')
        _d_per_group = channels // group
        dw_kernel_size = dw_kernel_size if dw_kernel_size is not None else kernel_size
        # you'd better set _d_per_group to a power of 2 which is more efficient in our CUDA implementation
        if not _is_power_of_2(_d_per_group):
            warnings.warn(
                "You'd better set channels in DCNv3 to make the dimension of each attention head a power of 2 "
                'which is more efficient in our CUDA implementation.')

        self.offset_scale = offset_scale
        self.channels = channels
        self.kernel_size = kernel_size
        self.dw_kernel_size = dw_kernel_size
        self.stride = stride
        self.dilation = dilation
        self.pad = pad
        self.group = group
        self.group_channels = channels // group
        self.offset_scale = offset_scale
        self.center_feature_scale = center_feature_scale
        self.remove_center = int(remove_center)

        self.dw_conv = nn.Sequential(
            nn.Conv2d(
                channels,
                channels,
                kernel_size=dw_kernel_size,
                stride=1,
                padding=(dw_kernel_size - 1) // 2,
                groups=channels),
            build_norm_layer(
                channels,
                norm_layer,
                'channels_first',
                'channels_last'),
            build_act_layer(act_layer))
        self.offset = nn.Linear(
            channels,
            group * (kernel_size * kernel_size - remove_center) * 2)
        self.mask = nn.Linear(
            channels,
            group * (kernel_size * kernel_size - remove_center))
        self.input_proj = nn.Linear(channels, channels)
        self.output_proj = nn.Linear(channels, channels)
        self._reset_parameters()

        if center_feature_scale:
            self.center_feature_scale_proj_weight = nn.Parameter(
                torch.zeros((group, channels), dtype=torch.float))
            self.center_feature_scale_proj_bias = nn.Parameter(
                torch.tensor(0.0, dtype=torch.float).view((1,)).repeat(group, ))
            self.center_feature_scale_module = CenterFeatureScaleModule()

    def _reset_parameters(self):
        constant_(self.offset.weight.data, 0.)
        constant_(self.offset.bias.data, 0.)
        constant_(self.mask.weight.data, 0.)
        constant_(self.mask.bias.data, 0.)
        xavier_uniform_(self.input_proj.weight.data)
        constant_(self.input_proj.bias.data, 0.)
        xavier_uniform_(self.output_proj.weight.data)
        constant_(self.output_proj.bias.data, 0.)

    def forward(self, input):
        """
        :param query                       (N, H, W, C)
        :return output                     (N, H, W, C)
        """
        N, H, W, _ = input.shape

        x = self.input_proj(input)
        x_proj = x

        x1 = input.permute(0, 3, 1, 2)
        x1 = self.dw_conv(x1)
        offset = self.offset(x1)
        mask = self.mask(x1).reshape(N, H, W, self.group, -1)
        mask = F.softmax(mask, -1).reshape(N, H, W, -1)

        x = dcnv3_core_pytorch(
            x, offset, mask,
            self.kernel_size, self.kernel_size,
            self.stride, self.stride,
            self.pad, self.pad,
            self.dilation, self.dilation,
            self.group, self.group_channels,
            self.offset_scale, self.remove_center)
        if self.center_feature_scale:
            center_feature_scale = self.center_feature_scale_module(
                x1, self.center_feature_scale_proj_weight, self.center_feature_scale_proj_bias)
            # N, H, W, groups -> N, H, W, groups, 1 -> N, H, W, groups, _d_per_group -> N, H, W, channels
            center_feature_scale = center_feature_scale[..., None].repeat(
                1, 1, 1, 1, self.channels // self.group).flatten(-2)
            x = x * (1 - center_feature_scale) + x_proj * center_feature_scale
        x = self.output_proj(x)

        return x


class DCNv3(nn.Module):
    def __init__(
        self,
        channels=64,
        kernel_size=3,
        dw_kernel_size=None,
        stride=1,
        pad=1,
        dilation=1,
        group=4,
        offset_scale=1.0,
        act_layer='GELU',
        norm_layer='LN',
        center_feature_scale=False,
        remove_center=False,
    ):
        """
        DCNv3 Module
        :param channels
        :param kernel_size
        :param stride
        :param pad
        :param dilation
        :param group
        :param offset_scale
        :param act_layer
        :param norm_layer
        """
        super().__init__()
        if channels % group != 0:
            raise ValueError(
                f'channels must be divisible by group, but got {channels} and {group}')
        _d_per_group = channels // group
        dw_kernel_size = dw_kernel_size if dw_kernel_size is not None else kernel_size
        # you'd better set _d_per_group to a power of 2 which is more efficient in our CUDA implementation
        if not _is_power_of_2(_d_per_group):
            warnings.warn(
                "You'd better set channels in DCNv3 to make the dimension of each attention head a power of 2 "
                'which is more efficient in our CUDA implementation.')

        self.offset_scale = offset_scale
        self.channels = channels
        self.kernel_size = kernel_size
        self.dw_kernel_size = dw_kernel_size
        self.stride = stride
        self.dilation = dilation
        self.pad = pad
        self.group = group
        self.group_channels = channels // group
        self.offset_scale = offset_scale
        self.center_feature_scale = center_feature_scale
        self.remove_center = int(remove_center)

        if self.remove_center and self.kernel_size % 2 == 0:
            raise ValueError('remove_center is only compatible with odd kernel size.')

        self.dw_conv = nn.Sequential(
            nn.Conv2d(
                channels,
                channels,
                kernel_size=dw_kernel_size,
                stride=1,
                padding=(dw_kernel_size - 1) // 2,
                groups=channels),
            build_norm_layer(
                channels,
                norm_layer,
                'channels_first',
                'channels_last'),
            build_act_layer(act_layer))
        self.offset = nn.Linear(
            channels,
            group * (kernel_size * kernel_size - remove_center) * 2)
        self.mask = nn.Linear(
            channels,
            group * (kernel_size * kernel_size - remove_center))
        self.input_proj = nn.Linear(channels, channels)
        self.output_proj = nn.Linear(channels, channels)
        self._reset_parameters()

        if center_feature_scale:
            self.center_feature_scale_proj_weight = nn.Parameter(
                torch.zeros((group, channels), dtype=torch.float))
            self.center_feature_scale_proj_bias = nn.Parameter(
                torch.tensor(0.0, dtype=torch.float).view((1,)).repeat(group, ))
            self.center_feature_scale_module = CenterFeatureScaleModule()

    def _reset_parameters(self):
        constant_(self.offset.weight.data, 0.)
        constant_(self.offset.bias.data, 0.)
        constant_(self.mask.weight.data, 0.)
        constant_(self.mask.bias.data, 0.)
        xavier_uniform_(self.input_proj.weight.data)
        constant_(self.input_proj.bias.data, 0.)
        xavier_uniform_(self.output_proj.weight.data)
        constant_(self.output_proj.bias.data, 0.)

    def forward(self, input):
        """
        :param query                       (N, H, W, C)
        :return output                     (N, H, W, C)
        """
        N, H, W, _ = input.shape

        x = self.input_proj(input)
        x_proj = x
        dtype = x.dtype

        x1 = input.permute(0, 3, 1, 2)
        x1 = self.dw_conv(x1)
        offset = self.offset(x1)
        mask = self.mask(x1).reshape(N, H, W, self.group, -1)
        mask = F.softmax(mask, -1)
        mask = mask.reshape(N, H, W, -1).type(dtype)

        x = DCNv3Function.apply(
            x, offset, mask,
            self.kernel_size, self.kernel_size,
            self.stride, self.stride,
            self.pad, self.pad,
            self.dilation, self.dilation,
            self.group, self.group_channels,
            self.offset_scale,
            256,
            self.remove_center)

        if self.center_feature_scale:
            center_feature_scale = self.center_feature_scale_module(
                x1, self.center_feature_scale_proj_weight, self.center_feature_scale_proj_bias)
            # N, H, W, groups -> N, H, W, groups, 1 -> N, H, W, groups, _d_per_group -> N, H, W, channels
            center_feature_scale = center_feature_scale[..., None].repeat(
                1, 1, 1, 1, self.channels // self.group).flatten(-2)
            x = x * (1 - center_feature_scale) + x_proj * center_feature_scale
        x = self.output_proj(x)

        return x