File size: 13,965 Bytes
f549064
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) OpenMMLab. All rights reserved.
from typing import Union

import torch
from mmcv.cnn import build_norm_layer
from mmcv.cnn.bricks.transformer import FFN, MultiheadAttention
from mmengine import ConfigDict
from mmengine.model import BaseModule, ModuleList
from torch import Tensor

from mmdet.utils import ConfigType, OptConfigType


class DetrTransformerEncoder(BaseModule):
    """Encoder of DETR.

    Args:
        num_layers (int): Number of encoder layers.
        layer_cfg (:obj:`ConfigDict` or dict): the config of each encoder
            layer. All the layers will share the same config.
        init_cfg (:obj:`ConfigDict` or dict, optional): the config to control
            the initialization. Defaults to None.
    """

    def __init__(self,
                 num_layers: int,
                 layer_cfg: ConfigType,
                 init_cfg: OptConfigType = None) -> None:

        super().__init__(init_cfg=init_cfg)
        self.num_layers = num_layers
        self.layer_cfg = layer_cfg
        self._init_layers()

    def _init_layers(self) -> None:
        """Initialize encoder layers."""
        self.layers = ModuleList([
            DetrTransformerEncoderLayer(**self.layer_cfg)
            for _ in range(self.num_layers)
        ])
        self.embed_dims = self.layers[0].embed_dims

    def forward(self, query: Tensor, query_pos: Tensor,
                key_padding_mask: Tensor, **kwargs) -> Tensor:
        """Forward function of encoder.

        Args:
            query (Tensor): Input queries of encoder, has shape
                (bs, num_queries, dim).
            query_pos (Tensor): The positional embeddings of the queries, has
                shape (bs, num_queries, dim).
            key_padding_mask (Tensor): The `key_padding_mask` of `self_attn`
                input. ByteTensor, has shape (bs, num_queries).

        Returns:
            Tensor: Has shape (bs, num_queries, dim) if `batch_first` is
            `True`, otherwise (num_queries, bs, dim).
        """
        for layer in self.layers:
            query = layer(query, query_pos, key_padding_mask, **kwargs)
        return query


class DetrTransformerDecoder(BaseModule):
    """Decoder of DETR.

    Args:
        num_layers (int): Number of decoder layers.
        layer_cfg (:obj:`ConfigDict` or dict): the config of each encoder
            layer. All the layers will share the same config.
        post_norm_cfg (:obj:`ConfigDict` or dict, optional): Config of the
            post normalization layer. Defaults to `LN`.
        return_intermediate (bool, optional): Whether to return outputs of
            intermediate layers. Defaults to `True`,
        init_cfg (:obj:`ConfigDict` or dict, optional): the config to control
            the initialization. Defaults to None.
    """

    def __init__(self,
                 num_layers: int,
                 layer_cfg: ConfigType,
                 post_norm_cfg: OptConfigType = dict(type='LN'),
                 return_intermediate: bool = True,
                 init_cfg: Union[dict, ConfigDict] = None) -> None:
        super().__init__(init_cfg=init_cfg)
        self.layer_cfg = layer_cfg
        self.num_layers = num_layers
        self.post_norm_cfg = post_norm_cfg
        self.return_intermediate = return_intermediate
        self._init_layers()

    def _init_layers(self) -> None:
        """Initialize decoder layers."""
        self.layers = ModuleList([
            DetrTransformerDecoderLayer(**self.layer_cfg)
            for _ in range(self.num_layers)
        ])
        self.embed_dims = self.layers[0].embed_dims
        self.post_norm = build_norm_layer(self.post_norm_cfg,
                                          self.embed_dims)[1]

    def forward(self, query: Tensor, key: Tensor, value: Tensor,
                query_pos: Tensor, key_pos: Tensor, key_padding_mask: Tensor,
                **kwargs) -> Tensor:
        """Forward function of decoder
        Args:
            query (Tensor): The input query, has shape (bs, num_queries, dim).
            key (Tensor): The input key, has shape (bs, num_keys, dim).
            value (Tensor): The input value with the same shape as `key`.
            query_pos (Tensor): The positional encoding for `query`, with the
                same shape as `query`.
            key_pos (Tensor): The positional encoding for `key`, with the
                same shape as `key`.
            key_padding_mask (Tensor): The `key_padding_mask` of `cross_attn`
                input. ByteTensor, has shape (bs, num_value).

        Returns:
            Tensor: The forwarded results will have shape
            (num_decoder_layers, bs, num_queries, dim) if
            `return_intermediate` is `True` else (1, bs, num_queries, dim).
        """
        intermediate = []
        for layer in self.layers:
            query = layer(
                query,
                key=key,
                value=value,
                query_pos=query_pos,
                key_pos=key_pos,
                key_padding_mask=key_padding_mask,
                **kwargs)
            if self.return_intermediate:
                intermediate.append(self.post_norm(query))
        query = self.post_norm(query)

        if self.return_intermediate:
            return torch.stack(intermediate)

        return query.unsqueeze(0)


class DetrTransformerEncoderLayer(BaseModule):
    """Implements encoder layer in DETR transformer.

    Args:
        self_attn_cfg (:obj:`ConfigDict` or dict, optional): Config for self
            attention.
        ffn_cfg (:obj:`ConfigDict` or dict, optional): Config for FFN.
        norm_cfg (:obj:`ConfigDict` or dict, optional): Config for
            normalization layers. All the layers will share the same
            config. Defaults to `LN`.
        init_cfg (:obj:`ConfigDict` or dict, optional): Config to control
            the initialization. Defaults to None.
    """

    def __init__(self,
                 self_attn_cfg: OptConfigType = dict(
                     embed_dims=256, num_heads=8, dropout=0.0),
                 ffn_cfg: OptConfigType = dict(
                     embed_dims=256,
                     feedforward_channels=1024,
                     num_fcs=2,
                     ffn_drop=0.,
                     act_cfg=dict(type='ReLU', inplace=True)),
                 norm_cfg: OptConfigType = dict(type='LN'),
                 init_cfg: OptConfigType = None) -> None:

        super().__init__(init_cfg=init_cfg)

        self.self_attn_cfg = self_attn_cfg
        if 'batch_first' not in self.self_attn_cfg:
            self.self_attn_cfg['batch_first'] = True
        else:
            assert self.self_attn_cfg['batch_first'] is True, 'First \
            dimension of all DETRs in mmdet is `batch`, \
            please set `batch_first` flag.'

        self.ffn_cfg = ffn_cfg
        self.norm_cfg = norm_cfg
        self._init_layers()

    def _init_layers(self) -> None:
        """Initialize self-attention, FFN, and normalization."""
        self.self_attn = MultiheadAttention(**self.self_attn_cfg)
        self.embed_dims = self.self_attn.embed_dims
        self.ffn = FFN(**self.ffn_cfg)
        norms_list = [
            build_norm_layer(self.norm_cfg, self.embed_dims)[1]
            for _ in range(2)
        ]
        self.norms = ModuleList(norms_list)

    def forward(self, query: Tensor, query_pos: Tensor,
                key_padding_mask: Tensor, **kwargs) -> Tensor:
        """Forward function of an encoder layer.

        Args:
            query (Tensor): The input query, has shape (bs, num_queries, dim).
            query_pos (Tensor): The positional encoding for query, with
                the same shape as `query`.
            key_padding_mask (Tensor): The `key_padding_mask` of `self_attn`
                input. ByteTensor. has shape (bs, num_queries).
        Returns:
            Tensor: forwarded results, has shape (bs, num_queries, dim).
        """
        query = self.self_attn(
            query=query,
            key=query,
            value=query,
            query_pos=query_pos,
            key_pos=query_pos,
            key_padding_mask=key_padding_mask,
            **kwargs)
        query = self.norms[0](query)
        query = self.ffn(query)
        query = self.norms[1](query)

        return query


class DetrTransformerDecoderLayer(BaseModule):
    """Implements decoder layer in DETR transformer.

    Args:
        self_attn_cfg (:obj:`ConfigDict` or dict, optional): Config for self
            attention.
        cross_attn_cfg (:obj:`ConfigDict` or dict, optional): Config for cross
            attention.
        ffn_cfg (:obj:`ConfigDict` or dict, optional): Config for FFN.
        norm_cfg (:obj:`ConfigDict` or dict, optional): Config for
            normalization layers. All the layers will share the same
            config. Defaults to `LN`.
        init_cfg (:obj:`ConfigDict` or dict, optional): Config to control
            the initialization. Defaults to None.
    """

    def __init__(self,
                 self_attn_cfg: OptConfigType = dict(
                     embed_dims=256,
                     num_heads=8,
                     dropout=0.0,
                     batch_first=True),
                 cross_attn_cfg: OptConfigType = dict(
                     embed_dims=256,
                     num_heads=8,
                     dropout=0.0,
                     batch_first=True),
                 ffn_cfg: OptConfigType = dict(
                     embed_dims=256,
                     feedforward_channels=1024,
                     num_fcs=2,
                     ffn_drop=0.,
                     act_cfg=dict(type='ReLU', inplace=True),
                 ),
                 norm_cfg: OptConfigType = dict(type='LN'),
                 init_cfg: OptConfigType = None) -> None:

        super().__init__(init_cfg=init_cfg)

        self.self_attn_cfg = self_attn_cfg
        self.cross_attn_cfg = cross_attn_cfg
        if 'batch_first' not in self.self_attn_cfg:
            self.self_attn_cfg['batch_first'] = True
        else:
            assert self.self_attn_cfg['batch_first'] is True, 'First \
            dimension of all DETRs in mmdet is `batch`, \
            please set `batch_first` flag.'

        if 'batch_first' not in self.cross_attn_cfg:
            self.cross_attn_cfg['batch_first'] = True
        else:
            assert self.cross_attn_cfg['batch_first'] is True, 'First \
            dimension of all DETRs in mmdet is `batch`, \
            please set `batch_first` flag.'

        self.ffn_cfg = ffn_cfg
        self.norm_cfg = norm_cfg
        self._init_layers()

    def _init_layers(self) -> None:
        """Initialize self-attention, FFN, and normalization."""
        self.self_attn = MultiheadAttention(**self.self_attn_cfg)
        self.cross_attn = MultiheadAttention(**self.cross_attn_cfg)
        self.embed_dims = self.self_attn.embed_dims
        self.ffn = FFN(**self.ffn_cfg)
        norms_list = [
            build_norm_layer(self.norm_cfg, self.embed_dims)[1]
            for _ in range(3)
        ]
        self.norms = ModuleList(norms_list)

    def forward(self,
                query: Tensor,
                key: Tensor = None,
                value: Tensor = None,
                query_pos: Tensor = None,
                key_pos: Tensor = None,
                self_attn_mask: Tensor = None,
                cross_attn_mask: Tensor = None,
                key_padding_mask: Tensor = None,
                **kwargs) -> Tensor:
        """
        Args:
            query (Tensor): The input query, has shape (bs, num_queries, dim).
            key (Tensor, optional): The input key, has shape (bs, num_keys,
                dim). If `None`, the `query` will be used. Defaults to `None`.
            value (Tensor, optional): The input value, has the same shape as
                `key`, as in `nn.MultiheadAttention.forward`. If `None`, the
                `key` will be used. Defaults to `None`.
            query_pos (Tensor, optional): The positional encoding for `query`,
                has the same shape as `query`. If not `None`, it will be added
                to `query` before forward function. Defaults to `None`.
            key_pos (Tensor, optional): The positional encoding for `key`, has
                the same shape as `key`. 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.
            self_attn_mask (Tensor, optional): ByteTensor mask, has shape
                (num_queries, num_keys), as in `nn.MultiheadAttention.forward`.
                Defaults to None.
            cross_attn_mask (Tensor, optional): ByteTensor mask, has shape
                (num_queries, num_keys), as in `nn.MultiheadAttention.forward`.
                Defaults to None.
            key_padding_mask (Tensor, optional): The `key_padding_mask` of
                `self_attn` input. ByteTensor, has shape (bs, num_value).
                Defaults to None.

        Returns:
            Tensor: forwarded results, has shape (bs, num_queries, dim).
        """

        query = self.self_attn(
            query=query,
            key=query,
            value=query,
            query_pos=query_pos,
            key_pos=query_pos,
            attn_mask=self_attn_mask,
            **kwargs)
        query = self.norms[0](query)
        query = self.cross_attn(
            query=query,
            key=key,
            value=value,
            query_pos=query_pos,
            key_pos=key_pos,
            attn_mask=cross_attn_mask,
            key_padding_mask=key_padding_mask,
            **kwargs)
        query = self.norms[1](query)
        query = self.ffn(query)
        query = self.norms[2](query)

        return query