File size: 15,577 Bytes
e98bd8c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
import copy
from abc import ABCMeta
from collections import defaultdict
from typing import Iterable, List, Optional, Union, Callable
import warnings
from inspect import getfullargspec
import functools
import torch.nn as nn

from .utils import is_model_wrapper
from .weight_init import PretrainedInit, initialize, update_init_info
from ..utils.activation import build_dropout
from ..utils.registry import MODELS


class BaseModule(nn.Module, metaclass=ABCMeta):
    """Base module for all modules in openmmlab. ``BaseModule`` is a wrapper of
    ``torch.nn.Module`` with additional functionality of parameter
    initialization. Compared with ``torch.nn.Module``, ``BaseModule`` mainly
    adds three attributes.

    - ``init_cfg``: the config to control the initialization.
    - ``init_weights``: The function of parameter initialization and recording
      initialization information.
    - ``_params_init_info``: Used to track the parameter initialization
      information. This attribute only exists during executing the
      ``init_weights``.

    Note:
        :obj:`PretrainedInit` has a higher priority than any other
        initializer. The loaded pretrained weights will overwrite
        the previous initialized weights.

    Args:
        init_cfg (dict or List[dict], optional): Initialization config dict.
    """

    def __init__(self, init_cfg: Union[dict, List[dict], None] = None):
        """Initialize BaseModule, inherited from `torch.nn.Module`"""

        # NOTE init_cfg can be defined in different levels, but init_cfg
        # in low levels has a higher priority.

        super().__init__()
        # define default value of init_cfg instead of hard code
        # in init_weights() function
        self._is_init = False

        self.init_cfg = copy.deepcopy(init_cfg)

        # Backward compatibility in derived classes
        # if pretrained is not None:
        #     warnings.warn('DeprecationWarning: pretrained is a deprecated \
        #         key, please consider using init_cfg')
        #     self.init_cfg = dict(type='Pretrained', checkpoint=pretrained)

    @property
    def is_init(self):
        return self._is_init

    @is_init.setter
    def is_init(self, value):
        self._is_init = value

    def init_weights(self):
        """Initialize the weights."""

        is_top_level_module = False
        # check if it is top-level module
        if not hasattr(self, '_params_init_info'):
            # The `_params_init_info` is used to record the initialization
            # information of the parameters
            # the key should be the obj:`nn.Parameter` of model and the value
            # should be a dict containing
            # - init_info (str): The string that describes the initialization.
            # - tmp_mean_value (FloatTensor): The mean of the parameter,
            #       which indicates whether the parameter has been modified.
            # this attribute would be deleted after all parameters
            # is initialized.
            self._params_init_info = defaultdict(dict)
            is_top_level_module = True

            # Initialize the `_params_init_info`,
            # When detecting the `tmp_mean_value` of
            # the corresponding parameter is changed, update related
            # initialization information
            for name, param in self.named_parameters():
                self._params_init_info[param][
                    'init_info'] = f'The value is the same before and ' \
                                   f'after calling `init_weights` ' \
                                   f'of {self.__class__.__name__} '
                self._params_init_info[param][
                    'tmp_mean_value'] = param.data.mean().cpu()

            # pass `params_init_info` to all submodules
            # All submodules share the same `params_init_info`,
            # so it will be updated when parameters are
            # modified at any level of the model.
            for sub_module in self.modules():
                sub_module._params_init_info = self._params_init_info

        module_name = self.__class__.__name__
        if not self._is_init:
            if self.init_cfg:

                init_cfgs = self.init_cfg
                if isinstance(self.init_cfg, dict):
                    init_cfgs = [self.init_cfg]

                # PretrainedInit has higher priority than any other init_cfg.
                # Therefore we initialize `pretrained_cfg` last to overwrite
                # the previous initialized weights.
                # See details in https://github.com/open-mmlab/mmengine/issues/691 # noqa E501
                other_cfgs = []
                pretrained_cfg = []
                for init_cfg in init_cfgs:
                    assert isinstance(init_cfg, dict)
                    if (init_cfg['type'] == 'Pretrained'
                            or init_cfg['type'] is PretrainedInit):
                        pretrained_cfg.append(init_cfg)
                    else:
                        other_cfgs.append(init_cfg)

                initialize(self, other_cfgs)

            for m in self.children():
                if is_model_wrapper(m) and not hasattr(m, 'init_weights'):
                    m = m.module
                if hasattr(m, 'init_weights') and not getattr(
                        m, 'is_init', False):
                    m.init_weights()
                    # users may overload the `init_weights`
                    update_init_info(
                        m,
                        init_info=f'Initialized by '
                        f'user-defined `init_weights`'
                        f' in {m.__class__.__name__} ')
            if self.init_cfg and pretrained_cfg:
                initialize(self, pretrained_cfg)
            self._is_init = True

        if is_top_level_module:
            self._dump_init_info()

            for sub_module in self.modules():
                del sub_module._params_init_info

    def __repr__(self):
        s = super().__repr__()
        if self.init_cfg:
            s += f'\ninit_cfg={self.init_cfg}'
        return s
    

def deprecated_api_warning(name_dict: dict,
                           cls_name: Optional[str] = None) -> Callable:
    """A decorator to check if some arguments are deprecate and try to replace
    deprecate src_arg_name to dst_arg_name.

    Args:
        name_dict(dict):
            key (str): Deprecate argument names.
            val (str): Expected argument names.

    Returns:
        func: New function.
    """

    def api_warning_wrapper(old_func):

        @functools.wraps(old_func)
        def new_func(*args, **kwargs):
            # get the arg spec of the decorated method
            args_info = getfullargspec(old_func)
            # get name of the function
            func_name = old_func.__name__
            if cls_name is not None:
                func_name = f'{cls_name}.{func_name}'
            if args:
                arg_names = args_info.args[:len(args)]
                for src_arg_name, dst_arg_name in name_dict.items():
                    if src_arg_name in arg_names:
                        warnings.warn(
                            f'"{src_arg_name}" is deprecated in '
                            f'`{func_name}`, please use "{dst_arg_name}" '
                            'instead', DeprecationWarning)
                        arg_names[arg_names.index(src_arg_name)] = dst_arg_name
            if kwargs:
                for src_arg_name, dst_arg_name in name_dict.items():
                    if src_arg_name in kwargs:
                        assert dst_arg_name not in kwargs, (
                            f'The expected behavior is to replace '
                            f'the deprecated key `{src_arg_name}` to '
                            f'new key `{dst_arg_name}`, but got them '
                            f'in the arguments at the same time, which '
                            f'is confusing. `{src_arg_name} will be '
                            f'deprecated in the future, please '
                            f'use `{dst_arg_name}` instead.')

                        warnings.warn(
                            f'"{src_arg_name}" is deprecated in '
                            f'`{func_name}`, please use "{dst_arg_name}" '
                            'instead', DeprecationWarning)
                        kwargs[dst_arg_name] = kwargs.pop(src_arg_name)

            # apply converted arguments to the decorated method
            output = old_func(*args, **kwargs)
            return output

        return new_func

    return api_warning_wrapper
    

@MODELS.register_module()
class MultiheadAttention(BaseModule):
    """A wrapper for ``torch.nn.MultiheadAttention``.

    This module implements MultiheadAttention with identity connection,
    and positional encoding  is also passed as input.

    Args:
        embed_dims (int): The embedding dimension.
        num_heads (int): Parallel attention heads.
        attn_drop (float): A Dropout layer on attn_output_weights.
            Default: 0.0.
        proj_drop (float): A Dropout layer after `nn.MultiheadAttention`.
            Default: 0.0.
        dropout_layer (obj:`ConfigDict`): The dropout_layer used
            when adding the shortcut.
        init_cfg (obj:`mmcv.ConfigDict`): The Config for initialization.
            Default: None.
        batch_first (bool): When it is True,  Key, Query and Value are shape of
            (batch, n, embed_dim), otherwise (n, batch, embed_dim).
             Default to False.
    """

    def __init__(self,
                 embed_dims,
                 num_heads,
                 attn_drop=0.,
                 proj_drop=0.,
                 dropout_layer=dict(type='Dropout', drop_prob=0.),
                 init_cfg=None,
                 batch_first=False,
                 **kwargs):
        super().__init__(init_cfg)
        if 'dropout' in kwargs:
            warnings.warn(
                'The arguments `dropout` in MultiheadAttention '
                'has been deprecated, now you can separately '
                'set `attn_drop`(float), proj_drop(float), '
                'and `dropout_layer`(dict) ', DeprecationWarning)
            attn_drop = kwargs['dropout']
            dropout_layer['drop_prob'] = kwargs.pop('dropout')

        self.embed_dims = embed_dims
        self.num_heads = num_heads
        self.batch_first = batch_first

        self.attn = nn.MultiheadAttention(embed_dims, num_heads, attn_drop,
                                          **kwargs)

        self.proj_drop = nn.Dropout(proj_drop)
        self.dropout_layer = build_dropout(
            dropout_layer) if dropout_layer else nn.Identity()

    @deprecated_api_warning({'residual': 'identity'},
                            cls_name='MultiheadAttention')
    def forward(self,
                query,
                key=None,
                value=None,
                identity=None,
                query_pos=None,
                key_pos=None,
                attn_mask=None,
                key_padding_mask=None,
                **kwargs):
        """Forward function for `MultiheadAttention`.

        **kwargs allow passing a more general data flow when combining
        with other operations in `transformerlayer`.

        Args:
            query (Tensor): The input query with shape [num_queries, bs,
                embed_dims] if self.batch_first is False, else
                [bs, num_queries embed_dims].
            key (Tensor): The key tensor with shape [num_keys, bs,
                embed_dims] if self.batch_first is False, else
                [bs, num_keys, embed_dims] .
                If None, the ``query`` will be used. Defaults to None.
            value (Tensor): The value tensor with same shape as `key`.
                Same in `nn.MultiheadAttention.forward`. Defaults to None.
                If None, the `key` will be used.
            identity (Tensor): This tensor, with the same shape as x,
                will be used for the identity link.
                If None, `x` will be used. Defaults to None.
            query_pos (Tensor): The positional encoding for query, with
                the same shape as `x`. If not None, it will
                be added to `x` before forward function. Defaults to None.
            key_pos (Tensor): The positional encoding for `key`, with the
                same shape as `key`. Defaults to None. 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.
            attn_mask (Tensor): ByteTensor mask with shape [num_queries,
                num_keys]. Same in `nn.MultiheadAttention.forward`.
                Defaults to None.
            key_padding_mask (Tensor): ByteTensor with shape [bs, num_keys].
                Defaults to None.

        Returns:
            Tensor: forwarded results with shape
            [num_queries, bs, embed_dims]
            if self.batch_first is False, else
            [bs, num_queries embed_dims].
        """

        if key is None:
            key = query
        if value is None:
            value = key
        if identity is None:
            identity = query
        if key_pos is None:
            if query_pos is not None:
                # use query_pos if key_pos is not available
                if query_pos.shape == key.shape:
                    key_pos = query_pos
        if query_pos is not None:
            query = query + query_pos
        if key_pos is not None:
            key = key + key_pos

        # Because the dataflow('key', 'query', 'value') of
        # ``torch.nn.MultiheadAttention`` is (num_query, batch,
        # embed_dims), We should adjust the shape of dataflow from
        # batch_first (batch, num_query, embed_dims) to num_query_first
        # (num_query ,batch, embed_dims), and recover ``attn_output``
        # from num_query_first to batch_first.
        if self.batch_first:
            query = query.transpose(0, 1)
            key = key.transpose(0, 1)
            value = value.transpose(0, 1)

        out = self.attn(
            query=query,
            key=key,
            value=value,
            attn_mask=attn_mask,
            key_padding_mask=key_padding_mask)[0]

        if self.batch_first:
            out = out.transpose(0, 1)

        return identity + self.dropout_layer(self.proj_drop(out))


class ModuleList(BaseModule, nn.ModuleList):
    """ModuleList in openmmlab.

    Ensures that all modules in ``ModuleList`` have a different initialization
    strategy than the outer model

    Args:
        modules (iterable, optional): An iterable of modules to add.
        init_cfg (dict, optional): Initialization config dict.
    """

    def __init__(self,
                 modules: Optional[Iterable] = None,
                 init_cfg: Optional[dict] = None):
        BaseModule.__init__(self, init_cfg)
        nn.ModuleList.__init__(self, modules)


class Sequential(BaseModule, nn.Sequential):
    """Sequential module in openmmlab.

    Ensures that all modules in ``Sequential`` have a different initialization
    strategy than the outer model

    Args:
        init_cfg (dict, optional): Initialization config dict.
    """

    def __init__(self, *args, init_cfg: Optional[dict] = None):
        BaseModule.__init__(self, init_cfg)
        nn.Sequential.__init__(self, *args)