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)