一文看懂DETR（二）

在这里插入图片描述
训练流程
1.输入图像经过CNN的backbone获得32倍下采样的深度特征；
2.将图片给拉直形成token，并添加位置编码送入encoder中；
3.将encoder的输出以及Object Query作为decoder的输入得到解码特征；
4.将解码后的特征传入FFN得到预测特征；
5.根据预测特征计算cost matrix，并由匈牙利算法匹配GT，获得正负样本；
6.根据正负样本计算分类与回归loss。

代码实现

书接上回，我们从代码层面讲解了训练步骤1，下面将详细解读DETR如何在encoder中处理token与位置编码的。def forward(self, x, mask, query_embed, pos_embed):可以看到输入transformer forwad函数中的x表示步骤1的32倍下采样特征（[2,256,22,38]维度），query_embed是用于decoder的可学习编码(即论文中的object query维度为[100,256]),pos_embed表示位置编码维度与x一致，mask表示图片有效区域（维度为[2,22,38]）。

首先，需要将x,pos_embed,mask的h，w二维结构拉直成h*w的一维结构，即将x特征拉直成tokens。然后，x(维度[836，2，256])作为query，query_pos=pos_embed(维度[836，2，256]),query_key_padding_mask=mask(维度[2，836])，将三者送入self.encoder中。

class Transformer(BaseModule):
    """Implements the DETR transformer.

    Following the official DETR implementation, this module copy-paste
    from torch.nn.Transformer with modifications:

        * positional encodings are passed in MultiheadAttention
        * extra LN at the end of encoder is removed
        * decoder returns a stack of activations from all decoding layers

    See `paper: End-to-End Object Detection with Transformers
    <https://arxiv.org/pdf/2005.12872>`_ for details.

    Args:
        encoder (`mmcv.ConfigDict` | Dict): Config of
            TransformerEncoder. Defaults to None.
        decoder ((`mmcv.ConfigDict` | Dict)): Config of
            TransformerDecoder. Defaults to None
        init_cfg (obj:`mmcv.ConfigDict`): The Config for initialization.
            Defaults to None.
    """

    def __init__(self, encoder=None, decoder=None, init_cfg=None):
        super(Transformer, self).__init__(init_cfg=init_cfg)
        self.encoder = build_transformer_layer_sequence(encoder)
        self.decoder = build_transformer_layer_sequence(decoder)
        self.embed_dims = self.encoder.embed_dims

    def init_weights(self):
        # follow the official DETR to init parameters
        for m in self.modules():
            if hasattr(m, 'weight') and m.weight.dim() > 1:
                xavier_init(m, distribution='uniform')
        self._is_init = True

    def forward(self, x, mask, query_embed, pos_embed):
        """Forward function for `Transformer`.

        Args:
            x (Tensor): Input query with shape [bs, c, h, w] where
                c = embed_dims.
            mask (Tensor): The key_padding_mask used for encoder and decoder,
                with shape [bs, h, w].
            query_embed (Tensor): The query embedding for decoder, with shape
                [num_query, c].
            pos_embed (Tensor): The positional encoding for encoder and
                decoder, with the same shape as `x`.

        Returns:
            tuple[Tensor]: results of decoder containing the following tensor.

                - out_dec: Output from decoder. If return_intermediate_dec \
                      is True output has shape [num_dec_layers, bs,
                      num_query, embed_dims], else has shape [1, bs, \
                      num_query, embed_dims].
                - memory: Output results from encoder, with shape \
                      [bs, embed_dims, h, w].
        """
        bs, c, h, w = x.shape
        # use `view` instead of `flatten` for dynamically exporting to ONNX
        x = x.view(bs, c, -1).permute(2, 0, 1)  # [bs, c, h, w] -> [h*w, bs, c]
        pos_embed = pos_embed.view(bs, c, -1).permute(2, 0, 1)
        query_embed = query_embed.unsqueeze(1).repeat(
            1, bs, 1)  # [num_query, dim] -> [num_query, bs, dim]
        mask = mask.view(bs, -1)  # [bs, h, w] -> [bs, h*w]
        memory = self.encoder(
            query=x,
            key=None,
            value=None,
            query_pos=pos_embed,
            query_key_padding_mask=mask)
        target = torch.zeros_like(query_embed)
        # out_dec: [num_layers, num_query, bs, dim]
        out_dec = self.decoder(
            query=target,
            key=memory,
            value=memory,
            key_pos=pos_embed,
            query_pos=query_embed,
            key_padding_mask=mask)
        out_dec = out_dec.transpose(1, 2)
        memory = memory.permute(1, 2, 0).reshape(bs, c, h, w)
        return out_dec, memory

self.encoder在transformer初始化时已经注册完毕，是来自class DetrTransformerEncoder(TransformerLayerSequence):的对象。从代码forward中可以看到，它没有单独实现这部分代码，而是直接使用父类DetrTransformerEncoder的forward。

class DetrTransformerEncoder(TransformerLayerSequence):
    """TransformerEncoder of DETR.

    Args:
        post_norm_cfg (dict): Config of last normalization layer. Default：
            `LN`. Only used when `self.pre_norm` is `True`
    """

    def __init__(self, *args, post_norm_cfg=dict(type='LN'), **kwargs):
        super(DetrTransformerEncoder, self).__init__(*args, **kwargs)
        if post_norm_cfg is not None:
            self.post_norm = build_norm_layer(
                post_norm_cfg, self.embed_dims)[1] if self.pre_norm else None
        else:
            assert not self.pre_norm, f'Use prenorm in ' \
                                      f'{self.__class__.__name__},' \
                                      f'Please specify post_norm_cfg'
            self.post_norm = None

    def forward(self, *args, **kwargs):
        """Forward function for `TransformerCoder`.

        Returns:
            Tensor: forwarded results with shape [num_query, bs, embed_dims].
        """
        x = super(DetrTransformerEncoder, self).forward(*args, **kwargs)
        if self.post_norm is not None:
            x = self.post_norm(x)
        return x

在对TransformerLayerSequence初始化时，通过num_layer（config中设置为6）设置encoder的层数，即self.layers的层数。transformerlayers是从config传入的dict，如下所示，构成self.layers的是BaseTransformerLayer这个类，其中self.pre_norm=False， self.embed_dims=256, 这里因为嵌套的类比较多，看起来有些复杂，我们慢慢剖析。

{‘type’: ‘BaseTransformerLayer’, ‘attn_cfgs’: [{‘type’: ‘MultiheadAttention’, ‘embed_dims’: 256, ‘num_heads’: 8, ‘dropout’: 0.1}], ‘feedforward_channels’: 2048, ‘ffn_dropout’: 0.1, ‘operation_order’: (‘self_attn’, ‘norm’, ‘ffn’, ‘norm’)}

class TransformerLayerSequence(BaseModule):
    """Base class for TransformerEncoder and TransformerDecoder in vision
    transformer.

    As base-class of Encoder and Decoder in vision transformer.
    Support customization such as specifying different kind
    of `transformer_layer` in `transformer_coder`.

    Args:
        transformerlayer (list[obj:`mmcv.ConfigDict`] |
            obj:`mmcv.ConfigDict`): Config of transformerlayer
            in TransformerCoder. If it is obj:`mmcv.ConfigDict`,
             it would be repeated `num_layer` times to a
             list[`mmcv.ConfigDict`]. Default: None.
        num_layers (int): The number of `TransformerLayer`. Default: None.
        init_cfg (obj:`mmcv.ConfigDict`): The Config for initialization.
            Default: None.
    """

    def __init__(self, transformerlayers=None, num_layers=None, init_cfg=None):
        super(TransformerLayerSequence, self).__init__(init_cfg)
        if isinstance(transformerlayers, dict):
            transformerlayers = [
                copy.deepcopy(transformerlayers) for _ in range(num_layers)
            ]
        else:
            assert isinstance(transformerlayers, list) and \
                   len(transformerlayers) == num_layers
        self.num_layers = num_layers
        self.layers = ModuleList()
        for i in range(num_layers):
            self.layers.append(build_transformer_layer(transformerlayers[i]))
        self.embed_dims = self.layers[0].embed_dims
        self.pre_norm = self.layers[0].pre_norm

    def forward(self,
                query,
                key,
                value,
                query_pos=None,
                key_pos=None,
                attn_masks=None,
                query_key_padding_mask=None,
                key_padding_mask=None,
                **kwargs):
        """Forward function for `TransformerCoder`.

        Args:
            query (Tensor): Input query with shape
                `(num_queries, bs, embed_dims)`.
            key (Tensor): The key tensor with shape
                `(num_keys, bs, embed_dims)`.
            value (Tensor): The value tensor with shape
                `(num_keys, bs, embed_dims)`.
            query_pos (Tensor): The positional encoding for `query`.
                Default: None.
            key_pos (Tensor): The positional encoding for `key`.
                Default: None.
            attn_masks (List[Tensor], optional): Each element is 2D Tensor
                which is used in calculation of corresponding attention in
                operation_order. Default: None.
            query_key_padding_mask (Tensor): ByteTensor for `query`, with
                shape [bs, num_queries]. Only used in self-attention
                Default: None.
            key_padding_mask (Tensor): ByteTensor for `query`, with
                shape [bs, num_keys]. Default: None.

        Returns:
            Tensor:  results with shape [num_queries, bs, embed_dims].
        """
        for layer in self.layers:
            query = layer(
                query,
                key,
                value,
                query_pos=query_pos,
                key_pos=key_pos,
                attn_masks=attn_masks,
                query_key_padding_mask=query_key_padding_mask,
                key_padding_mask=key_padding_mask,
                **kwargs)
        return query

在TransformerLayerSequence的forward中，我们看到在init初始化好的self.layers，它是个Modulelist，里面重复了6次layer，layer是BaseTransformerLayer这个类的对象。

在BaseTransformerLayer初始化里，self.batch_first=False，operation_order里面按序存放了算子名称，如下所示，self_attn指MultiheadAttention，norm表示layer norm，ffn就是FFN。num_attn表示operation_order的self_attn自注意力和cross_attn交叉注意力出现的次数。在encoder中，operation_order存放了一个self_attn，self_attn与cross_attn具体区别会在下面讲到。

(‘self_attn’, ‘norm’, ‘ffn’, ‘norm’)

attn_cfgs存放了如下参数，添加batch_first属性为False，并使用attn_cfgs初始化attention，attention是来自MultiheadAttention的对象，其中token的维度为256，多头数为8，dropout系数是0.1。

[{‘type’: ‘MultiheadAttention’, ‘embed_dims’: 256, ‘num_heads’: 8, ‘dropout’: 0.1}]

接下来num_ffns表示operation_order中FFN的个数，同样构建ffn_cfgs存放FFN需要的参数，其中embed_dims与attention保持一致即256，self.norms用来指向layer norm。

class BaseTransformerLayer(BaseModule):
    """Base `TransformerLayer` for vision transformer.

    It can be built from `mmcv.ConfigDict` and support more flexible
    customization, for example, using any number of `FFN or LN ` and
    use different kinds of `attention` by specifying a list of `ConfigDict`
    named `attn_cfgs`. It is worth mentioning that it supports `prenorm`
    when you specifying `norm` as the first element of `operation_order`.
    More details about the `prenorm`: `On Layer Normalization in the
    Transformer Architecture <https://arxiv.org/abs/2002.04745>`_ .

    Args:
        attn_cfgs (list[`mmcv.ConfigDict`] | obj:`mmcv.ConfigDict` | None )):
            Configs for `self_attention` or `cross_attention` modules,
            The order of the configs in the list should be consistent with
            corresponding attentions in operation_order.
            If it is a dict, all of the attention modules in operation_order
            will be built with this config. Default: None.
        ffn_cfgs (list[`mmcv.ConfigDict`] | obj:`mmcv.ConfigDict` | None )):
            Configs for FFN, The order of the configs in the list should be
            consistent with corresponding ffn in operation_order.
            If it is a dict, all of the attention modules in operation_order
            will be built with this config.
        operation_order (tuple[str]): The execution order of operation
            in transformer. Such as ('self_attn', 'norm', 'ffn', 'norm').
            Support `prenorm` when you specifying first element as `norm`.
            Default：None.
        norm_cfg (dict): Config dict for normalization layer.
            Default: dict(type='LN').
        init_cfg (obj:`mmcv.ConfigDict`): The Config for initialization.
            Default: None.
        batch_first (bool): Key, Query and Value are shape
            of (batch, n, embed_dim)
            or (n, batch, embed_dim). Default to False.
    """

    def __init__(self,
                 attn_cfgs=None,
                 ffn_cfgs=dict(
                     type='FFN',
                     embed_dims=256,
                     feedforward_channels=1024,
                     num_fcs=2,
                     ffn_drop=0.,
                     act_cfg=dict(type='ReLU', inplace=True),
                 ),
                 operation_order=None,
                 norm_cfg=dict(type='LN'),
                 init_cfg=None,
                 batch_first=False,
                 **kwargs):
        super(BaseTransformerLayer, self).__init__(init_cfg)

        self.batch_first = batch_first
        num_attn = operation_order.count('self_attn') + operation_order.count(
            'cross_attn')
        if isinstance(attn_cfgs, dict):
            attn_cfgs = [copy.deepcopy(attn_cfgs) for _ in range(num_attn)]
        else:
            assert num_attn == len(attn_cfgs), f'The length ' \
                f'of attn_cfg {num_attn} is ' \
                f'not consistent with the number of attention' \
                f'in operation_order {operation_order}.'

        self.num_attn = num_attn
        self.operation_order = operation_order
        self.norm_cfg = norm_cfg
        self.pre_norm = operation_order[0] == 'norm'
        self.attentions = ModuleList()

        index = 0
        for operation_name in operation_order:
            if operation_name in ['self_attn', 'cross_attn']:
                if 'batch_first' in attn_cfgs[index]:
                    assert self.batch_first == attn_cfgs[index]['batch_first']
                else:
                    attn_cfgs[index]['batch_first'] = self.batch_first
                attention = build_attention(attn_cfgs[index])
                # Some custom attentions used as `self_attn`
                # or `cross_attn` can have different behavior.
                attention.operation_name = operation_name
                self.attentions.append(attention)
                index += 1

        self.embed_dims = self.attentions[0].embed_dims

        self.ffns = ModuleList()
        num_ffns = operation_order.count('ffn')
        if isinstance(ffn_cfgs, dict):
            ffn_cfgs = ConfigDict(ffn_cfgs)
        if isinstance(ffn_cfgs, dict):
            ffn_cfgs = [copy.deepcopy(ffn_cfgs) for _ in range(num_ffns)]
        assert len(ffn_cfgs) == num_ffns
        for ffn_index in range(num_ffns):
            if 'embed_dims' not in ffn_cfgs[ffn_index]:
                ffn_cfgs[ffn_index]['embed_dims'] = self.embed_dims
            else:
                assert ffn_cfgs[ffn_index]['embed_dims'] == self.embed_dims
            self.ffns.append(
                build_feedforward_network(ffn_cfgs[ffn_index],
                                          dict(type='FFN')))

        self.norms = ModuleList()
        num_norms = operation_order.count('norm')
        for _ in range(num_norms):
            self.norms.append(build_norm_layer(norm_cfg, self.embed_dims)[1])

BaseTransformerLayer初始化介绍完之后，程序继续来到其forward中，在上面分析self.encoder的forward时，我们已经知道传到这里的参数：x作为query，query_pos=pos_embed,query_key_padding_mask=mask。我们在self.operation_order的循环里首先进入self_attn这个if分支，即进入MultiheadAttention的对象中，这里要注意 temp_key = temp_value = query，即qkv都是x。

 def forward(self,
                query,
                key=None,
                value=None,
                query_pos=None,
                key_pos=None,
                attn_masks=None,
                query_key_padding_mask=None,
                key_padding_mask=None,
                **kwargs):
        """Forward function for `TransformerDecoderLayer`.

        **kwargs contains some specific arguments of attentions.

        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] .
            value (Tensor): The value tensor with same shape as `key`.
            query_pos (Tensor): The positional encoding for `query`.
                Default: None.
            key_pos (Tensor): The positional encoding for `key`.
                Default: None.
            attn_masks (List[Tensor] | None): 2D Tensor used in
                calculation of corresponding attention. The length of
                it should equal to the number of `attention` in
                `operation_order`. Default: None.
            query_key_padding_mask (Tensor): ByteTensor for `query`, with
                shape [bs, num_queries]. Only used in `self_attn` layer.
                Defaults to None.
            key_padding_mask (Tensor): ByteTensor for `query`, with
                shape [bs, num_keys]. Default: None.

        Returns:
            Tensor: forwarded results with shape [num_queries, bs, embed_dims].
        """

        norm_index = 0
        attn_index = 0
        ffn_index = 0
        identity = query
        if attn_masks is None:
            attn_masks = [None for _ in range(self.num_attn)]
        elif isinstance(attn_masks, torch.Tensor):
            attn_masks = [
                copy.deepcopy(attn_masks) for _ in range(self.num_attn)
            ]
            warnings.warn(f'Use same attn_mask in all attentions in '
                          f'{self.__class__.__name__} ')
        else:
            assert len(attn_masks) == self.num_attn, f'The length of ' \
                        f'attn_masks {len(attn_masks)} must be equal ' \
                        f'to the number of attention in ' \
                        f'operation_order {self.num_attn}'

        for layer in self.operation_order:
            if layer == 'self_attn':
                temp_key = temp_value = query
                query = self.attentions[attn_index](
                    query,
                    temp_key,
                    temp_value,
                    identity if self.pre_norm else None,
                    query_pos=query_pos,
                    key_pos=query_pos,
                    attn_mask=attn_masks[attn_index],
                    key_padding_mask=query_key_padding_mask,
                    **kwargs)
                attn_index += 1
                identity = query

            elif layer == 'norm':
                query = self.norms[norm_index](query)
                norm_index += 1

            elif layer == 'cross_attn':
                query = self.attentions[attn_index](
                    query,
                    key,
                    value,
                    identity if self.pre_norm else None,
                    query_pos=query_pos,
                    key_pos=key_pos,
                    attn_mask=attn_masks[attn_index],
                    key_padding_mask=key_padding_mask,
                    **kwargs)
                attn_index += 1
                identity = query

            elif layer == 'ffn':
                query = self.ffns[ffn_index](
                    query, identity if self.pre_norm else None)
                ffn_index += 1

        return query

看到MultiheadAttention的初始化，上面已经讲过这些参数，这里就没啥好介绍的，略过了。。。

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(MultiheadAttention, self).__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()

在看到MultiheadAttention的forward，qkv都是x(维度[836，2，256])，query_pos=key_pos=pos_embed,在q，k上都加上位置编码，将图片位置信息与token结合。self.attn就是多头自注意力，细节在ViT中讲过，这里就不赘述了，最后，return的时候加上了残差以及droupout。

这里进行的是self_attn自注意力，即qkv=x，同时q，k需要加上位置编码。

从MultiheadAttention的forward出来后，又重新进入BaseTransformerLayer的forward，继续进入FFN以及layer norm。这个过程会重复6次。至此，self.encoder的工作就结束了。

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
                else:
                    warnings.warn(f'position encoding of key is'
                                  f'missing in {self.__class__.__name__}.')
        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))

程序返回到transformer的forward中，由self.encoder输出的memory维度为[836,2,256],它重新作为self.decoder的k与v。在self.decoder中，它的q=target，target是维度为[836,2,256]且值为0的tensor，key_pos=pos_embed，query_pos=query_embed, key_padding_mask=mask，这些参数传入到BaseTransformerLayer的forward中。

		memory = self.encoder(
            query=x,
            key=None,
            value=None,
            query_pos=pos_embed,
            query_key_padding_mask=mask)
        target = torch.zeros_like(query_embed)
        # out_dec: [num_layers, num_query, bs, dim]
        out_dec = self.decoder(
            query=target,
            key=memory,
            value=memory,
            key_pos=pos_embed,
            query_pos=query_embed,
            key_padding_mask=mask)
        out_dec = out_dec.transpose(1, 2)
        memory = memory.permute(1, 2, 0).reshape(bs, c, h, w)
        return out_dec, memory

self.decoder中的self.operation_order如下所示，包含了self_attn自注意力与cross_attn交叉注意力模块，num_layers=6。

(‘self_attn’, ‘norm’, ‘cross_attn’, ‘norm’, ‘ffn’, ‘norm’)

self_attn自注意力模块上面介绍过，需要注意的是，self.decoder第一次self_attn时，q=target（target维度为[836，2，256]，值为0），temp_key = temp_value = query，在进入self.attentions后，qkv=0。

进入cross_attn分支，query来自self_attn，key=value=memory，即k，v是来自self.encoder生成的特征（维度为[836，2，256]）。cross_attn交叉注意力与self_attn自注意力的区别在于，交叉注意力的query不等于key，value，qkv来源不同。 需要注意的是，这里的query_pos是object query（可学习编码[100,256]），在attention中，query需要与object query相加生成新的query。可以这么理解，在多头注意力中，加入了object query后，就如同添加100个可学习的anchor，这些anchor会与k计算相似度，并通过softmax归一化后作为value的系数，形成新的特征。通过学习，object query可以分别总结归纳出他们各自感兴趣的位置信息，从而帮助DETR完成目标检测的任务。

if query_pos is not None:
query = query + query_pos
if key_pos is not None:
key = key + key_pos

			if layer == 'self_attn':
                temp_key = temp_value = query
                query = self.attentions[attn_index](
                    query,
                    temp_key,
                    temp_value,
                    identity if self.pre_norm else None,
                    query_pos=query_pos,
                    key_pos=query_pos,
                    attn_mask=attn_masks[attn_index],
                    key_padding_mask=query_key_padding_mask,
                    **kwargs)
                attn_index += 1
                identity = query

            elif layer == 'norm':
                query = self.norms[norm_index](query)
                norm_index += 1

            elif layer == 'cross_attn':
                query = self.attentions[attn_index](
                    query,
                    key,
                    value,
                    identity if self.pre_norm else None,
                    query_pos=query_pos,
                    key_pos=key_pos,
                    attn_mask=attn_masks[attn_index],
                    key_padding_mask=key_padding_mask,
                    **kwargs)
                attn_index += 1
                identity = query

由于DETR没有先验anchor，收敛难度大。为了更好优化梯度，将6层decoder输出的feature都保存下来，并分别分配给 auxiliary decoding losses。至此，self.decoder讲解完毕。

class DetrTransformerDecoder(TransformerLayerSequence):
    """Implements the decoder in DETR transformer.

    Args:
        return_intermediate (bool): Whether to return intermediate outputs.
        post_norm_cfg (dict): Config of last normalization layer. Default：
            `LN`.
    """

    def __init__(self,
                 *args,
                 post_norm_cfg=dict(type='LN'),
                 return_intermediate=False,
                 **kwargs):

        super(DetrTransformerDecoder, self).__init__(*args, **kwargs)
        self.return_intermediate = return_intermediate
        if post_norm_cfg is not None:
            self.post_norm = build_norm_layer(post_norm_cfg,
                                              self.embed_dims)[1]
        else:
            self.post_norm = None

    def forward(self, query, *args, **kwargs):
        """Forward function for `TransformerDecoder`.

        Args:
            query (Tensor): Input query with shape
                `(num_query, bs, embed_dims)`.

        Returns:
            Tensor: Results with shape [1, num_query, bs, embed_dims] when
                return_intermediate is `False`, otherwise it has shape
                [num_layers, num_query, bs, embed_dims].
        """
        if not self.return_intermediate:
            x = super().forward(query, *args, **kwargs)
            if self.post_norm:
                x = self.post_norm(x)[None]
            return x

        intermediate = []
        for layer in self.layers:
            query = layer(query, *args, **kwargs)
            if self.return_intermediate:
                if self.post_norm is not None:
                    intermediate.append(self.post_norm(query))
                else:
                    intermediate.append(query)
        return torch.stack(intermediate)