引言

Deformable-DETR的主要贡献：
1，结合可变形卷积的稀疏空间采用和Transformer的全局关系建模能力，提出可变形注意力机制模型，使其计算量降低，收敛加快。
2，使用多层级特征，但不使用FPN，对小目标有较好效果。

改进与创新

可变形注意力

可变形注意力提出的初衷是为了解决Transformer的Q，K的运算数据量巨大问题。作者认为Q没必要与所有的K都计算内积，而是只需要选择几个重要的K即可。
如下图，在该论文中，作者设定找4个K即可，而4个K的位置可以不断进行偏移，偏移过程如下图所示：

因此要解决的问题就是：（1）确定reference point(参考点）。（2）确定每个reference point的偏移量(offset)。（3）确定注意力权重矩阵 A_mqk，其中在Encoder和Decoder中实现方法不太一样。Deformable的计算方式如下：

在Encoder部分，输入的Query Feature ( z_q )为加入了位置编码的特征图(src+pos)， value(x)的计算方法只使用了src而没有位置编码(value_proj函数)。

1. reference point确定方法为用了torch.meshgrid方法，调用函数 get_reference_points，有一个细节就是参考点归一化到0和1之间，因此取值的时候要用到双线性插值的方法。而在Decoder中，参考点的获取方法为object queries通过一个nn.Linear得到每个对应的reference point。

def get_reference_points(spatial_shapes, valid_ratios, device):reference_points_list = []for lvl, (H_, W_) in enumerate(spatial_shapes):# 从0.5到H-0.5采样H个点，W同理 这个操作的目的也就是为了特征图的对齐ref_y, ref_x = torch.meshgrid(torch.linspace(0.5, H_ - 0.5, H_, dtype=torch.float32, device=device),torch.linspace(0.5, W_ - 0.5, W_, dtype=torch.float32, device=device))ref_y = ref_y.reshape(-1)[None] / (valid_ratios[:, None, lvl, 1] * H_)ref_x = ref_x.reshape(-1)[None] / (valid_ratios[:, None, lvl, 0] * W_)ref = torch.stack((ref_x, ref_y), -1)reference_points_list.append(ref)reference_points = torch.cat(reference_points_list, 1)reference_points = reference_points[:, :, None] * valid_ratios[:, None]return reference_points

（2）计算offset的方法为对 z_q 过一个nn.Linear，得到多组偏移量，每组偏移量的维度为参考点的个数，组数为注意力头的数量。

（3）计算注意力权重矩阵 A_mqk 的方法为 z_q 过一个nn.Linear和一个F.softmax，得到每个头的注意力权重。
如下图所示：

如上图所示：分头计算完的注意力最终会拼接到一起，然后最后过一个nn.Linear得到输入x 的最终输出。

多层级特征融合（Multi-Scale Deformable Attention Module）

多尺度的Deformable Attention模块也是在多尺度特征图上计算的。多尺度的特征融合方法则是取了骨干网(ResNet)最后三层的特征图C3，C4，C5，并且用了一个Conv3x3 Stride2的卷积得到了一个C6构成了四层特征图。随后会通过卷积操作将通道数量统一为256(也就是token的数量)，然后在这四个特征图上运行Deformable Attention Module并且进行直接相加得到最终输出。其中

Deformable Attention Module算子的pytorch实现如下：
def ms_deform_attn_core_pytorch(value, value_spatial_shapes, sampling_locations, attention_weights):# for debug and test only,# need to use cuda version insteadN_, S_, M_, D_ = value.shape # batch size, number token, number head, head dims# Lq_: number query, L_: level number, P_: sampling number采样点数_, Lq_, M_, L_, P_, _ = sampling_locations.shape# 按照level划分valuevalue_list = value.split([H_ * W_ for H_, W_ in value_spatial_shapes], dim=1)# [0, 1] -> [-1, 1] 因为要满足F.grid_sample的输入要求sampling_grids = 2 * sampling_locations - 1sampling_value_list = []for lid_, (H_, W_) in enumerate(value_spatial_shapes):# N_, H_*W_, M_, D_ -> N_, H_*W_, M_*D_ -> N_, M_*D_, H_*W_ -> N_*M_, D_, H_, W_value_l_ = value_list[lid_].flatten(2).transpose(1, 2).reshape(N_*M_, D_, H_, W_)# N_, Lq_, M_, P_, 2 -> N_, M_, Lq_, P_, 2 -> N_*M_, Lq_, P_, 2sampling_grid_l_ = sampling_grids[:, :, :, lid_].transpose(1, 2).flatten(0, 1)# N_*M_, D_, Lq_, P_# 用双线性插值从feature map上获取value，因为mask的原因越界所以要zeros的方法进行填充sampling_value_l_ = F.grid_sample(value_l_, sampling_grid_l_,mode='bilinear', padding_mode='zeros', align_corners=False)sampling_value_list.append(sampling_value_l_)# (N_, Lq_, M_, L_, P_) -> (N_, M_, Lq_, L_, P_) -> (N_, M_, 1, Lq_, L_*P_)attention_weights = attention_weights.transpose(1, 2).reshape(N_*M_, 1, Lq_, L_*P_)# 不同scale计算出的multi head attention 进行相加，返回output后还需要过一个Linear层output = (torch.stack(sampling_value_list, dim=-2).flatten(-2) * attention_weights).sum(-1).view(N_, M_*D_, Lq_)return output.transpose(1, 2).contiguous()

Two-Stage Deformable DETR

这里的两阶段是受two-stage object detectors的启发，当然这里的改动其实很小：将Encoder输出的memory送入了FFN（前馈神经网络负责类别预测与box预测）将其进行修正后再送入Decoder。

其他改进

其他方面，Deformable相较于DETR修改了query-num的数量，改为300，但在推理过程中其会仍使用top100的预测框，此外在匈牙利匹配的cost矩阵构建时class的损失由原本的softmax简单运算变为了Focus loss。

模型结构

Encoder

Encoder加入了参考点计算，，此外改动了DerormableAttention计算。

class DeformableTransformerEncoderLayer(nn.Module):def __init__(self,d_model=256, d_ffn=1024,dropout=0.1, activation="relu",n_levels=4, n_heads=8, n_points=4):super().__init__()# self attentionself.self_attn = MSDeformAttn(d_model, n_levels, n_heads, n_points)self.dropout1 = nn.Dropout(dropout)self.norm1 = nn.LayerNorm(d_model)# ffnself.linear1 = nn.Linear(d_model, d_ffn)self.activation = _get_activation_fn(activation)self.dropout2 = nn.Dropout(dropout)self.linear2 = nn.Linear(d_ffn, d_model)self.dropout3 = nn.Dropout(dropout)self.norm2 = nn.LayerNorm(d_model)@staticmethoddef with_pos_embed(tensor, pos):return tensor if pos is None else tensor + posdef forward_ffn(self, src):src2 = self.linear2(self.dropout2(self.activation(self.linear1(src))))src = src + self.dropout3(src2)src = self.norm2(src)return srcdef forward(self, src, pos, reference_points, spatial_shapes, level_start_index, padding_mask=None):# self attentionsrc2 = self.self_attn(self.with_pos_embed(src, pos), reference_points, src, spatial_shapes, level_start_index, padding_mask)src = src + self.dropout1(src2)src = self.norm1(src)# ffnsrc = self.forward_ffn(src)return srcclass DeformableTransformerEncoder(nn.Module):def __init__(self, encoder_layer, num_layers):super().__init__()self.layers = _get_clones(encoder_layer, num_layers)self.num_layers = num_layers@staticmethoddef get_reference_points(spatial_shapes, valid_ratios, device):reference_points_list = []for lvl, (H_, W_) in enumerate(spatial_shapes):# 从0.5到H-0.5采样H个点，W同理 这个操作的目的也就是为了特征图的对齐ref_y, ref_x = torch.meshgrid(torch.linspace(0.5, H_ - 0.5, H_, dtype=torch.float32, device=device),torch.linspace(0.5, W_ - 0.5, W_, dtype=torch.float32, device=device))ref_y = ref_y.reshape(-1)[None] / (valid_ratios[:, None, lvl, 1] * H_)ref_x = ref_x.reshape(-1)[None] / (valid_ratios[:, None, lvl, 0] * W_)ref = torch.stack((ref_x, ref_y), -1)reference_points_list.append(ref)reference_points = torch.cat(reference_points_list, 1)reference_points = reference_points[:, :, None] * valid_ratios[:, None]return reference_pointsdef forward(self, src, spatial_shapes, level_start_index, valid_ratios, pos=None, padding_mask=None):output = srcreference_points = self.get_reference_points(spatial_shapes, valid_ratios, device=src.device)for _, layer in enumerate(self.layers):output = layer(output, pos, reference_points, spatial_shapes, level_start_index, padding_mask)return output

Decoder

详细代码注释如下，这里要控制是否使用iterative bounding box refinement和two stage技巧。iterative bounding box refinement其实就是对参考点的位置进行微调。two stage方法其实就是通过参考点直接生成anchor但是只取最高置信度的前几个，然后再送入decoder进行调整。intermediate数组是一个trick，每层Decoder都是可以输出bbox和分类信息的，如果都利用起来算损失则成为auxiliary loss。

class DeformableTransformerDecoderLayer(nn.Module):def __init__(self, d_model=256, d_ffn=1024,dropout=0.1, activation="relu",n_levels=4, n_heads=8, n_points=4):super().__init__()# cross attentionself.cross_attn = MSDeformAttn(d_model, n_levels, n_heads, n_points)self.dropout1 = nn.Dropout(dropout)self.norm1 = nn.LayerNorm(d_model)# self attentionself.self_attn = nn.MultiheadAttention(d_model, n_heads, dropout=dropout)self.dropout2 = nn.Dropout(dropout)self.norm2 = nn.LayerNorm(d_model)# ffnself.linear1 = nn.Linear(d_model, d_ffn)self.activation = _get_activation_fn(activation)self.dropout3 = nn.Dropout(dropout)self.linear2 = nn.Linear(d_ffn, d_model)self.dropout4 = nn.Dropout(dropout)self.norm3 = nn.LayerNorm(d_model)@staticmethoddef with_pos_embed(tensor, pos):return tensor if pos is None else tensor + posdef forward_ffn(self, tgt):tgt2 = self.linear2(self.dropout3(self.activation(self.linear1(tgt))))tgt = tgt + self.dropout4(tgt2)tgt = self.norm3(tgt)return tgtdef forward(self, tgt, query_pos, reference_points, src, src_spatial_shapes, level_start_index, src_padding_mask=None):# self attentionq = k = self.with_pos_embed(tgt, query_pos)tgt2 = self.self_attn(q.transpose(0, 1), k.transpose(0, 1), tgt.transpose(0, 1))[0].transpose(0, 1)tgt = tgt + self.dropout2(tgt2)tgt = self.norm2(tgt)# cross attentiontgt2 = self.cross_attn(self.with_pos_embed(tgt, query_pos),reference_points,src, src_spatial_shapes, level_start_index, src_padding_mask)tgt = tgt + self.dropout1(tgt2)tgt = self.norm1(tgt)# ffntgt = self.forward_ffn(tgt)return tgtclass DeformableTransformerDecoder(nn.Module):def __init__(self, decoder_layer, num_layers, return_intermediate=False):super().__init__()self.layers = _get_clones(decoder_layer, num_layers)self.num_layers = num_layersself.return_intermediate = return_intermediate# hack implementation for iterative bounding box refinement and two-stage Deformable DETRself.bbox_embed = Noneself.class_embed = Nonedef forward(self, tgt, reference_points, src, src_spatial_shapes, src_level_start_index, src_valid_ratios,query_pos=None, src_padding_mask=None):output = tgt# 用来存储中间decoder输出的 可以考虑是否用auxiliary lossintermediate = []intermediate_reference_points = []for lid, layer in enumerate(self.layers):if reference_points.shape[-1] == 4:reference_points_input = reference_points[:, :, None] \* torch.cat([src_valid_ratios, src_valid_ratios], -1)[:, None]else:assert reference_points.shape[-1] == 2reference_points_input = reference_points[:, :, None] * src_valid_ratios[:, None]output = layer(output, query_pos, reference_points_input, src, src_spatial_shapes, src_level_start_index, src_padding_mask)# hack implementation for iterative bounding box refinement# iterative refinement是对decoder中的参考点进行微调，类似cascade rcnn思想if self.bbox_embed is not None:tmp = self.bbox_embed[lid](output)if reference_points.shape[-1] == 4:new_reference_points = tmp + inverse_sigmoid(reference_points)new_reference_points = new_reference_points.sigmoid()else:assert reference_points.shape[-1] == 2new_reference_points = tmpnew_reference_points[..., :2] = tmp[..., :2] + inverse_sigmoid(reference_points)new_reference_points = new_reference_points.sigmoid()reference_points = new_reference_points.detach()if self.return_intermediate:intermediate.append(output)intermediate_reference_points.append(reference_points)if self.return_intermediate:return torch.stack(intermediate), torch.stack(intermediate_reference_points)return output, reference_points

Deformable Transformer

class DeformableTransformer(nn.Module):def __init__(self, d_model=256, nhead=8,num_encoder_layers=6, num_decoder_layers=6, dim_feedforward=1024, dropout=0.1,activation="relu", return_intermediate_dec=False,num_feature_levels=4, dec_n_points=4,  enc_n_points=4,two_stage=False, two_stage_num_proposals=300):super().__init__()self.d_model = d_modelself.nhead = nheadself.two_stage = two_stageself.two_stage_num_proposals = two_stage_num_proposalsencoder_layer = DeformableTransformerEncoderLayer(d_model, dim_feedforward,dropout, activation,num_feature_levels, nhead, enc_n_points)self.encoder = DeformableTransformerEncoder(encoder_layer, num_encoder_layers)decoder_layer = DeformableTransformerDecoderLayer(d_model, dim_feedforward,dropout, activation,num_feature_levels, nhead, dec_n_points)self.decoder = DeformableTransformerDecoder(decoder_layer, num_decoder_layers, return_intermediate_dec)self.level_embed = nn.Parameter(torch.Tensor(num_feature_levels, d_model))if two_stage:self.enc_output = nn.Linear(d_model, d_model)self.enc_output_norm = nn.LayerNorm(d_model)self.pos_trans = nn.Linear(d_model * 2, d_model * 2)self.pos_trans_norm = nn.LayerNorm(d_model * 2)else:self.reference_points = nn.Linear(d_model, 2)self._reset_parameters()def _reset_parameters(self):for p in self.parameters():if p.dim() > 1:nn.init.xavier_uniform_(p)for m in self.modules():if isinstance(m, MSDeformAttn):m._reset_parameters()if not self.two_stage:xavier_uniform_(self.reference_points.weight.data, gain=1.0)constant_(self.reference_points.bias.data, 0.)normal_(self.level_embed)def get_proposal_pos_embed(self, proposals):num_pos_feats = 128temperature = 10000scale = 2 * math.pidim_t = torch.arange(num_pos_feats, dtype=torch.float32, device=proposals.device)dim_t = temperature ** (2 * (dim_t // 2) / num_pos_feats)# N, L, 4proposals = proposals.sigmoid() * scale# N, L, 4, 128pos = proposals[:, :, :, None] / dim_t# N, L, 4, 64, 2pos = torch.stack((pos[:, :, :, 0::2].sin(), pos[:, :, :, 1::2].cos()), dim=4).flatten(2)return posdef gen_encoder_output_proposals(self, memory, memory_padding_mask, spatial_shapes):N_, S_, C_ = memory.shapebase_scale = 4.0proposals = []_cur = 0for lvl, (H_, W_) in enumerate(spatial_shapes):mask_flatten_ = memory_padding_mask[:, _cur:(_cur + H_ * W_)].view(N_, H_, W_, 1)valid_H = torch.sum(~mask_flatten_[:, :, 0, 0], 1)valid_W = torch.sum(~mask_flatten_[:, 0, :, 0], 1)grid_y, grid_x = torch.meshgrid(torch.linspace(0, H_ - 1, H_, dtype=torch.float32, device=memory.device),torch.linspace(0, W_ - 1, W_, dtype=torch.float32, device=memory.device))grid = torch.cat([grid_x.unsqueeze(-1), grid_y.unsqueeze(-1)], -1)scale = torch.cat([valid_W.unsqueeze(-1), valid_H.unsqueeze(-1)], 1).view(N_, 1, 1, 2)grid = (grid.unsqueeze(0).expand(N_, -1, -1, -1) + 0.5) / scalewh = torch.ones_like(grid) * 0.05 * (2.0 ** lvl)proposal = torch.cat((grid, wh), -1).view(N_, -1, 4)proposals.append(proposal)_cur += (H_ * W_)output_proposals = torch.cat(proposals, 1)output_proposals_valid = ((output_proposals > 0.01) & (output_proposals < 0.99)).all(-1, keepdim=True)output_proposals = torch.log(output_proposals / (1 - output_proposals))output_proposals = output_proposals.masked_fill(memory_padding_mask.unsqueeze(-1), float('inf'))output_proposals = output_proposals.masked_fill(~output_proposals_valid, float('inf'))output_memory = memoryoutput_memory = output_memory.masked_fill(memory_padding_mask.unsqueeze(-1), float(0))output_memory = output_memory.masked_fill(~output_proposals_valid, float(0))output_memory = self.enc_output_norm(self.enc_output(output_memory))return output_memory, output_proposalsdef get_valid_ratio(self, mask):_, H, W = mask.shapevalid_H = torch.sum(~mask[:, :, 0], 1)valid_W = torch.sum(~mask[:, 0, :], 1)valid_ratio_h = valid_H.float() / Hvalid_ratio_w = valid_W.float() / Wvalid_ratio = torch.stack([valid_ratio_w, valid_ratio_h], -1)return valid_ratiodef forward(self, srcs, masks, pos_embeds, query_embed=None):assert self.two_stage or query_embed is not None# prepare input for encodersrc_flatten = []mask_flatten = []lvl_pos_embed_flatten = []spatial_shapes = []for lvl, (src, mask, pos_embed) in enumerate(zip(srcs, masks, pos_embeds)):# 得到每一层feature map的batch size 通道数量 高宽bs, c, h, w = src.shapespatial_shape = (h, w)spatial_shapes.append(spatial_shape)# 将每层的feature map、mask、位置编码拉平，并且加入到相关数组中src = src.flatten(2).transpose(1, 2)mask = mask.flatten(1)pos_embed = pos_embed.flatten(2).transpose(1, 2)# 位置编码和可学习的每层编码相加，表征类似 3D positionlvl_pos_embed = pos_embed + self.level_embed[lvl].view(1, 1, -1) lvl_pos_embed_flatten.append(lvl_pos_embed)src_flatten.append(src)mask_flatten.append(mask)# 在hidden_dim维度上进行拼接，也就是number token数量一样的那个维度src_flatten = torch.cat(src_flatten, 1)mask_flatten = torch.cat(mask_flatten, 1)lvl_pos_embed_flatten = torch.cat(lvl_pos_embed_flatten, 1)spatial_shapes = torch.as_tensor(spatial_shapes, dtype=torch.long, device=src_flatten.device)# 记录每个level开始的索引以及有效的长宽(因为有mask存在，raw image的分辨率可能不统一) 具体查看get_valid_ratio函数# prod(1)计算h*w，cumsum(0)计算前缀和level_start_index = torch.cat((spatial_shapes.new_zeros((1, )), spatial_shapes.prod(1).cumsum(0)[:-1]))valid_ratios = torch.stack([self.get_valid_ratio(m) for m in masks], 1)# encoder memory = self.encoder(src_flatten, spatial_shapes, level_start_index, valid_ratios, lvl_pos_embed_flatten, mask_flatten)# prepare input for decoderbs, _, c = memory.shape# 是否使用两阶段模式if self.two_stage:output_memory, output_proposals = self.gen_encoder_output_proposals(memory, mask_flatten, spatial_shapes)# hack implementation for two-stage Deformable DETRenc_outputs_class = self.decoder.class_embed[self.decoder.num_layers](output_memory)enc_outputs_coord_unact = self.decoder.bbox_embed[self.decoder.num_layers](output_memory) + output_proposalstopk = self.two_stage_num_proposalstopk_proposals = torch.topk(enc_outputs_class[..., 0], topk, dim=1)[1]topk_coords_unact = torch.gather(enc_outputs_coord_unact, 1, topk_proposals.unsqueeze(-1).repeat(1, 1, 4))topk_coords_unact = topk_coords_unact.detach()reference_points = topk_coords_unact.sigmoid()init_reference_out = reference_pointspos_trans_out = self.pos_trans_norm(self.pos_trans(self.get_proposal_pos_embed(topk_coords_unact)))query_embed, tgt = torch.split(pos_trans_out, c, dim=2)else:# 这是非双阶段版本的Deformable DETR# 将query_embed划分为query_embed和tgt两部分query_embed, tgt = torch.split(query_embed, c, dim=1)# 复制bs份query_embed = query_embed.unsqueeze(0).expand(bs, -1, -1)tgt = tgt.unsqueeze(0).expand(bs, -1, -1)# nn.Linear得到每个object queries对应的reference point, 这是decoder参考点的方法!!!reference_points = self.reference_points(query_embed).sigmoid()init_reference_out = reference_points# decoderhs, inter_references = self.decoder(tgt, reference_points, memory,spatial_shapes, level_start_index, valid_ratios, query_embed, mask_flatten)inter_references_out = inter_referencesif self.two_stage:return hs, init_reference_out, inter_references_out, enc_outputs_class, enc_outputs_coord_unactreturn hs, init_reference_out, inter_references_out, None, None

Deformable DETR效率高并且收敛快，核心是Multi-Scale Deformable Attention Module。解决了DETR中收敛慢以及小目标性能低的问题。

本文链接：https://my.lmcjl.com/post/5702.html