詳細(xì)代碼注釋如下，這里要控制是否使用iterative bounding box refinement和two stage技巧。iterative bounding box refinement其實(shí)就是對(duì)參考點(diǎn)的位置進(jìn)行微調(diào)。two stage方法其實(shí)就是通過參考點(diǎn)直接生成anchor但是只取最高置信度的前幾個(gè)，然后再送入decoder進(jìn)行調(diào)整。intermediate數(shù)組是一個(gè)trick，每層Decoder都是可以輸出bbox和分類信息的，如果都利用起來算損失則成為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 attention
        self.cross_attn = MSDeformAttn(d_model, n_levels, n_heads, n_points)
        self.dropout1 = nn.Dropout(dropout)
        self.norm1 = nn.LayerNorm(d_model)

        # self attention
        self.self_attn = nn.MultiheadAttention(d_model, n_heads, dropout=dropout)
        self.dropout2 = nn.Dropout(dropout)
        self.norm2 = nn.LayerNorm(d_model)

        # ffn
        self.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)

    @staticmethod
    def with_pos_embed(tensor, pos):
        return tensor if pos is None else tensor + pos

    def forward_ffn(self, tgt):
        tgt2 = self.linear2(self.dropout3(self.activation(self.linear1(tgt))))
        tgt = tgt + self.dropout4(tgt2)
        tgt = self.norm3(tgt)
        return tgt

    def forward(self, tgt, query_pos, reference_points, src, src_spatial_shapes, level_start_index, src_padding_mask=None):
        # self attention
        q = 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 attention
        tgt2 = 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)

        # ffn
        tgt = self.forward_ffn(tgt)

        return tgt


class 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_layers
        self.return_intermediate = return_intermediate
        # hack implementation for iterative bounding box refinement and two-stage Deformable DETR
        self.bbox_embed = None
        self.class_embed = None

    def 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

        # 用來存儲(chǔ)中間decoder輸出的 可以考慮是否用auxiliary loss
        intermediate = []
        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] == 2
                reference_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是對(duì)decoder中的參考點(diǎn)進(jìn)行微調(diào)，類似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] == 2
                    new_reference_points = tmp
                    new_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

綜合模塊代碼如下

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_model
        self.nhead = nhead
        self.two_stage = two_stage
        self.two_stage_num_proposals = two_stage_num_proposals

        encoder_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 = 128
        temperature = 10000
        scale = 2 * math.pi

        dim_t = torch.arange(num_pos_feats, dtype=torch.float32, device=proposals.device)
        dim_t = temperature ** (2 * (dim_t // 2) / num_pos_feats)
        # N, L, 4
        proposals = proposals.sigmoid() * scale
        # N, L, 4, 128
        pos = proposals[:, :, :, None] / dim_t
        # N, L, 4, 64, 2
        pos = torch.stack((pos[:, :, :, 0::2].sin(), pos[:, :, :, 1::2].cos()), dim=4).flatten(2)
        return pos

    def gen_encoder_output_proposals(self, memory, memory_padding_mask, spatial_shapes):
        N_, S_, C_ = memory.shape
        base_scale = 4.0
        proposals = []
        _cur = 0
        for 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) / scale
            wh = 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 = memory
        output_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_proposals

    def get_valid_ratio(self, mask):
        _, H, W = mask.shape
        valid_H = torch.sum(~mask[:, :, 0], 1)
        valid_W = torch.sum(~mask[:, 0, :], 1)
        valid_ratio_h = valid_H.float() / H
        valid_ratio_w = valid_W.float() / W
        valid_ratio = torch.stack([valid_ratio_w, valid_ratio_h], -1)
        return valid_ratio

    def forward(self, srcs, masks, pos_embeds, query_embed=None):
        assert self.two_stage or query_embed is not None

        # prepare input for encoder
        src_flatten = []
        mask_flatten = []
        lvl_pos_embed_flatten = []
        spatial_shapes = []
        for lvl, (src, mask, pos_embed) in enumerate(zip(srcs, masks, pos_embeds)):
            # 得到每一層feature map的batch size 通道數(shù)量 高寬
            bs, c, h, w = src.shape
            spatial_shape = (h, w)
            spatial_shapes.append(spatial_shape)
            # 將每層的feature map、mask、位置編碼拉平，并且加入到相關(guān)數(shù)組中
            src = src.flatten(2).transpose(1, 2)
            mask = mask.flatten(1)
            pos_embed = pos_embed.flatten(2).transpose(1, 2)
            # 位置編碼和可學(xué)習(xí)的每層編碼相加，表征類似 3D position
            lvl_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維度上進(jìn)行拼接，也就是number token數(shù)量一樣的那個(gè)維度
        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)
        # 記錄每個(gè)level開始的索引以及有效的長(zhǎng)寬(因?yàn)橛衜ask存在，raw image的分辨率可能不統(tǒng)一) 具體查看get_valid_ratio函數(shù)
        # prod(1)計(jì)算h*w，cumsum(0)計(jì)算前綴和
        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 decoder
        bs, _, c = memory.shape
        # 是否使用兩階段模式
        if self.two_stage:
            output_memory, output_proposals = self.gen_encoder_output_proposals(memory, mask_flatten, spatial_shapes)

            # hack implementation for two-stage Deformable DETR
            enc_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_proposals

            topk = self.two_stage_num_proposals
            topk_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_points
            pos_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:
            # 這是非雙階段版本的Deformable DETR
            # 將query_embed劃分為query_embed和tgt兩部分
            query_embed, tgt = torch.split(query_embed, c, dim=1)
            # 復(fù)制bs份
            query_embed = query_embed.unsqueeze(0).expand(bs, -1, -1)
            tgt = tgt.unsqueeze(0).expand(bs, -1, -1)
            # nn.Linear得到每個(gè)object queries對(duì)應(yīng)的reference point, 這是decoder參考點(diǎn)的方法!!!
            reference_points = self.reference_points(query_embed).sigmoid()
            init_reference_out = reference_points

        # decoder
        hs, inter_references = self.decoder(tgt, reference_points, memory,
                                            spatial_shapes, level_start_index, valid_ratios, query_embed, mask_flatten)

        inter_references_out = inter_references
        if self.two_stage:
            return hs, init_reference_out, inter_references_out, enc_outputs_class, enc_outputs_coord_unact
        return hs, init_reference_out, inter_references_out, None, None

圖4. Deformable DETR性能對(duì)比

由圖4可知，Deformable DETR不僅收斂速率比DETR快并且小目標(biāo)精度也高了許多。

Deformable DETR效率高并且收斂快，核心是Multi-Scale Deformable Attention Module。解決了DETR中收斂慢以及小目標(biāo)性能低的問題。

Reference

Deformable DETR：https://arxiv.org/pdf/2010.04159v4

官方代碼倉(cāng)庫(kù)：https://github.com/fundamentalvision/Deformable-DETR

DCNv2：https://arxiv.org/pdf/2008.13535v2.pdf