学习资源站

RT-DETR改进策略【Conv和Transformer】CVPR-2023BiFormer稀疏自注意力，减少内存占用_mhsartdetr-

RT-DETR改进策略【Conv和Transformer】| CVPR-2023 BiFormer 稀疏自注意力，减少内存占用

一、本文介绍

本文记录的是 利用 BiFormer 双级路由注意力机制优化 RT-DETR 的目标检测网络模型 。传统的 多头自注意力（MHSA） 复杂度高，随着输入空间分辨率增加，计算量呈平方增长，导致严重的可扩展性问题。 而本文所使用的 BiFormer 在获取全局信息的同时，通过区域到区域路由和令牌到令牌注意力的合理设计，减少了不必要的计算量。

二、BiFormer介绍

BiFormer: Vision Transformer with Bi-Level Routing Attention

2.1 出发点

  • 解决MHSA的可扩展性问题 ：传统的多head自注意力（MHSA）复杂度高，随着输入空间分辨率增加，计算量呈平方增长，导致严重的可扩展性问题。例如，在图像识别任务中，高分辨率图像会使MHSA的计算负担过重。
  • 实现动态、查询感知的稀疏注意力 ：现有的稀疏注意力机制要么使用手工制作的静态模式，要么在所有查询中共享采样的键值对子集。而实际中不同查询在不同语义区域应关注不同的键值对，所以需要一种能根据查询动态确定关注的键值对的机制。

2.2 原理

2.2.1 区域到区域路由（Region-to-Region Routing）

  • 粗粒度区域过滤 ：首先在粗粒度区域级别过滤掉最不相关的键值对。对于给定的二维输入特征图 X ∈ R H × W × C X\in\mathbb{R}^{H\times W\times C} X ∈ R H × W × C ，将其划分为 S × S S\times S S × S 个非重叠区域，通过计算区域级别的查询和键的邻接矩阵 A r A^{r} A r ，并保留每个区域的前 k k k 个连接，得到路由索引矩阵 I r I^{r} I r 。
  • 确定关注区域 ：通过上述步骤，每个区域只需要关注前 k k k 个路由区域，从而减少了需要计算的区域数量，实现了初步的稀疏化。

2.2.2 令牌到令牌注意力（Token-to-Token Attention）

  • 收集键值对 ：在确定了关注区域后，由于这些区域可能在整个特征图中是分散的，直接进行稀疏矩阵乘法在现代GPU上效率低下。所以先收集这些区域对应的键和值张量 K g = g a t h e r ( K , I r ) K^{g}=gather(K, I^{r}) K g = g a t h er ( K , I r ) ， V g = g a t h e r ( V , I r ) V^{g}=gather(V, I^{r}) V g = g a t h er ( V , I r ) 。
  • 计算注意力 ：然后对收集到的键值对应用注意力机制 O = A t t e n t i o n ( Q , K g , V g ) + L C E ( V ) O = Attention(Q, K^{g}, V^{g})+LCE(V) O = A tt e n t i o n ( Q , K g , V g ) + L CE ( V ) ，其中还引入了局部上下文增强项 L C E ( V ) LCE(V) L CE ( V ) ，它是用深度卷积参数化的。

2.3 结构

2.3.1 整体架构

BiFormer 采用四阶段金字塔结构，与一些先进的视觉 Transformer 类似。在每个阶段，首先使用重叠的补丁嵌入（在第一阶段）或补丁合并模块（在第二到第四阶段）来降低输入空间分辨率同时增加通道数，然后使用 N i N_{i} N i ​ 个连续的BiFormer块来转换特征。

2.3.1 BiFormer块内部

在每个 BiFormer块 中，首先使用 3 × 3 3\times3 3 × 3 深度卷积来隐式编码相对位置信息，然后依次应用 Bi - Level Routing Attention（BRA）模块 和2层 MLP模块 ，分别用于跨位置关系建模和逐位置嵌入。

在这里插入图片描述

2.4 优势

  • 计算复杂度降低 ：与普通注意力的 O ( ( H W ) 2 ) O((H W)^{2}) O (( H W ) 2 ) 复杂度和准全局轴向注意力的 O ( ( H W ) 3 2 ) O((H W)^{\frac{3}{2}}) O (( H W ) 2 3 ​ ) 复杂度相比， BRA 在适当的区域划分下具有 O ( ( H W ) 4 3 ) O((H W)^{\frac{4}{3}}) O (( H W ) 3 4 ​ ) 的复杂度。通过区域到区域路由和令牌到令牌注意力的合理设计，减少了不必要的计算量。
  • 性能优越
    • 图像分类 ：在ImageNet - 1K数据集上进行图像分类实验，不同计算量组的模型中，BiFormer - T、BiFormer - S和BiFormer - B分别取得了优异的成绩，超过了很多相关方法和现有技术水平。例如BiFormer - S在不使用额外训练数据或训练技巧的情况下，取得了83.8%的top - 1准确率，使用标记标记技术后可进一步提高到84.3%。
    • 目标检测和实例分割 ：在COCO 2017数据集上，使用RetinaNet进行目标检测和Mask R - CNN进行实例分割，BiFormer在小物体检测上表现突出，整体性能也与一些先进方法相当，在实例分割的各项指标上有明显优势。
    • 语义分割 ：在ADE20K数据集上进行语义分割实验，在Semantic FPN和UperNet框架下，BiFormer - S和BiFormer - B分别取得了较好的成绩，相比一些现有方法提高了mIoU值。

论文： https://arxiv.org/pdf/2303.08810
源码： https://github.com/rayleizhu/BiFormer

三、BiFormer的实现代码

BiFormer模块 的实现代码如下： 

from typing import Tuple, Optional

import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange
from torch import Tensor, LongTensor

from ultralytics.nn.modules.conv import LightConv
from ultralytics.utils.torch_utils import fuse_conv_and_bn

class TopkRouting(nn.Module):
    """
    differentiable topk routing with scaling
    Args:
        qk_dim: int, feature dimension of query and key
        topk: int, the 'topk'
        qk_scale: int or None, temperature (multiply) of softmax activation
        with_param: bool, wether inorporate learnable params in routing unit
        diff_routing: bool, wether make routing differentiable
        soft_routing: bool, wether make output value multiplied by routing weights
    """

    def __init__(self, qk_dim, topk=4, qk_scale=None, param_routing=False, diff_routing=False):
        super().__init__()
        self.topk = topk
        self.qk_dim = qk_dim
        self.scale = qk_scale or qk_dim ** -0.5
        self.diff_routing = diff_routing
        # TODO: norm layer before/after linear?
        self.emb = nn.Linear(qk_dim, qk_dim) if param_routing else nn.Identity()
        # routing activation
        self.routing_act = nn.Softmax(dim=-1)

    def forward(self, query: Tensor, key: Tensor) -> Tuple[Tensor]:
        """
        Args:
            q, k: (n, p^2, c) tensor
        Return:
            r_weight, topk_index: (n, p^2, topk) tensor
        """
        if not self.diff_routing:
            query, key = query.detach(), key.detach()
        query_hat, key_hat = self.emb(query), self.emb(key)  # per-window pooling -> (n, p^2, c)
        attn_logit = (query_hat * self.scale) @ key_hat.transpose(-2, -1)  # (n, p^2, p^2)
        topk_attn_logit, topk_index = torch.topk(attn_logit, k=self.topk, dim=-1)  # (n, p^2, k), (n, p^2, k)
        r_weight = self.routing_act(topk_attn_logit)  # (n, p^2, k)

        return r_weight, topk_index

class KVGather(nn.Module):
    def __init__(self, mul_weight='none'):
        super().__init__()
        assert mul_weight in ['none', 'soft', 'hard']
        self.mul_weight = mul_weight

    def forward(self, r_idx: Tensor, r_weight: Tensor, kv: Tensor):
        """
        r_idx: (n, p^2, topk) tensor
        r_weight: (n, p^2, topk) tensor
        kv: (n, p^2, w^2, c_kq+c_v)

        Return:
            (n, p^2, topk, w^2, c_kq+c_v) tensor
        """
        # select kv according to routing index
        n, p2, w2, c_kv = kv.size()
        topk = r_idx.size(-1)
        # print(r_idx.size(), r_weight.size())
        # FIXME: gather consumes much memory (topk times redundancy), write cuda kernel?
        topk_kv = torch.gather(kv.view(n, 1, p2, w2, c_kv).expand(-1, p2, -1, -1, -1),
                               # (n, p^2, p^2, w^2, c_kv) without mem cpy
                               dim=2,
                               index=r_idx.view(n, p2, topk, 1, 1).expand(-1, -1, -1, w2, c_kv)
                               # (n, p^2, k, w^2, c_kv)
                               )

        if self.mul_weight == 'soft':
            topk_kv = r_weight.view(n, p2, topk, 1, 1) * topk_kv  # (n, p^2, k, w^2, c_kv)
        elif self.mul_weight == 'hard':
            raise NotImplementedError('differentiable hard routing TBA')
        # else: #'none'
        #     topk_kv = topk_kv # do nothing

        return topk_kv

class QKVLinear(nn.Module):
    def __init__(self, dim, qk_dim, bias=True):
        super().__init__()
        self.dim = dim
        self.qk_dim = qk_dim
        self.qkv = nn.Linear(dim, qk_dim + qk_dim + dim, bias=bias)

    def forward(self, x):
        q, kv = self.qkv(x).split([self.qk_dim, self.qk_dim + self.dim], dim=-1)
        return q, kv
        # q, k, v = self.qkv(x).split([self.qk_dim, self.qk_dim, self.dim], dim=-1)
        # return q, k, v

class BiLevelRoutingAttention(nn.Module):
    """
    n_win: number of windows in one side (so the actual number of windows is n_win*n_win)
    kv_per_win: for kv_downsample_mode='ada_xxxpool' only, number of key/values per window. Similar to n_win, the actual number is kv_per_win*kv_per_win.
    topk: topk for window filtering
    param_attention: 'qkvo'-linear for q,k,v and o, 'none': param free attention
    param_routing: extra linear for routing
    diff_routing: wether to set routing differentiable
    soft_routing: wether to multiply soft routing weights
    """

    def __init__(self, dim, n_win=7, num_heads=8, qk_dim=None, qk_scale=None,
                 kv_per_win=4, kv_downsample_ratio=4, kv_downsample_kernel=None, kv_downsample_mode='identity',
                 topk=4, param_attention="qkvo", param_routing=False, diff_routing=False, soft_routing=False,
                 side_dwconv=3,
                 auto_pad=True):
        super().__init__()
        # local attention setting
        self.dim = dim
        self.n_win = n_win  # Wh, Ww
        self.num_heads = num_heads
        self.qk_dim = qk_dim or dim
        assert self.qk_dim % num_heads == 0 and self.dim % num_heads == 0, 'qk_dim and dim must be divisible by num_heads!'
        self.scale = qk_scale or self.qk_dim ** -0.5

        ################side_dwconv (i.e. LCE in ShuntedTransformer)###########
        self.lepe = nn.Conv2d(dim, dim, kernel_size=side_dwconv, stride=1, padding=side_dwconv // 2,
                              groups=dim) if side_dwconv > 0 else \
            lambda x: torch.zeros_like(x)

        ################ global routing setting #################
        self.topk = topk
        self.param_routing = param_routing
        self.diff_routing = diff_routing
        self.soft_routing = soft_routing
        # router
        assert not (self.param_routing and not self.diff_routing)  # cannot be with_param=True and diff_routing=False
        self.router = TopkRouting(qk_dim=self.qk_dim,
                                  qk_scale=self.scale,
                                  topk=self.topk,
                                  diff_routing=self.diff_routing,
                                  param_routing=self.param_routing)
        if self.soft_routing:  # soft routing, always diffrentiable (if no detach)
            mul_weight = 'soft'
        elif self.diff_routing:  # hard differentiable routing
            mul_weight = 'hard'
        else:  # hard non-differentiable routing
            mul_weight = 'none'
        self.kv_gather = KVGather(mul_weight=mul_weight)

        # qkv mapping (shared by both global routing and local attention)
        self.param_attention = param_attention
        if self.param_attention == 'qkvo':
            self.qkv = QKVLinear(self.dim, self.qk_dim)
            self.wo = nn.Linear(dim, dim)
        elif self.param_attention == 'qkv':
            self.qkv = QKVLinear(self.dim, self.qk_dim)
            self.wo = nn.Identity()
        else:
            raise ValueError(f'param_attention mode {self.param_attention} is not surpported!')

        self.kv_downsample_mode = kv_downsample_mode
        self.kv_per_win = kv_per_win
        self.kv_downsample_ratio = kv_downsample_ratio
        self.kv_downsample_kenel = kv_downsample_kernel
        if self.kv_downsample_mode == 'ada_avgpool':
            assert self.kv_per_win is not None
            self.kv_down = nn.AdaptiveAvgPool2d(self.kv_per_win)
        elif self.kv_downsample_mode == 'ada_maxpool':
            assert self.kv_per_win is not None
            self.kv_down = nn.AdaptiveMaxPool2d(self.kv_per_win)
        elif self.kv_downsample_mode == 'maxpool':
            assert self.kv_downsample_ratio is not None
            self.kv_down = nn.MaxPool2d(self.kv_downsample_ratio) if self.kv_downsample_ratio > 1 else nn.Identity()
        elif self.kv_downsample_mode == 'avgpool':
            assert self.kv_downsample_ratio is not None
            self.kv_down = nn.AvgPool2d(self.kv_downsample_ratio) if self.kv_downsample_ratio > 1 else nn.Identity()
        elif self.kv_downsample_mode == 'identity':  # no kv downsampling
            self.kv_down = nn.Identity()
        elif self.kv_downsample_mode == 'fracpool':
            # assert self.kv_downsample_ratio is not None
            # assert self.kv_downsample_kenel is not None
            # TODO: fracpool
            # 1. kernel size should be input size dependent
            # 2. there is a random factor, need to avoid independent sampling for k and v
            raise NotImplementedError('fracpool policy is not implemented yet!')
        elif kv_downsample_mode == 'conv':
            # TODO: need to consider the case where k != v so that need two downsample modules
            raise NotImplementedError('conv policy is not implemented yet!')
        else:
            raise ValueError(f'kv_down_sample_mode {self.kv_downsaple_mode} is not surpported!')

        # softmax for local attention
        self.attn_act = nn.Softmax(dim=-1)

        self.auto_pad = auto_pad

    def forward(self, x, ret_attn_mask=False):
        """
        x: NHWC tensor

        Return:
            NHWC tensor
        """
        x = rearrange(x, "n c h w -> n h w c")
        # NOTE: use padding for semantic segmentation
        ###################################################
        if self.auto_pad:
            N, H_in, W_in, C = x.size()

            pad_l = pad_t = 0
            pad_r = (self.n_win - W_in % self.n_win) % self.n_win
            pad_b = (self.n_win - H_in % self.n_win) % self.n_win
            x = F.pad(x, (0, 0,  # dim=-1
                          pad_l, pad_r,  # dim=-2
                          pad_t, pad_b))  # dim=-3
            _, H, W, _ = x.size()  # padded size
        else:
            N, H, W, C = x.size()
            assert H % self.n_win == 0 and W % self.n_win == 0  #
        ###################################################

        # patchify, (n, p^2, w, w, c), keep 2d window as we need 2d pooling to reduce kv size
        x = rearrange(x, "n (j h) (i w) c -> n (j i) h w c", j=self.n_win, i=self.n_win)

        #################qkv projection###################
        # q: (n, p^2, w, w, c_qk)
        # kv: (n, p^2, w, w, c_qk+c_v)
        # NOTE: separte kv if there were memory leak issue caused by gather
        q, kv = self.qkv(x)

        # pixel-wise qkv
        # q_pix: (n, p^2, w^2, c_qk)
        # kv_pix: (n, p^2, h_kv*w_kv, c_qk+c_v)
        q_pix = rearrange(q, 'n p2 h w c -> n p2 (h w) c')
        kv_pix = self.kv_down(rearrange(kv, 'n p2 h w c -> (n p2) c h w'))
        kv_pix = rearrange(kv_pix, '(n j i) c h w -> n (j i) (h w) c', j=self.n_win, i=self.n_win)

        q_win, k_win = q.mean([2, 3]), kv[..., 0:self.qk_dim].mean(
            [2, 3])  # window-wise qk, (n, p^2, c_qk), (n, p^2, c_qk)

        ##################side_dwconv(lepe)##################
        # NOTE: call contiguous to avoid gradient warning when using ddp
        lepe = self.lepe(rearrange(kv[..., self.qk_dim:], 'n (j i) h w c -> n c (j h) (i w)', j=self.n_win,
                                   i=self.n_win).contiguous())
        lepe = rearrange(lepe, 'n c (j h) (i w) -> n (j h) (i w) c', j=self.n_win, i=self.n_win)

        ############ gather q dependent k/v #################

        r_weight, r_idx = self.router(q_win, k_win)  # both are (n, p^2, topk) tensors

        kv_pix_sel = self.kv_gather(r_idx=r_idx, r_weight=r_weight, kv=kv_pix)  # (n, p^2, topk, h_kv*w_kv, c_qk+c_v)
        k_pix_sel, v_pix_sel = kv_pix_sel.split([self.qk_dim, self.dim], dim=-1)
        # kv_pix_sel: (n, p^2, topk, h_kv*w_kv, c_qk)
        # v_pix_sel: (n, p^2, topk, h_kv*w_kv, c_v)

        ######### do attention as normal ####################
        k_pix_sel = rearrange(k_pix_sel, 'n p2 k w2 (m c) -> (n p2) m c (k w2)',
                              m=self.num_heads)  # flatten to BMLC, (n*p^2, m, topk*h_kv*w_kv, c_kq//m) transpose here?
        v_pix_sel = rearrange(v_pix_sel, 'n p2 k w2 (m c) -> (n p2) m (k w2) c',
                              m=self.num_heads)  # flatten to BMLC, (n*p^2, m, topk*h_kv*w_kv, c_v//m)
        q_pix = rearrange(q_pix, 'n p2 w2 (m c) -> (n p2) m w2 c',
                          m=self.num_heads)  # to BMLC tensor (n*p^2, m, w^2, c_qk//m)

        # param-free multihead attention
        attn_weight = (
                              q_pix * self.scale) @ k_pix_sel  # (n*p^2, m, w^2, c) @ (n*p^2, m, c, topk*h_kv*w_kv) -> (n*p^2, m, w^2, topk*h_kv*w_kv)
        attn_weight = self.attn_act(attn_weight)
        out = attn_weight @ v_pix_sel  # (n*p^2, m, w^2, topk*h_kv*w_kv) @ (n*p^2, m, topk*h_kv*w_kv, c) -> (n*p^2, m, w^2, c)
        out = rearrange(out, '(n j i) m (h w) c -> n (j h) (i w) (m c)', j=self.n_win, i=self.n_win,
                        h=H // self.n_win, w=W // self.n_win)

        out = out + lepe
        # output linear
        out = self.wo(out)

        # NOTE: use padding for semantic segmentation
        # crop padded region
        if self.auto_pad and (pad_r > 0 or pad_b > 0):
            out = out[:, :H_in, :W_in, :].contiguous()

        if ret_attn_mask:
            return out, r_weight, r_idx, attn_weight
        else:
            return rearrange(out, "n h w c -> n c h w")

# class Attention(nn.Module):
#     """
#     vanilla attention
#     """

#     def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0.):
#         super().__init__()
#         self.num_heads = num_heads
#         head_dim = dim // num_heads
#         # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
#         self.scale = qk_scale or head_dim ** -0.5

#         self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
#         self.attn_drop = nn.Dropout(attn_drop)
#         self.proj = nn.Linear(dim, dim)
#         self.proj_drop = nn.Dropout(proj_drop)

#     def forward(self, x):
#         """
#         args:
#             x: NCHW tensor
#         return:
#             NCHW tensor
#         """
#         _, _, H, W = x.size()
#         x = rearrange(x, 'n c h w -> n (h w) c')

#         #######################################
#         B, N, C = x.shape
#         qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
#         q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)

#         attn = (q @ k.transpose(-2, -1)) * self.scale
#         attn = attn.softmax(dim=-1)
#         attn = self.attn_drop(attn)

#         x = (attn @ v).transpose(1, 2).reshape(B, N, C)
#         x = self.proj(x)
#         x = self.proj_drop(x)
#         #######################################

#         x = rearrange(x, 'n (h w) c -> n c h w', h=H, w=W)
#         return x

class AttentionLePE(nn.Module):
    """
    vanilla attention
    """

    def __init__(self, dim, num_heads=8, qkv_bias=False, qk_scale=None, attn_drop=0., proj_drop=0., side_dwconv=5):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // num_heads
        # NOTE scale factor was wrong in my original version, can set manually to be compat with prev weights
        self.scale = qk_scale or head_dim ** -0.5

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)
        self.lepe = nn.Conv2d(dim, dim, kernel_size=side_dwconv, stride=1, padding=side_dwconv // 2,
                              groups=dim) if side_dwconv > 0 else \
            lambda x: torch.zeros_like(x)

    def forward(self, x):
        """
        args:
            x: NCHW tensor
        return:
            NCHW tensor
        """
        _, _, H, W = x.size()
        x = rearrange(x, 'n c h w -> n (h w) c')

        #######################################
        B, N, C = x.shape
        qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)

        lepe = self.lepe(rearrange(x, 'n (h w) c -> n c h w', h=H, w=W))
        lepe = rearrange(lepe, 'n c h w -> n (h w) c')

        attn = (q @ k.transpose(-2, -1)) * self.scale
        attn = attn.softmax(dim=-1)
        attn = self.attn_drop(attn)

        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
        x = x + lepe

        x = self.proj(x)
        x = self.proj_drop(x)
        #######################################

        x = rearrange(x, 'n (h w) c -> n c h w', h=H, w=W)
        return x

def _grid2seq(x: Tensor, region_size: Tuple[int], num_heads: int):
    """
    Args:
        x: BCHW tensor
        region size: int
        num_heads: number of attention heads
    Return:
        out: rearranged x, has a shape of (bs, nhead, nregion, reg_size, head_dim)
        region_h, region_w: number of regions per col/row
    """
    B, C, H, W = x.size()
    region_h, region_w = H // region_size[0], W // region_size[1]
    x = x.view(B, num_heads, C // num_heads, region_h, region_size[0], region_w, region_size[1])
    x = torch.einsum('bmdhpwq->bmhwpqd', x).flatten(2, 3).flatten(-3, -2)  # (bs, nhead, nregion, reg_size, head_dim)
    return x, region_h, region_w

def _seq2grid(x: Tensor, region_h: int, region_w: int, region_size: Tuple[int]):
    """
    Args:
        x: (bs, nhead, nregion, reg_size^2, head_dim)
    Return:
        x: (bs, C, H, W)
    """
    bs, nhead, nregion, reg_size_square, head_dim = x.size()
    x = x.view(bs, nhead, region_h, region_w, region_size[0], region_size[1], head_dim)
    x = torch.einsum('bmhwpqd->bmdhpwq', x).reshape(bs, nhead * head_dim,
                                                    region_h * region_size[0], region_w * region_size[1])
    return x

def regional_routing_attention_torch(
        query: Tensor, key: Tensor, value: Tensor, scale: float,
        region_graph: LongTensor, region_size: Tuple[int],
        kv_region_size: Optional[Tuple[int]] = None,
        auto_pad=True) -> Tensor:
    """
    Args:
        query, key, value: (B, C, H, W) tensor
        scale: the scale/temperature for dot product attention
        region_graph: (B, nhead, h_q*w_q, topk) tensor, topk <= h_k*w_k
        region_size: region/window size for queries, (rh, rw)
        key_region_size: optional, if None, key_region_size=region_size
        auto_pad: required to be true if the input sizes are not divisible by the region_size
    Return:
        output: (B, C, H, W) tensor
        attn: (bs, nhead, q_nregion, reg_size, topk*kv_region_size) attention matrix
    """
    kv_region_size = kv_region_size or region_size
    bs, nhead, q_nregion, topk = region_graph.size()

    # Auto pad to deal with any input size
    q_pad_b, q_pad_r, kv_pad_b, kv_pad_r = 0, 0, 0, 0
    if auto_pad:
        _, _, Hq, Wq = query.size()
        q_pad_b = (region_size[0] - Hq % region_size[0]) % region_size[0]
        q_pad_r = (region_size[1] - Wq % region_size[1]) % region_size[1]
        if (q_pad_b > 0 or q_pad_r > 0):
            query = F.pad(query, (0, q_pad_r, 0, q_pad_b))  # zero padding

        _, _, Hk, Wk = key.size()
        kv_pad_b = (kv_region_size[0] - Hk % kv_region_size[0]) % kv_region_size[0]
        kv_pad_r = (kv_region_size[1] - Wk % kv_region_size[1]) % kv_region_size[1]
        if (kv_pad_r > 0 or kv_pad_b > 0):
            key = F.pad(key, (0, kv_pad_r, 0, kv_pad_b))  # zero padding
            value = F.pad(value, (0, kv_pad_r, 0, kv_pad_b))  # zero padding

    # to sequence format, i.e. (bs, nhead, nregion, reg_size, head_dim)
    query, q_region_h, q_region_w = _grid2seq(query, region_size=region_size, num_heads=nhead)
    key, _, _ = _grid2seq(key, region_size=kv_region_size, num_heads=nhead)
    value, _, _ = _grid2seq(value, region_size=kv_region_size, num_heads=nhead)

    # gather key and values.
    # TODO: is seperate gathering slower than fused one (our old version) ?
    # torch.gather does not support broadcasting, hence we do it manually
    bs, nhead, kv_nregion, kv_region_size, head_dim = key.size()
    broadcasted_region_graph = region_graph.view(bs, nhead, q_nregion, topk, 1, 1). \
        expand(-1, -1, -1, -1, kv_region_size, head_dim)
    key_g = torch.gather(key.view(bs, nhead, 1, kv_nregion, kv_region_size, head_dim). \
                         expand(-1, -1, query.size(2), -1, -1, -1), dim=3,
                         index=broadcasted_region_graph)  # (bs, nhead, q_nregion, topk, kv_region_size, head_dim)
    value_g = torch.gather(value.view(bs, nhead, 1, kv_nregion, kv_region_size, head_dim). \
                           expand(-1, -1, query.size(2), -1, -1, -1), dim=3,
                           index=broadcasted_region_graph)  # (bs, nhead, q_nregion, topk, kv_region_size, head_dim)

    # token-to-token attention
    # (bs, nhead, q_nregion, reg_size, head_dim) @ (bs, nhead, q_nregion, head_dim, topk*kv_region_size)
    # -> (bs, nhead, q_nregion, reg_size, topk*kv_region_size)
    # TODO: mask padding region
    attn = (query * scale) @ key_g.flatten(-3, -2).transpose(-1, -2)
    attn = torch.softmax(attn, dim=-1)
    # (bs, nhead, q_nregion, reg_size, topk*kv_region_size) @ (bs, nhead, q_nregion, topk*kv_region_size, head_dim)
    # -> (bs, nhead, q_nregion, reg_size, head_dim)
    output = attn @ value_g.flatten(-3, -2)

    # to BCHW format
    output = _seq2grid(output, region_h=q_region_h, region_w=q_region_w, region_size=region_size)

    # remove paddings if needed
    if auto_pad and (q_pad_b > 0 or q_pad_r > 0):
        output = output[:, :, :Hq, :Wq]

    return output, attn

class BiLevelRoutingAttention_nchw(nn.Module):
    """Bi-Level Routing Attention that takes nchw input

    Compared to legacy version, this implementation:
    * removes unused args and components
    * uses nchw input format to avoid frequent permutation

    When the size of inputs is not divisible by the region size, there is also a numerical difference
    than legacy implementation, due to:
    * different way to pad the input feature map (padding after linear projection)
    * different pooling behavior (count_include_pad=False)

    Current implementation is more reasonable, hence we do not keep backward numerical compatiability
    """

    def __init__(self, dim, num_heads=1, n_win=7, qk_scale=None, topk=4, side_dwconv=3, auto_pad=False,
                 attn_backend='torch'):
        super().__init__()
        # local attention setting
        self.dim = dim
        self.num_heads = num_heads
        assert self.dim % num_heads == 0, 'dim must be divisible by num_heads!'
        self.head_dim = self.dim // self.num_heads
        self.scale = qk_scale or self.dim ** -0.5  # NOTE: to be consistent with old models.

        ################side_dwconv (i.e. LCE in Shunted Transformer)###########
        self.lepe = nn.Conv2d(dim, dim, kernel_size=side_dwconv, stride=1, padding=side_dwconv // 2,
                              groups=dim) if side_dwconv > 0 else \
            lambda x: torch.zeros_like(x)

        ################ regional routing setting #################
        self.topk = topk
        self.n_win = n_win  # number of windows per row/col

        ##########################################

        self.qkv_linear = nn.Conv2d(self.dim, 3 * self.dim, kernel_size=1)
        self.output_linear = nn.Conv2d(self.dim, self.dim, kernel_size=1)

        if attn_backend == 'torch':
            self.attn_fn = regional_routing_attention_torch
        else:
            raise ValueError('CUDA implementation is not available yet. Please stay tuned.')

    def forward(self, x: Tensor, ret_attn_mask=False):
        """
        Args:
            x: NCHW tensor, better to be channel_last (https://pytorch.org/tutorials/intermediate/memory_format_tutorial.html)
        Return:
            NCHW tensor
        """
        N, C, H, W = x.size()
        region_size = (H // self.n_win, W // self.n_win)

        # STEP 1: linear projection
        qkv = self.qkv_linear.forward(x)  # ncHW
        q, k, v = qkv.chunk(3, dim=1)  # ncHW

        # STEP 2: region-to-region routing
        # NOTE: ceil_mode=True, count_include_pad=False = auto padding
        # NOTE: gradients backward through token-to-token attention. See Appendix A for the intuition.
        q_r = F.avg_pool2d(q.detach(), kernel_size=region_size, ceil_mode=True, count_include_pad=False)
        k_r = F.avg_pool2d(k.detach(), kernel_size=region_size, ceil_mode=True, count_include_pad=False)  # nchw
        q_r: Tensor = q_r.permute(0, 2, 3, 1).flatten(1, 2)  # n(hw)c
        k_r: Tensor = k_r.flatten(2, 3)  # nc(hw)
        a_r = q_r @ k_r  # n(hw)(hw), adj matrix of regional graph
        _, idx_r = torch.topk(a_r, k=self.topk, dim=-1)  # n(hw)k long tensor
        idx_r: LongTensor = idx_r.unsqueeze_(1).expand(-1, self.num_heads, -1, -1)

        # STEP 3: token to token attention (non-parametric function)
        output, attn_mat = self.attn_fn(query=q, key=k, value=v, scale=self.scale,
                                        region_graph=idx_r, region_size=region_size
                                        )

        output = output + self.lepe(v)  # ncHW
        output = self.output_linear(output)  # ncHW

        if ret_attn_mask:
            return output, attn_mat

        return output

def autopad(k, p=None, d=1):  # kernel, padding, dilation
    """Pad to 'same' shape outputs."""
    if d > 1:
        k = d * (k - 1) + 1 if isinstance(k, int) else [d * (x - 1) + 1 for x in k]  # actual kernel-size
    if p is None:
        p = k // 2 if isinstance(k, int) else [x // 2 for x in k]  # auto-pad
    return p

class Conv(nn.Module):
    """Standard convolution with args(ch_in, ch_out, kernel, stride, padding, groups, dilation, activation)."""
 
    default_act = nn.SiLU()  # default activation
 
    def __init__(self, c1, c2, k=1, s=1, p=None, g=1, d=1, act=True):
        """Initialize Conv layer with given arguments including activation."""
        super().__init__()
        self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p, d), groups=g, dilation=d, bias=False)
        self.bn = nn.BatchNorm2d(c2)
        self.act = self.default_act if act is True else act if isinstance(act, nn.Module) else nn.Identity()
 
    def forward(self, x):
        """Apply convolution, batch normalization and activation to input tensor."""
        return self.act(self.bn(self.conv(x)))
 
    def forward_fuse(self, x):
        """Perform transposed convolution of 2D data."""
        return self.act(self.conv(x))

class HGBlock_Biformer(nn.Module):
    """
    HG_Block of PPHGNetV2 with 2 convolutions and LightConv.

    https://github.com/PaddlePaddle/PaddleDetection/blob/develop/ppdet/modeling/backbones/hgnet_v2.py
    """

    def __init__(self, c1, cm, c2, k=3, n=6, lightconv=False, shortcut=False, act=nn.ReLU()):
        """Initializes a CSP Bottleneck with 1 convolution using specified input and output channels."""
        super().__init__()
        block = LightConv if lightconv else Conv
        self.m = nn.ModuleList(block(c1 if i == 0 else cm, cm, k=k, act=act) for i in range(n))
        self.sc = Conv(c1 + n * cm, c2 // 2, 1, 1, act=act)  # squeeze conv
        self.ec = Conv(c2 // 2, c2, 1, 1, act=act)  # excitation conv
        self.add = shortcut and c1 == c2
        self.cv = BiLevelRoutingAttention_nchw(c2)
        
    def forward(self, x):
        """Forward pass of a PPHGNetV2 backbone layer."""
        y = [x]
        y.extend(m(y[-1]) for m in self.m)
        y = self.cv(self.ec(self.sc(torch.cat(y, 1))))
        return y + x if self.add else y

四、创新模块

4.1 改进点1

模块改进方法 1️⃣：直接加入 BiLevelRoutingAttention_nchw模块 。( 第五节讲解添加步骤 )
BiLevelRoutingAttention_nchw模块 添加后如下：

在这里插入图片描述

注意❗：在 5.2和5.3小节 中需要声明的模块名称为： BiLevelRoutingAttention_nchw 。

4.2 改进点2⭐

模块改进方法 2️⃣：基于 BiLevelRoutingAttention_nchw 的 HGBlock 。( 第五节讲解添加步骤 )

改进方法是对 RT-DETR 中的 HGBlock模块 进行改进。在将 BiLevelRoutingAttention_nchw模块 与 HGBlock 结合后，在进行特征提取时， 能够获取远距离信息，并且通过区域到区域路由和令牌到令牌注意力的合理设计，相比普通的自注意力，减少了不必要的计算量。

改进代码如下：

首先添加如下代码改进 HGBlock 模块，加入 BiLevelRoutingAttention_nchw ，并将 HGBlock 重命名为 HGBlock_Biformer 

class HGBlock_Biformer(nn.Module):
    """
    HG_Block of PPHGNetV2 with 2 convolutions and LightConv.

    https://github.com/PaddlePaddle/PaddleDetection/blob/develop/ppdet/modeling/backbones/hgnet_v2.py
    """

    def __init__(self, c1, cm, c2, k=3, n=6, lightconv=False, shortcut=False, act=nn.ReLU()):
        """Initializes a CSP Bottleneck with 1 convolution using specified input and output channels."""
        super().__init__()
        block = LightConv if lightconv else Conv
        self.m = nn.ModuleList(block(c1 if i == 0 else cm, cm, k=k, act=act) for i in range(n))
        self.sc = Conv(c1 + n * cm, c2 // 2, 1, 1, act=act)  # squeeze conv
        self.ec = Conv(c2 // 2, c2, 1, 1, act=act)  # excitation conv
        self.add = shortcut and c1 == c2
        self.cv = BiLevelRoutingAttention_nchw(c2)
        
    def forward(self, x):
        """Forward pass of a PPHGNetV2 backbone layer."""
        y = [x]
        y.extend(m(y[-1]) for m in self.m)
        y = self.cv(self.ec(self.sc(torch.cat(y, 1))))
        return y + x if self.add else y

在这里插入图片描述

注意❗：在 4.2小节 中的 yolo.py 文件中需要声明的模块名称为： HGBlock_Biformer 。

五、添加步骤

5.1 修改一

① 在 ultralytics/nn/ 目录下新建 AddModules 文件夹用于存放模块代码

② 在 AddModules 文件夹下新建 Biformer.py ，将 第三节 中的代码粘贴到此处

在这里插入图片描述

5.2 修改二

在 AddModules 文件夹下新建 __init__.py （已有则不用新建），在文件内导入模块： from .Biformer import *

在这里插入图片描述

5.3 修改三

在 ultralytics/nn/modules/tasks.py 文件中，需要在两处位置添加各模块类名称。

首先：导入模块

在这里插入图片描述

其次：在 parse_model函数 中注册 BiLevelRoutingAttention_nchw 和 HGBlock_Biformer 模块

在这里插入图片描述

在这里插入图片描述

在这里插入图片描述

六、yaml模型文件

6.1 模型改进版本1

此处以 ultralytics/cfg/models/rt-detr/rtdetr-l.yaml 为例，在同目录下创建一个用于自己数据集训练的模型文件 rtdetr-l-Biformer.yaml 。

将 rtdetr-l.yaml 中的内容复制到 rtdetr-l-BiFormer.yaml 文件下，修改 nc 数量等于自己数据中目标的数量。

📌 模型的修改方法是将骨干网络中的部分 HGBlock模块 替换成 BiLevelRoutingAttention_nchw模块 。 

# Ultralytics YOLO 🚀, AGPL-3.0 license
# RT-DETR-l object detection model with P3-P5 outputs. For details see https://docs.ultralytics.com/models/rtdetr

# Parameters
nc: 1 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov8n-cls.yaml' will call yolov8-cls.yaml with scale 'n'
  # [depth, width, max_channels]
  l: [1.00, 1.00, 1024]

backbone:
  # [from, repeats, module, args]
  - [-1, 1, HGStem, [32, 48]] # 0-P2/4
  - [-1, 6, HGBlock, [48, 128, 3]] # stage 1

  - [-1, 1, DWConv, [128, 3, 2, 1, False]] # 2-P3/8
  - [-1, 6, HGBlock, [96, 512, 3]] # stage 2

  - [-1, 1, DWConv, [512, 3, 2, 1, False]] # 4-P4/16
  - [-1, 6, HGBlock, [192, 1024, 5, True, False]] # cm, c2, k, light, shortcut
  - [-1, 6, HGBlock, [192, 1024, 5, True, True]]
  - [-1, 6, HGBlock, [192, 1024, 5, True, True]] # stage 3

  - [-1, 1, DWConv, [1024, 3, 2, 1, False]] # 8-P5/32
  - [-1, 6, BiLevelRoutingAttention_nchw, [1024]] # stage 4

head:
  - [-1, 1, Conv, [256, 1, 1, None, 1, 1, False]] # 10 input_proj.2
  - [-1, 1, AIFI, [1024, 8]]
  - [-1, 1, Conv, [256, 1, 1]] # 12, Y5, lateral_convs.0

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [7, 1, Conv, [256, 1, 1, None, 1, 1, False]] # 14 input_proj.1
  - [[-2, -1], 1, Concat, [1]]
  - [-1, 3, RepC3, [256]] # 16, fpn_blocks.0
  - [-1, 1, Conv, [256, 1, 1]] # 17, Y4, lateral_convs.1

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [3, 1, Conv, [256, 1, 1, None, 1, 1, False]] # 19 input_proj.0
  - [[-2, -1], 1, Concat, [1]] # cat backbone P4
  - [-1, 3, RepC3, [256]] # X3 (21), fpn_blocks.1

  - [-1, 1, Conv, [256, 3, 2]] # 22, downsample_convs.0
  - [[-1, 17], 1, Concat, [1]] # cat Y4
  - [-1, 3, RepC3, [256]] # F4 (24), pan_blocks.0

  - [-1, 1, Conv, [256, 3, 2]] # 25, downsample_convs.1
  - [[-1, 12], 1, Concat, [1]] # cat Y5
  - [-1, 3, RepC3, [256]] # F5 (27), pan_blocks.1

  - [[21, 24, 27], 1, RTDETRDecoder, [nc]] # Detect(P3, P4, P5)

6.2 模型改进版本2⭐

此处以 ultralytics/cfg/models/rt-detr/rtdetr-l.yaml 为例，在同目录下创建一个用于自己数据集训练的模型文件 rtdetr-l-HGBlock_Biformer.yaml 。

将 rtdetr-l.yaml 中的内容复制到 rtdetr-l-HGBlock_Biformer.yaml 文件下，修改 nc 数量等于自己数据中目标的数量。

📌 模型的修改方法是将 骨干网络 中的部分 HGBlock模块 替换成 HGBlock_Biformer模块 。 

# Ultralytics YOLO 🚀, AGPL-3.0 license
# RT-DETR-l object detection model with P3-P5 outputs. For details see https://docs.ultralytics.com/models/rtdetr

# Parameters
nc: 1 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolov8n-cls.yaml' will call yolov8-cls.yaml with scale 'n'
  # [depth, width, max_channels]
  l: [1.00, 1.00, 1024]

backbone:
  # [from, repeats, module, args]
  - [-1, 1, HGStem, [32, 48]] # 0-P2/4
  - [-1, 6, HGBlock, [48, 128, 3]] # stage 1

  - [-1, 1, DWConv, [128, 3, 2, 1, False]] # 2-P3/8
  - [-1, 6, HGBlock, [96, 512, 3]] # stage 2

  - [-1, 1, DWConv, [512, 3, 2, 1, False]] # 4-P4/16
  - [-1, 6, HGBlock_Biformer, [192, 1024, 5, True, False]] # cm, c2, k, light, shortcut
  - [-1, 6, HGBlock_Biformer, [192, 1024, 5, True, True]]
  - [-1, 6, HGBlock_Biformer, [192, 1024, 5, True, True]] # stage 3

  - [-1, 1, DWConv, [1024, 3, 2, 1, False]] # 8-P5/32
  - [-1, 6, HGBlock, [384, 2048, 5, True, False]] # stage 4

head:
  - [-1, 1, Conv, [256, 1, 1, None, 1, 1, False]] # 10 input_proj.2
  - [-1, 1, AIFI, [1024, 8]]
  - [-1, 1, Conv, [256, 1, 1]] # 12, Y5, lateral_convs.0

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [7, 1, Conv, [256, 1, 1, None, 1, 1, False]] # 14 input_proj.1
  - [[-2, -1], 1, Concat, [1]]
  - [-1, 3, RepC3, [256]] # 16, fpn_blocks.0
  - [-1, 1, Conv, [256, 1, 1]] # 17, Y4, lateral_convs.1

  - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  - [3, 1, Conv, [256, 1, 1, None, 1, 1, False]] # 19 input_proj.0
  - [[-2, -1], 1, Concat, [1]] # cat backbone P4
  - [-1, 3, RepC3, [256]] # X3 (21), fpn_blocks.1

  - [-1, 1, Conv, [256, 3, 2]] # 22, downsample_convs.0
  - [[-1, 17], 1, Concat, [1]] # cat Y4
  - [-1, 3, RepC3, [256]] # F4 (24), pan_blocks.0

  - [-1, 1, Conv, [256, 3, 2]] # 25, downsample_convs.1
  - [[-1, 12], 1, Concat, [1]] # cat Y5
  - [-1, 3, RepC3, [256]] # F5 (27), pan_blocks.1

  - [[21, 24, 27], 1, RTDETRDecoder, [nc]] # Detect(P3, P4, P5)

七、成功运行结果

打印网络模型可以看到 BiLevelRoutingAttention_nchw 和 HGBlock_Biformer 已经加入到模型中，并可以进行训练了。

rtdetr-l-Biformer ：

rtdetr-l-Biformer summary: 650 layers, 51,089,347 parameters, 51,089,347 gradients, 107.5 GFLOPs 

                   from  n    params  module                                       arguments                     
  0                  -1  1     25248  ultralytics.nn.modules.block.HGStem          [3, 32, 48]                   
  1                  -1  6    155072  ultralytics.nn.modules.block.HGBlock         [48, 48, 128, 3, 6]           
  2                  -1  1      1408  ultralytics.nn.modules.conv.DWConv           [128, 128, 3, 2, 1, False]    
  3                  -1  6    839296  ultralytics.nn.modules.block.HGBlock         [128, 96, 512, 3, 6]          
  4                  -1  1      5632  ultralytics.nn.modules.conv.DWConv           [512, 512, 3, 2, 1, False]    
  5                  -1  6   1695360  ultralytics.nn.modules.block.HGBlock         [512, 192, 1024, 5, 6, True, False]
  6                  -1  6   2055808  ultralytics.nn.modules.block.HGBlock         [1024, 192, 1024, 5, 6, True, True]
  7                  -1  6   2055808  ultralytics.nn.modules.block.HGBlock         [1024, 192, 1024, 5, 6, True, True]
  8                  -1  1     11264  ultralytics.nn.modules.conv.DWConv           [1024, 1024, 3, 2, 1, False]  
  9                  -1  6  25251840  ultralytics.nn.AddModules.Biformer.BiLevelRoutingAttention_nchw[1024, 1024]                  
 10                  -1  1    262656  ultralytics.nn.modules.conv.Conv             [1024, 256, 1, 1, None, 1, 1, False]
 11                  -1  1    789760  ultralytics.nn.modules.transformer.AIFI      [256, 1024, 8]                
 12                  -1  1     66048  ultralytics.nn.modules.conv.Conv             [256, 256, 1, 1]              
 13                  -1  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']          
 14                   7  1    262656  ultralytics.nn.modules.conv.Conv             [1024, 256, 1, 1, None, 1, 1, False]
 15            [-2, -1]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 16                  -1  3   2232320  ultralytics.nn.modules.block.RepC3           [512, 256, 3]                 
 17                  -1  1     66048  ultralytics.nn.modules.conv.Conv             [256, 256, 1, 1]              
 18                  -1  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']          
 19                   3  1    131584  ultralytics.nn.modules.conv.Conv             [512, 256, 1, 1, None, 1, 1, False]
 20            [-2, -1]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 21                  -1  3   2232320  ultralytics.nn.modules.block.RepC3           [512, 256, 3]                 
 22                  -1  1    590336  ultralytics.nn.modules.conv.Conv             [256, 256, 3, 2]              
 23            [-1, 17]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 24                  -1  3   2232320  ultralytics.nn.modules.block.RepC3           [512, 256, 3]                 
 25                  -1  1    590336  ultralytics.nn.modules.conv.Conv             [256, 256, 3, 2]              
 26            [-1, 12]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 27                  -1  3   2232320  ultralytics.nn.modules.block.RepC3           [512, 256, 3]                 
 28        [21, 24, 27]  1   7303907  ultralytics.nn.modules.head.RTDETRDecoder    [1, [256, 256, 256]]          
rtdetr-l-Biformer summary: 650 layers, 51,089,347 parameters, 51,089,347 gradients, 107.5 GFLOPs

rtdetr-l-HGBlock_Biformer ：

rtdetr-l-HGBlock_Biformer summary: 694 layers, 45,434,051 parameters, 45,434,051 gradients, 118.1 GFLOPs 

                   from  n    params  module                                       arguments                     
  0                  -1  1     25248  ultralytics.nn.modules.block.HGStem          [3, 32, 48]                   
  1                  -1  6    155072  ultralytics.nn.modules.block.HGBlock         [48, 48, 128, 3, 6]           
  2                  -1  1      1408  ultralytics.nn.modules.conv.DWConv           [128, 128, 3, 2, 1, False]    
  3                  -1  6    839296  ultralytics.nn.modules.block.HGBlock         [128, 96, 512, 3, 6]          
  4                  -1  1      5632  ultralytics.nn.modules.conv.DWConv           [512, 512, 3, 2, 1, False]    
  5                  -1  6   5904000  ultralytics.nn.AddModules.Biformer.HGBlock_Biformer[512, 192, 1024, 5, 6, True, False]
  6                  -1  6   6264448  ultralytics.nn.AddModules.Biformer.HGBlock_Biformer[1024, 192, 1024, 5, 6, True, True]
  7                  -1  6   6264448  ultralytics.nn.AddModules.Biformer.HGBlock_Biformer[1024, 192, 1024, 5, 6, True, True]
  8                  -1  1     11264  ultralytics.nn.modules.conv.DWConv           [1024, 1024, 3, 2, 1, False]  
  9                  -1  6   6708480  ultralytics.nn.modules.block.HGBlock         [1024, 384, 2048, 5, 6, True, False]
 10                  -1  1    524800  ultralytics.nn.modules.conv.Conv             [2048, 256, 1, 1, None, 1, 1, False]
 11                  -1  1    789760  ultralytics.nn.modules.transformer.AIFI      [256, 1024, 8]                
 12                  -1  1     66048  ultralytics.nn.modules.conv.Conv             [256, 256, 1, 1]              
 13                  -1  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']          
 14                   7  1    262656  ultralytics.nn.modules.conv.Conv             [1024, 256, 1, 1, None, 1, 1, False]
 15            [-2, -1]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 16                  -1  3   2232320  ultralytics.nn.modules.block.RepC3           [512, 256, 3]                 
 17                  -1  1     66048  ultralytics.nn.modules.conv.Conv             [256, 256, 1, 1]              
 18                  -1  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']          
 19                   3  1    131584  ultralytics.nn.modules.conv.Conv             [512, 256, 1, 1, None, 1, 1, False]
 20            [-2, -1]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 21                  -1  3   2232320  ultralytics.nn.modules.block.RepC3           [512, 256, 3]                 
 22                  -1  1    590336  ultralytics.nn.modules.conv.Conv             [256, 256, 3, 2]              
 23            [-1, 17]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 24                  -1  3   2232320  ultralytics.nn.modules.block.RepC3           [512, 256, 3]                 
 25                  -1  1    590336  ultralytics.nn.modules.conv.Conv             [256, 256, 3, 2]              
 26            [-1, 12]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 27                  -1  3   2232320  ultralytics.nn.modules.block.RepC3           [512, 256, 3]                 
 28        [21, 24, 27]  1   7303907  ultralytics.nn.modules.head.RTDETRDecoder    [1, [256, 256, 256]]          
rtdetr-l-HGBlock_Biformer summary: 694 layers, 45,434,051 parameters, 45,434,051 gradients, 118.1 GFLOPs