一、本文介绍

本文记录的是 基于SwinTransformer的RT-DETR目标检测改进方法研究 。本文利用 SwinTransformer 替换 RT-DETR 的骨干网络， Swin Transformer 的作用在于 同时包含层次化特征表示和基于移位窗口的自注意力机制 ，克服了常见 Transformer 模型在视觉任务中面临的 尺度差异 和 分辨率差异 问题。本文将其应用到 RT-DETR 中，使其在保持推理高效的同时，具备大模型的特点。

---

---

二、Swin Transformer模型轻量化设计

Swin Transformer: Hierarchical Vision Transformer using Shifted Windows

2.1 出发点

• 从语言到视觉的迁移挑战
  • 尺度差异 ：语言Transformer以单词为基本处理单元，而视觉元素的尺度变化很大，如在目标检测中，不同物体大小不一。现有基于Transformer的模型中，token都是固定尺度，不适合视觉应用。
  • 分辨率差异 ：图像像素分辨率远高于文本中的单词，对于语义分割等需要在像素级别进行密集预测的视觉任务，Transformer的自注意力计算复杂度与图像大小呈二次方关系，处理高分辨率图像会面临计算困难。

2.2 原理

2.2.1 层次化Transformer

• 通过从小尺寸的图像块开始，逐步在更深的 Transformer 层中合并相邻的块，构建出层次化的特征表示。例如，从最初的 $4\times 4$ 大小的图像块开始，随着网络层数增加，通过合并相邻块来减少 token 数量，同时改变特征维度，形成不同分辨率的特征图。

2.2.2 基于移位窗口的自注意力机制

• 局部窗口内的自注意力计算 ：为了提高效率，在不重叠的局部窗口内计算自注意力。假设每个窗口包含 $M\times M$ 个图像块，当 $M$ 固定时，计算复杂度与图像块数量呈线性关系，而传统的全局自注意力计算复杂度是二次方的。
• 移位窗口分区 ：在连续的自注意力层之间采用移位窗口分区方法。第一个模块使用常规窗口分区策略，下一个模块的窗口配置相对于前一层进行移位。这种方法引入了跨窗口的连接，同时保持了不重叠窗口计算的高效性。

• 相对位置偏差 ：在计算自注意力时，引入相对位置偏差 $B$ ，用于衡量每个头在计算相似度时的相对位置关系。这有助于模型更好地学习视觉元素之间的位置关系，提高模型性能。

2.3 结构

2.3.1 整体架构

• 图像块划分模块 ：首先将输入的RGB图像通过图像块划分模块划分为不重叠的图像块，每个图像块的特征是原始像素RGB值的拼接。例如，采用 $4\times 4$ 的图像块大小，每个图像块的特征维度为 $4\times 4\times 3=48$ 。
• 线性嵌入层 ：对原始值特征应用线性嵌入层，将其投影到任意维度 $C$ 。
• Transformer块 ：包含多个带有改进自注意力计算的Transformer块（Swin Transformer块），这些块保持token数量不变，与线性嵌入层一起构成“Stage 1”。
• 层次化结构构建 ：随着网络加深，通过图像块合并层减少token数量，构建层次化表示。例如，第一个图像块合并层将每组 $2\times 2$ 相邻图像块的特征进行拼接，并应用线性层，输出维度设置为 $2C$ ，然后再应用Swin Transformer块进行特征变换，形成不同分辨率的阶段，如“Stage 2”“Stage 3”“Stage 4”等。

2.3.2 Swin Transformer块

• 由基于移位窗口的多头自注意力（MSA）模块、带有GELU非线性的2层多层感知机（MLP）组成，在每个MSA模块和MLP之前应用层归一化（LN）层，并在每个模块之后应用残差连接。

2.4 优势

• 计算效率高
  • 基于移位窗口的自注意力机制使得计算复杂度与图像大小呈线性关系，相比传统的全局自注意力机制或一些基于滑动窗口的方法，在处理高分辨率图像时计算效率更高。
• 层次化特征表示灵活
  • 能够构建不同分辨率的层次化特征图，适合多种视觉任务。例如，在图像分类、目标检测和语义分割等任务中，可以方便地利用这些特征图进行不同尺度的信息提取和处理。
• 性能优越
  • 在多个视觉任务上取得了优异的性能，如在ImageNet - 1K图像分类任务中达到了较高的准确率，在COCO目标检测和ADE20K语义分割任务上超越了之前的 state - of - the - art方法。具体来说，在COCO测试集上，其box AP和mask AP分别超越之前的最佳结果 $+2.7$ 和 $+2.6$ ；在ADE20K验证集上，mIoU超越之前的最佳结果 $+3.2$ 。

论文： https://arxiv.org/pdf/2103.14030.pdf
源码： https://github.com/microsoft/Swin-Transformer

三、Swin Transformer的实现代码

Swin Transformer模块 的实现代码如下：

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint as checkpoint
import numpy as np
from timm.models.layers import DropPath, to_2tuple, trunc_normal_

__all__ = ['SwinTransformer']

class Mlp(nn.Module):
    """ Multilayer perceptron."""

    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.GELU, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = act_layer()
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x

def window_partition(x, window_size):
    """
    Args:
        x: (B, H, W, C)
        window_size (int): window size
    Returns:
        windows: (num_windows*B, window_size, window_size, C)
    """
    B, H, W, C = x.shape
    x = x.view(B, H // window_size, window_size, W // window_size, window_size, C)
    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
    return windows

def window_reverse(windows, window_size, H, W):
    """
    Args:
        windows: (num_windows*B, window_size, window_size, C)
        window_size (int): Window size
        H (int): Height of image
        W (int): Width of image
    Returns:
        x: (B, H, W, C)
    """
    B = int(windows.shape[0] / (H * W / window_size / window_size))
    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
    return x

class WindowAttention(nn.Module):
    """ Window based multi-head self attention (W-MSA) module with relative position bias.
    It supports both of shifted and non-shifted window.
    Args:
        dim (int): Number of input channels.
        window_size (tuple[int]): The height and width of the window.
        num_heads (int): Number of attention heads.
        qkv_bias (bool, optional):  If True, add a learnable bias to query, key, value. Default: True
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
    """

    def __init__(self, dim, window_size, num_heads, qkv_bias=True, qk_scale=None, attn_drop=0., proj_drop=0.):

        super().__init__()
        self.dim = dim
        self.window_size = window_size  # Wh, Ww
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = qk_scale or head_dim ** -0.5

        # define a parameter table of relative position bias
        self.relative_position_bias_table = nn.Parameter(
            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH

        # get pair-wise relative position index for each token inside the window
        coords_h = torch.arange(self.window_size[0])
        coords_w = torch.arange(self.window_size[1])
        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
        relative_coords[:, :, 1] += self.window_size[1] - 1
        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
        self.register_buffer("relative_position_index", relative_position_index)

        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)

        trunc_normal_(self.relative_position_bias_table, std=.02)
        self.softmax = nn.Softmax(dim=-1)

    def forward(self, x, mask=None):
        """ Forward function.
        Args:
            x: input features with shape of (num_windows*B, N, C)
            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
        """
        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)

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

        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
        attn = attn + relative_position_bias.unsqueeze(0)

        if mask is not None:
            nW = mask.shape[0]
            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
            attn = attn.view(-1, self.num_heads, N, N)
            attn = self.softmax(attn)
        else:
            attn = self.softmax(attn)

        attn = self.attn_drop(attn)

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

class SwinTransformerBlock(nn.Module):
    """ Swin Transformer Block.
    Args:
        dim (int): Number of input channels.
        num_heads (int): Number of attention heads.
        window_size (int): Window size.
        shift_size (int): Shift size for SW-MSA.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
        drop (float, optional): Dropout rate. Default: 0.0
        attn_drop (float, optional): Attention dropout rate. Default: 0.0
        drop_path (float, optional): Stochastic depth rate. Default: 0.0
        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
    """

    def __init__(self, dim, num_heads, window_size=7, shift_size=0,
                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
                 act_layer=nn.GELU, norm_layer=nn.LayerNorm):
        super().__init__()
        self.dim = dim
        self.num_heads = num_heads
        self.window_size = window_size
        self.shift_size = shift_size
        self.mlp_ratio = mlp_ratio
        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"

        self.norm1 = norm_layer(dim)
        self.attn = WindowAttention(
            dim, window_size=to_2tuple(self.window_size), num_heads=num_heads,
            qkv_bias=qkv_bias, qk_scale=qk_scale, attn_drop=attn_drop, proj_drop=drop)

        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        self.norm2 = norm_layer(dim)
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, act_layer=act_layer, drop=drop)

        self.H = None
        self.W = None

    def forward(self, x, mask_matrix):
        """ Forward function.
        Args:
            x: Input feature, tensor size (B, H*W, C).
            H, W: Spatial resolution of the input feature.
            mask_matrix: Attention mask for cyclic shift.
        """
        B, L, C = x.shape
        H, W = self.H, self.W
        assert L == H * W, "input feature has wrong size"

        shortcut = x
        x = self.norm1(x)
        x = x.view(B, H, W, C)

        # pad feature maps to multiples of window size
        pad_l = pad_t = 0
        pad_r = (self.window_size - W % self.window_size) % self.window_size
        pad_b = (self.window_size - H % self.window_size) % self.window_size
        x = F.pad(x, (0, 0, pad_l, pad_r, pad_t, pad_b))
        _, Hp, Wp, _ = x.shape

        # cyclic shift
        if self.shift_size > 0:
            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
            attn_mask = mask_matrix.type(x.dtype)
        else:
            shifted_x = x
            attn_mask = None

        # partition windows
        x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
        x_windows = x_windows.view(-1, self.window_size * self.window_size, C)  # nW*B, window_size*window_size, C

        # W-MSA/SW-MSA
        attn_windows = self.attn(x_windows, mask=attn_mask)  # nW*B, window_size*window_size, C

        # merge windows
        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)
        shifted_x = window_reverse(attn_windows, self.window_size, Hp, Wp)  # B H' W' C

        # reverse cyclic shift
        if self.shift_size > 0:
            x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
        else:
            x = shifted_x

        if pad_r > 0 or pad_b > 0:
            x = x[:, :H, :W, :].contiguous()

        x = x.view(B, H * W, C)

        # FFN
        x = shortcut + self.drop_path(x)
        x = x + self.drop_path(self.mlp(self.norm2(x)))

        return x

class PatchMerging(nn.Module):
    """ Patch Merging Layer
    Args:
        dim (int): Number of input channels.
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
    """

    def __init__(self, dim, norm_layer=nn.LayerNorm):
        super().__init__()
        self.dim = dim
        self.reduction = nn.Linear(4 * dim, 2 * dim, bias=False)
        self.norm = norm_layer(4 * dim)

    def forward(self, x, H, W):
        """ Forward function.
        Args:
            x: Input feature, tensor size (B, H*W, C).
            H, W: Spatial resolution of the input feature.
        """
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size"

        x = x.view(B, H, W, C)

        # padding
        pad_input = (H % 2 == 1) or (W % 2 == 1)
        if pad_input:
            x = F.pad(x, (0, 0, 0, W % 2, 0, H % 2))

        x0 = x[:, 0::2, 0::2, :]  # B H/2 W/2 C
        x1 = x[:, 1::2, 0::2, :]  # B H/2 W/2 C
        x2 = x[:, 0::2, 1::2, :]  # B H/2 W/2 C
        x3 = x[:, 1::2, 1::2, :]  # B H/2 W/2 C
        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C
        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C

        x = self.norm(x)
        x = self.reduction(x)

        return x

class BasicLayer(nn.Module):
    """ A basic Swin Transformer layer for one stage.
    Args:
        dim (int): Number of feature channels
        depth (int): Depths of this stage.
        num_heads (int): Number of attention head.
        window_size (int): Local window size. Default: 7.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
        drop (float, optional): Dropout rate. Default: 0.0
        attn_drop (float, optional): Attention dropout rate. Default: 0.0
        drop_path (float | tuple[float], optional): Stochastic depth rate. Default: 0.0
        norm_layer (nn.Module, optional): Normalization layer. Default: nn.LayerNorm
        downsample (nn.Module | None, optional): Downsample layer at the end of the layer. Default: None
        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
    """

    def __init__(self,
                 dim,
                 depth,
                 num_heads,
                 window_size=7,
                 mlp_ratio=4.,
                 qkv_bias=True,
                 qk_scale=None,
                 drop=0.,
                 attn_drop=0.,
                 drop_path=0.,
                 norm_layer=nn.LayerNorm,
                 downsample=None,
                 use_checkpoint=False):
        super().__init__()
        self.window_size = window_size
        self.shift_size = window_size // 2
        self.depth = depth
        self.use_checkpoint = use_checkpoint

        # build blocks
        self.blocks = nn.ModuleList([
            SwinTransformerBlock(
                dim=dim,
                num_heads=num_heads,
                window_size=window_size,
                shift_size=0 if (i % 2 == 0) else window_size // 2,
                mlp_ratio=mlp_ratio,
                qkv_bias=qkv_bias,
                qk_scale=qk_scale,
                drop=drop,
                attn_drop=attn_drop,
                drop_path=drop_path[i] if isinstance(drop_path, list) else drop_path,
                norm_layer=norm_layer)
            for i in range(depth)])

        # patch merging layer
        if downsample is not None:
            self.downsample = downsample(dim=dim, norm_layer=norm_layer)
        else:
            self.downsample = None

    def forward(self, x, H, W):
        """ Forward function.
        Args:
            x: Input feature, tensor size (B, H*W, C).
            H, W: Spatial resolution of the input feature.
        """

        # calculate attention mask for SW-MSA
        Hp = int(np.ceil(H / self.window_size)) * self.window_size
        Wp = int(np.ceil(W / self.window_size)) * self.window_size
        img_mask = torch.zeros((1, Hp, Wp, 1), device=x.device)  # 1 Hp Wp 1
        h_slices = (slice(0, -self.window_size),
                    slice(-self.window_size, -self.shift_size),
                    slice(-self.shift_size, None))
        w_slices = (slice(0, -self.window_size),
                    slice(-self.window_size, -self.shift_size),
                    slice(-self.shift_size, None))
        cnt = 0
        for h in h_slices:
            for w in w_slices:
                img_mask[:, h, w, :] = cnt
                cnt += 1

        mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))

        for blk in self.blocks:
            blk.H, blk.W = H, W
            if self.use_checkpoint:
                x = checkpoint.checkpoint(blk, x, attn_mask)
            else:
                x = blk(x, attn_mask)
        if self.downsample is not None:
            x_down = self.downsample(x, H, W)
            Wh, Ww = (H + 1) // 2, (W + 1) // 2
            return x, H, W, x_down, Wh, Ww
        else:
            return x, H, W, x, H, W

class PatchEmbed(nn.Module):
    """ Image to Patch Embedding
    Args:
        patch_size (int): Patch token size. Default: 4.
        in_chans (int): Number of input image channels. Default: 3.
        embed_dim (int): Number of linear projection output channels. Default: 96.
        norm_layer (nn.Module, optional): Normalization layer. Default: None
    """

    def __init__(self, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
        super().__init__()
        patch_size = to_2tuple(patch_size)
        self.patch_size = patch_size

        self.in_chans = in_chans
        self.embed_dim = embed_dim

        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size)
        if norm_layer is not None:
            self.norm = norm_layer(embed_dim)
        else:
            self.norm = None

    def forward(self, x):
        """Forward function."""
        # padding
        _, _, H, W = x.size()
        if W % self.patch_size[1] != 0:
            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1]))
        if H % self.patch_size[0] != 0:
            x = F.pad(x, (0, 0, 0, self.patch_size[0] - H % self.patch_size[0]))

        x = self.proj(x)  # B C Wh Ww
        if self.norm is not None:
            Wh, Ww = x.size(2), x.size(3)
            x = x.flatten(2).transpose(1, 2)
            x = self.norm(x)
            x = x.transpose(1, 2).view(-1, self.embed_dim, Wh, Ww)

        return x

class SwinTransformer(nn.Module):
    """ Swin Transformer backbone.
        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
          https://arxiv.org/pdf/2103.14030
    Args:
        pretrain_img_size (int): Input image size for training the pretrained model,
            used in absolute postion embedding. Default 224.
        patch_size (int | tuple(int)): Patch size. Default: 4.
        in_chans (int): Number of input image channels. Default: 3.
        embed_dim (int): Number of linear projection output channels. Default: 96.
        depths (tuple[int]): Depths of each Swin Transformer stage.
        num_heads (tuple[int]): Number of attention head of each stage.
        window_size (int): Window size. Default: 7.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim. Default: 4.
        qkv_bias (bool): If True, add a learnable bias to query, key, value. Default: True
        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set.
        drop_rate (float): Dropout rate.
        attn_drop_rate (float): Attention dropout rate. Default: 0.
        drop_path_rate (float): Stochastic depth rate. Default: 0.2.
        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False.
        patch_norm (bool): If True, add normalization after patch embedding. Default: True.
        out_indices (Sequence[int]): Output from which stages.
        frozen_stages (int): Stages to be frozen (stop grad and set eval mode).
            -1 means not freezing any parameters.
        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False.
    """

    def __init__(self,
                 pretrain_img_size=224,
                 patch_size=4,
                 in_chans=3,
                 embed_dim=96,
                 depths=[2, 2, 6, 2],
                 num_heads=[3, 6, 12, 24],
                 window_size=7,
                 mlp_ratio=4.,
                 qkv_bias=True,
                 qk_scale=None,
                 drop_rate=0.,
                 attn_drop_rate=0.,
                 drop_path_rate=0.2,
                 norm_layer=nn.LayerNorm,
                 ape=False,
                 patch_norm=True,
                 out_indices=(0, 1, 2, 3),
                 frozen_stages=-1,
                 use_checkpoint=False):
        super().__init__()

        self.pretrain_img_size = pretrain_img_size
        self.num_layers = len(depths)
        self.embed_dim = embed_dim
        self.ape = ape
        self.patch_norm = patch_norm
        self.out_indices = out_indices
        self.frozen_stages = frozen_stages

        # split image into non-overlapping patches
        self.patch_embed = PatchEmbed(
            patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim,
            norm_layer=norm_layer if self.patch_norm else None)

        # absolute position embedding
        if self.ape:
            pretrain_img_size = to_2tuple(pretrain_img_size)
            patch_size = to_2tuple(patch_size)
            patches_resolution = [pretrain_img_size[0] // patch_size[0], pretrain_img_size[1] // patch_size[1]]

            self.absolute_pos_embed = nn.Parameter(
                torch.zeros(1, embed_dim, patches_resolution[0], patches_resolution[1]))
            trunc_normal_(self.absolute_pos_embed, std=.02)

        self.pos_drop = nn.Dropout(p=drop_rate)

        # stochastic depth
        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]  # stochastic depth decay rule

        # build layers
        self.layers = nn.ModuleList()
        for i_layer in range(self.num_layers):
            layer = BasicLayer(
                dim=int(embed_dim * 2 ** i_layer),
                depth=depths[i_layer],
                num_heads=num_heads[i_layer],
                window_size=window_size,
                mlp_ratio=mlp_ratio,
                qkv_bias=qkv_bias,
                qk_scale=qk_scale,
                drop=drop_rate,
                attn_drop=attn_drop_rate,
                drop_path=dpr[sum(depths[:i_layer]):sum(depths[:i_layer + 1])],
                norm_layer=norm_layer,
                downsample=PatchMerging if (i_layer < self.num_layers - 1) else None,
                use_checkpoint=use_checkpoint)
            self.layers.append(layer)

        num_features = [int(embed_dim * 2 ** i) for i in range(self.num_layers)]
        self.num_features = num_features

        # add a norm layer for each output
        for i_layer in out_indices:
            layer = norm_layer(num_features[i_layer])
            layer_name = f'norm{i_layer}'
            self.add_module(layer_name, layer)
        self.width_list = [i.size(1) for i in self.forward(torch.randn(1, 3, 640, 640))]

    def forward(self, x):
        """Forward function."""
        x = self.patch_embed(x)

        Wh, Ww = x.size(2), x.size(3)
        if self.ape:
            # interpolate the position embedding to the corresponding size
            absolute_pos_embed = F.interpolate(self.absolute_pos_embed, size=(Wh, Ww), mode='bicubic')
            x = (x + absolute_pos_embed).flatten(2).transpose(1, 2)  # B Wh*Ww C
        else:
            x = x.flatten(2).transpose(1, 2)
        x = self.pos_drop(x)

        outs = []
        for i in range(self.num_layers):
            layer = self.layers[i]
            x_out, H, W, x, Wh, Ww = layer(x, Wh, Ww)

            if i in self.out_indices:
                norm_layer = getattr(self, f'norm{i}')
                x_out = norm_layer(x_out)

                out = x_out.view(-1, H, W, self.num_features[i]).permute(0, 3, 1, 2).contiguous()
                outs.append(out)

        return outs
 

---

四、修改步骤

4.1 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 SwinTransformer.py ，将 第三节 中的代码粘贴到此处

4.2 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 .SwinTransformer import *

4.3 修改三

在 ultralytics/nn/modules/tasks.py 文件中，需要添加各模块类。

① 首先：导入模块

② 在BaseModel类的predict函数中，在如下两处位置中去掉 embed 参数：

③ 在BaseModel类的_predict_once函数，替换如下代码：

    def _predict_once(self, x, profile=False, visualize=False):
        """
        Perform a forward pass through the network.

        Args:
            x (torch.Tensor): The input tensor to the model.
            profile (bool):  Print the computation time of each layer if True, defaults to False.
            visualize (bool): Save the feature maps of the model if True, defaults to False.

        Returns:
            (torch.Tensor): The last output of the model.
        """
        y, dt = [], []  # outputs
        for m in self.model:
            if m.f != -1:  # if not from previous layer
                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers
            if profile:
                self._profile_one_layer(m, x, dt)
            x = m(x)  # run
            y.append(x if m.i in self.save else None)  # save output
            if visualize:
                feature_visualization(x, m.type, m.i, save_dir=visualize)
        return x

④ 将 RTDETRDetectionModel类 中的 predict函数 完整替换：

    def predict(self, x, profile=False, visualize=False, batch=None, augment=False):
        """
        Perform a forward pass through the model.

        Args:
            x (torch.Tensor): The input tensor.
            profile (bool, optional): If True, profile the computation time for each layer. Defaults to False.
            visualize (bool, optional): If True, save feature maps for visualization. Defaults to False.
            batch (dict, optional): Ground truth data for evaluation. Defaults to None.
            augment (bool, optional): If True, perform data augmentation during inference. Defaults to False.

        Returns:
            (torch.Tensor): Model's output tensor.
        """
        y, dt = [], []  # outputs
        for m in self.model[:-1]:  # except the head part
            if m.f != -1:  # if not from previous layer
                x = y[m.f] if isinstance(m.f, int) else [x if j == -1 else y[j] for j in m.f]  # from earlier layers
            if profile:
                self._profile_one_layer(m, x, dt)
            if hasattr(m, 'backbone'):
                x = m(x)
                for _ in range(5 - len(x)):
                    x.insert(0, None)
                for i_idx, i in enumerate(x):
                    if i_idx in self.save:
                        y.append(i)
                    else:
                        y.append(None)
                # for i in x:
                #     if i is not None:
                #         print(i.size())
                x = x[-1]
            else:
                x = m(x)  # run
                y.append(x if m.i in self.save else None)  # save output
            if visualize:
                feature_visualization(x, m.type, m.i, save_dir=visualize)
        head = self.model[-1]
        x = head([y[j] for j in head.f], batch)  # head inference
        return x

⑤ 在 parse_model函数 如下位置替换如下代码:

    if verbose:
        LOGGER.info(f"\n{'':>3}{'from':>20}{'n':>3}{'params':>10}  {'module':<45}{'arguments':<30}")
    ch = [ch]
    layers, save, c2 = [], [], ch[-1]  # layers, savelist, ch out
    is_backbone = False
    for i, (f, n, m, args) in enumerate(d['backbone'] + d['head']):  # from, number, module, args
        try:
            if m == 'node_mode':
                m = d[m]
                if len(args) > 0:
                    if args[0] == 'head_channel':
                        args[0] = int(d[args[0]])
            t = m
            m = getattr(torch.nn, m[3:]) if 'nn.' in m else globals()[m]  # get module
        except:
            pass
        for j, a in enumerate(args):
            if isinstance(a, str):
                with contextlib.suppress(ValueError):
                    try:
                        args[j] = locals()[a] if a in locals() else ast.literal_eval(a)
                    except:
                        args[j] = a

替换后如下：

⑥ 在 parse_model 函数，添加如下代码。

elif m in {
           SwinTransformer
           }:
    m = m(*args)
    c2 = m.width_list  

⑦ 在 parse_model函数 如下位置替换如下代码:

    	if isinstance(c2, list):
            is_backbone = True
            m_ = m
            m_.backbone = True
        else:
            m_ = nn.Sequential(*(m(*args) for _ in range(n))) if n > 1 else m(*args)  # module
            t = str(m)[8:-2].replace('__main__.', '')  # module type
        
        m_.np = sum(x.numel() for x in m_.parameters())  # number params
        m_.i, m_.f, m_.type = i + 4 if is_backbone else i, f, t  # attach index, 'from' index, type
        if verbose:
            LOGGER.info(f'{i:>3}{str(f):>20}{n_:>3}{m_.np:10.0f}  {t:<45}{str(args):<30}')  # print
        save.extend(x % (i + 4 if is_backbone else i) for x in ([f] if isinstance(f, int) else f) if x != -1)  # append to savelist
        layers.append(m_)
        if i == 0:
            ch = []
        if isinstance(c2, list):
            ch.extend(c2)
            for _ in range(5 - len(ch)):
                ch.insert(0, 0)
        else:
            ch.append(c2)
    return nn.Sequential(*layers), sorted(save)

⑧ 在 ultralytics\nn\autobackend.py 文件的 AutoBackend类 中的 forward函数 ，完整替换如下代码：

    def forward(self, im, augment=False, visualize=False):
        """
        Runs inference on the YOLOv8 MultiBackend model.

        Args:
            im (torch.Tensor): The image tensor to perform inference on.
            augment (bool): whether to perform data augmentation during inference, defaults to False
            visualize (bool): whether to visualize the output predictions, defaults to False

        Returns:
            (tuple): Tuple containing the raw output tensor, and processed output for visualization (if visualize=True)
        """
        b, ch, h, w = im.shape  # batch, channel, height, width
        if self.fp16 and im.dtype != torch.float16:
            im = im.half()  # to FP16
        if self.nhwc:
            im = im.permute(0, 2, 3, 1)  # torch BCHW to numpy BHWC shape(1,320,192,3)

        if self.pt or self.nn_module:  # PyTorch
            y = self.model(im, augment=augment, visualize=visualize) if augment or visualize else self.model(im)
        elif self.jit:  # TorchScript
            y = self.model(im)
        elif self.dnn:  # ONNX OpenCV DNN
            im = im.cpu().numpy()  # torch to numpy
            self.net.setInput(im)
            y = self.net.forward()
        elif self.onnx:  # ONNX Runtime
            im = im.cpu().numpy()  # torch to numpy
            y = self.session.run(self.output_names, {self.session.get_inputs()[0].name: im})
        elif self.xml:  # OpenVINO
            im = im.cpu().numpy()  # FP32
            y = list(self.ov_compiled_model(im).values())
        elif self.engine:  # TensorRT
            if self.dynamic and im.shape != self.bindings['images'].shape:
                i = self.model.get_binding_index('images')
                self.context.set_binding_shape(i, im.shape)  # reshape if dynamic
                self.bindings['images'] = self.bindings['images']._replace(shape=im.shape)
                for name in self.output_names:
                    i = self.model.get_binding_index(name)
                    self.bindings[name].data.resize_(tuple(self.context.get_binding_shape(i)))
            s = self.bindings['images'].shape
            assert im.shape == s, f"input size {im.shape} {'>' if self.dynamic else 'not equal to'} max model size {s}"
            self.binding_addrs['images'] = int(im.data_ptr())
            self.context.execute_v2(list(self.binding_addrs.values()))
            y = [self.bindings[x].data for x in sorted(self.output_names)]
        elif self.coreml:  # CoreML
            im = im[0].cpu().numpy()
            im_pil = Image.fromarray((im * 255).astype('uint8'))
            # im = im.resize((192, 320), Image.BILINEAR)
            y = self.model.predict({'image': im_pil})  # coordinates are xywh normalized
            if 'confidence' in y:
                raise TypeError('Ultralytics only supports inference of non-pipelined CoreML models exported with '
                                f"'nms=False', but 'model={w}' has an NMS pipeline created by an 'nms=True' export.")
                # TODO: CoreML NMS inference handling
                # from ultralytics.utils.ops import xywh2xyxy
                # box = xywh2xyxy(y['coordinates'] * [[w, h, w, h]])  # xyxy pixels
                # conf, cls = y['confidence'].max(1), y['confidence'].argmax(1).astype(np.float32)
                # y = np.concatenate((box, conf.reshape(-1, 1), cls.reshape(-1, 1)), 1)
            elif len(y) == 1:  # classification model
                y = list(y.values())
            elif len(y) == 2:  # segmentation model
                y = list(reversed(y.values()))  # reversed for segmentation models (pred, proto)
        elif self.paddle:  # PaddlePaddle
            im = im.cpu().numpy().astype(np.float32)
            self.input_handle.copy_from_cpu(im)
            self.predictor.run()
            y = [self.predictor.get_output_handle(x).copy_to_cpu() for x in self.output_names]
        elif self.ncnn:  # ncnn
            mat_in = self.pyncnn.Mat(im[0].cpu().numpy())
            ex = self.net.create_extractor()
            input_names, output_names = self.net.input_names(), self.net.output_names()
            ex.input(input_names[0], mat_in)
            y = []
            for output_name in output_names:
                mat_out = self.pyncnn.Mat()
                ex.extract(output_name, mat_out)
                y.append(np.array(mat_out)[None])
        elif self.triton:  # NVIDIA Triton Inference Server
            im = im.cpu().numpy()  # torch to numpy
            y = self.model(im)
        else:  # TensorFlow (SavedModel, GraphDef, Lite, Edge TPU)
            im = im.cpu().numpy()
            if self.saved_model:  # SavedModel
                y = self.model(im, training=False) if self.keras else self.model(im)
                if not isinstance(y, list):
                    y = [y]
            elif self.pb:  # GraphDef
                y = self.frozen_func(x=self.tf.constant(im))
                if len(y) == 2 and len(self.names) == 999:  # segments and names not defined
                    ip, ib = (0, 1) if len(y[0].shape) == 4 else (1, 0)  # index of protos, boxes
                    nc = y[ib].shape[1] - y[ip].shape[3] - 4  # y = (1, 160, 160, 32), (1, 116, 8400)
                    self.names = {i: f'class{i}' for i in range(nc)}
            else:  # Lite or Edge TPU
                details = self.input_details[0]
                integer = details['dtype'] in (np.int8, np.int16)  # is TFLite quantized int8 or int16 model
                if integer:
                    scale, zero_point = details['quantization']
                    im = (im / scale + zero_point).astype(details['dtype'])  # de-scale
                self.interpreter.set_tensor(details['index'], im)
                self.interpreter.invoke()
                y = []
                for output in self.output_details:
                    x = self.interpreter.get_tensor(output['index'])
                    if integer:
                        scale, zero_point = output['quantization']
                        x = (x.astype(np.float32) - zero_point) * scale  # re-scale
                    if x.ndim > 2:  # if task is not classification
                        # Denormalize xywh by image size. See https://github.com/ultralytics/ultralytics/pull/1695
                        # xywh are normalized in TFLite/EdgeTPU to mitigate quantization error of integer models
                        x[:, [0, 2]] *= w
                        x[:, [1, 3]] *= h
                    y.append(x)
            # TF segment fixes: export is reversed vs ONNX export and protos are transposed
            if len(y) == 2:  # segment with (det, proto) output order reversed
                if len(y[1].shape) != 4:
                    y = list(reversed(y))  # should be y = (1, 116, 8400), (1, 160, 160, 32)
                y[1] = np.transpose(y[1], (0, 3, 1, 2))  # should be y = (1, 116, 8400), (1, 32, 160, 160)
            y = [x if isinstance(x, np.ndarray) else x.numpy() for x in y]

        # for x in y:
        #     print(type(x), len(x)) if isinstance(x, (list, tuple)) else print(type(x), x.shape)  # debug shapes
        if isinstance(y, (list, tuple)):
            return self.from_numpy(y[0]) if len(y) == 1 else [self.from_numpy(x) for x in y]
        else:
            return self.from_numpy(y)

至此就修改完成了，可以配置模型开始训练了

---

五、yaml模型文件

5.1 模型改进⭐

在代码配置完成后，配置模型的YAML文件。

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-SwinTransformer.yaml 。

将 rtdetr-l.yaml 中的内容复制到 rtdetr-SwinTransformer.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 替换成 SwinTransformer 。

# 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, SwinTransformer, []]  # 4

head:
  - [-1, 1, Conv, [256, 1, 1, None, 1, 1, False]]  # 5 input_proj.2
  - [-1, 1, AIFI, [1024, 8]] # 6
  - [-1, 1, Conv, [256, 1, 1]]  # 7, Y5, lateral_convs.0

  - [-1, 1, nn.Upsample, [None, 2, 'nearest']] # 8
  - [3, 1, Conv, [256, 1, 1, None, 1, 1, False]]  # 9 input_proj.1
  - [[-2, -1], 1, Concat, [1]] # 10
  - [-1, 3, RepC3, [256]]  # 11, fpn_blocks.0
  - [-1, 1, Conv, [256, 1, 1]]   # 12, Y4, lateral_convs.1

  - [-1, 1, nn.Upsample, [None, 2, 'nearest']] # 13
  - [2, 1, Conv, [256, 1, 1, None, 1, 1, False]]  # 14 input_proj.0
  - [[-2, -1], 1, Concat, [1]]  # 15 cat backbone P4
  - [-1, 3, RepC3, [256]]    # X3 (16), fpn_blocks.1

  - [-1, 1, Conv, [256, 3, 2]]   # 17, downsample_convs.0
  - [[-1, 12], 1, Concat, [1]]  # 18 cat Y4
  - [-1, 3, RepC3, [256]]    # F4 (19), pan_blocks.0

  - [-1, 1, Conv, [256, 3, 2]]   # 20, downsample_convs.1
  - [[-1, 7], 1, Concat, [1]]  # 21 cat Y5
  - [-1, 3, RepC3, [256]]    # F5 (22), pan_blocks.1

  - [[16, 19, 22], 1, RTDETRDecoder, [nc]]  # Detect(P3, P4, P5)

---

六、成功运行结果

分别打印网络模型可以看到 SwinTransformer模块 已经加入到模型中，并可以进行训练了。

rtdetr-SwinTransformer ：

rtdetr-SwinTransformer summary: 577 layers, 46,202,013 parameters, 46,202,013 gradients, 134.6 GFLOPs

                   from  n    params  module                                       arguments                     
  0                  -1  1  27520698  SwinTransformer                              []                            
  1                  -1  1    197120  ultralytics.nn.modules.conv.Conv             [768, 256, 1, 1, None, 1, 1, False]
  2                  -1  1    789760  ultralytics.nn.modules.transformer.AIFI      [256, 1024, 8]                
  3                  -1  1     66048  ultralytics.nn.modules.conv.Conv             [256, 256, 1, 1]              
  4                  -1  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']          
  5                   3  1     98816  ultralytics.nn.modules.conv.Conv             [384, 256, 1, 1, None, 1, 1, False]
  6            [-2, -1]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
  7                  -1  3   2232320  ultralytics.nn.modules.block.RepC3           [512, 256, 3]                 
  8                  -1  1     66048  ultralytics.nn.modules.conv.Conv             [256, 256, 1, 1]              
  9                  -1  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']          
 10                   2  1     49664  ultralytics.nn.modules.conv.Conv             [192, 256, 1, 1, None, 1, 1, False]
 11            [-2, -1]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 12                  -1  3   2232320  ultralytics.nn.modules.block.RepC3           [512, 256, 3]                 
 13                  -1  1    590336  ultralytics.nn.modules.conv.Conv             [256, 256, 3, 2]              
 14            [-1, 12]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 15                  -1  3   2232320  ultralytics.nn.modules.block.RepC3           [512, 256, 3]                 
 16                  -1  1    590336  ultralytics.nn.modules.conv.Conv             [256, 256, 3, 2]              
 17             [-1, 7]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 18                  -1  3   2232320  ultralytics.nn.modules.block.RepC3           [512, 256, 3]                 
 19        [16, 19, 22]  1   7303907  ultralytics.nn.modules.head.RTDETRDecoder    [1, [256, 256, 256]]          
rtdetr-SwinTransformer summary: 577 layers, 46,202,013 parameters, 46,202,013 gradients, 134.6 GFLOPs