.\YOLO-World\yolo_world\models\detectors\__init__.py

# 版权声明，版权归腾讯公司所有
# 导入yolo_world模块中的YOLOWorldDetector类
from .yolo_world import YOLOWorldDetector

# 导出YOLOWorldDetector类，供外部使用
__all__ = ['YOLOWorldDetector']

.\YOLO-World\yolo_world\models\layers\yolo_bricks.py

# 版权声明，版权归腾讯公司所有
from typing import List  # 导入 List 类型

import torch  # 导入 torch 库
import torch.nn as nn  # 导入 torch.nn 模块
from torch import Tensor  # 导入 Tensor 类型
import torch.nn.functional as F  # 导入 torch.nn.functional 模块
from mmcv.cnn import ConvModule, DepthwiseSeparableConvModule, Linear  # 导入 mmcv.cnn 模块中的 ConvModule、DepthwiseSeparableConvModule、Linear 类
from mmdet.utils import ConfigType, OptConfigType, OptMultiConfig  # 导入 mmdet.utils 模块中的 ConfigType、OptConfigType、OptMultiConfig 类
from mmengine.model import BaseModule  # 导入 mmengine.model 模块中的 BaseModule 类
from mmyolo.registry import MODELS  # 导入 mmyolo.registry 模块中的 MODELS 注册器
from mmyolo.models.layers import CSPLayerWithTwoConv  # 导入 mmyolo.models.layers 模块中的 CSPLayerWithTwoConv 类

@MODELS.register_module()  # 使用 MODELS 注册器注册该类
class MaxSigmoidAttnBlock(BaseModule):  # 定义 MaxSigmoidAttnBlock 类，继承自 BaseModule
    """Max Sigmoid attention block."""  # 类的简要描述
    # 初始化函数，定义了模型的各种参数
    def __init__(self,
                 in_channels: int,
                 out_channels: int,
                 guide_channels: int,
                 embed_channels: int,
                 kernel_size: int = 3,
                 padding: int = 1,
                 num_heads: int = 1,
                 use_depthwise: bool = False,
                 with_scale: bool = False,
                 conv_cfg: OptConfigType = None,
                 norm_cfg: ConfigType = dict(type='BN',
                                             momentum=0.03,
                                             eps=0.001),
                 init_cfg: OptMultiConfig = None) -> None:
        # 调用父类的初始化函数
        super().__init__(init_cfg=init_cfg)
        # 根据是否使用深度可分离卷积选择不同的卷积模块
        conv = DepthwiseSeparableConvModule if use_depthwise else ConvModule

        # 检查输出通道数和嵌入通道数是否能被头数整除
        assert (out_channels % num_heads == 0 and
                embed_channels % num_heads == 0), \
            'out_channels and embed_channels should be divisible by num_heads.'
        # 设置头数和每个头的通道数
        self.num_heads = num_heads
        self.head_channels = out_channels // num_heads

        # 如果嵌入通道数不等于输入通道数，则定义一个卷积模块用于嵌入
        self.embed_conv = ConvModule(
            in_channels,
            embed_channels,
            1,
            conv_cfg=conv_cfg,
            norm_cfg=norm_cfg,
            act_cfg=None) if embed_channels != in_channels else None
        # 定义一个全连接层用于引导通道到嵌入通道的映射
        self.guide_fc = Linear(guide_channels, embed_channels)
        # 定义一个偏置参数
        self.bias = nn.Parameter(torch.zeros(num_heads))
        # 如果设置了缩放参数，则定义一个缩放参数
        if with_scale:
            self.scale = nn.Parameter(torch.ones(1, num_heads, 1, 1))
        else:
            self.scale = 1.0

        # 定义一个卷积模块用于将输入通道映射到输出通道
        self.project_conv = conv(in_channels,
                                 out_channels,
                                 kernel_size,
                                 stride=1,
                                 padding=padding,
                                 conv_cfg=conv_cfg,
                                 norm_cfg=norm_cfg,
                                 act_cfg=None)
    # 定义一个前向传播函数，接受输入张量 x 和引导张量 guide，返回处理后的张量
    def forward(self, x: Tensor, guide: Tensor) -> Tensor:
        """Forward process."""
        # 获取输入张量 x 的形状信息
        B, _, H, W = x.shape

        # 使用引导张量 guide 经过全连接层处理
        guide = self.guide_fc(guide)
        # 重新调整 guide 的形状，将其分成多个头部
        guide = guide.reshape(B, -1, self.num_heads, self.head_channels)
        # 如果存在嵌入卷积层，对输入张量 x 进行卷积操作，否则直接使用 x
        embed = self.embed_conv(x) if self.embed_conv is not None else x
        # 调整嵌入结果的形状，分成多个头部
        embed = embed.reshape(B, self.num_heads, self.head_channels, H, W)

        # 使用 einsum 函数计算注意力权重
        attn_weight = torch.einsum('bmchw,bnmc->bmhwn', embed, guide)
        # 在最后一个维度上取最大值
        attn_weight = attn_weight.max(dim=-1)[0]
        # 归一化注意力权重
        attn_weight = attn_weight / (self.head_channels**0.5)
        # 添加偏置项
        attn_weight = attn_weight + self.bias[None, :, None, None]
        # 对注意力权重进行 sigmoid 激活函数处理，并乘以缩放因子
        attn_weight = attn_weight.sigmoid() * self.scale

        # 对输入张量进行投影卷积
        x = self.project_conv(x)
        # 调整投影结果的形状，分成多个头部
        x = x.reshape(B, self.num_heads, -1, H, W)
        # 将投影结果与注意力权重相乘
        x = x * attn_weight.unsqueeze(2)
        # 调整结果的形状
        x = x.reshape(B, -1, H, W)
        # 返回处理后的张量
        return x
# 使用 @MODELS.register_module() 装饰器注册 MaxSigmoidCSPLayerWithTwoConv 类
@MODELS.register_module()
# 定义 MaxSigmoidCSPLayerWithTwoConv 类，继承自 CSPLayerWithTwoConv 类
class MaxSigmoidCSPLayerWithTwoConv(CSPLayerWithTwoConv):
    """Sigmoid-attention based CSP layer with two convolution layers."""
    # 类的说明文档，描述该类是基于 Sigmoid-attention 的 CSP 层，包含两个卷积层
    # 初始化函数，定义了网络结构的各种参数
    def __init__(
            self,
            in_channels: int,  # 输入通道数
            out_channels: int,  # 输出通道数
            guide_channels: int,  # 引导通道数
            embed_channels: int,  # 嵌入通道数
            num_heads: int = 1,  # 多头注意力机制的头数，默认为1
            expand_ratio: float = 0.5,  # 扩展比例，默认为0.5
            num_blocks: int = 1,  # 块的数量，默认为1
            with_scale: bool = False,  # 是否使用缩放，默认为False
            add_identity: bool = True,  # 是否添加身份连接，默认为True
            conv_cfg: OptConfigType = None,  # 卷积配置，默认为None
            norm_cfg: ConfigType = dict(type='BN', momentum=0.03, eps=0.001),  # 归一化配置，默认为BatchNorm
            act_cfg: ConfigType = dict(type='SiLU', inplace=True),  # 激活函数配置，默认为SiLU
            init_cfg: OptMultiConfig = None) -> None:  # 初始化配置，默认为None，返回None
        # 调用父类的初始化函数，传入各种参数
        super().__init__(in_channels=in_channels,
                         out_channels=out_channels,
                         expand_ratio=expand_ratio,
                         num_blocks=num_blocks,
                         add_identity=add_identity,
                         conv_cfg=conv_cfg,
                         norm_cfg=norm_cfg,
                         act_cfg=act_cfg,
                         init_cfg=init_cfg)

        # 定义最终的卷积层，输入通道数为(3 + num_blocks) * self.mid_channels，输出通道数为out_channels
        self.final_conv = ConvModule((3 + num_blocks) * self.mid_channels,
                                     out_channels,
                                     1,
                                     conv_cfg=conv_cfg,
                                     norm_cfg=norm_cfg,
                                     act_cfg=act_cfg)

        # 定义注意力块，输入通道数为self.mid_channels，输出通道数为self.mid_channels
        self.attn_block = MaxSigmoidAttnBlock(self.mid_channels,
                                              self.mid_channels,
                                              guide_channels=guide_channels,
                                              embed_channels=embed_channels,
                                              num_heads=num_heads,
                                              with_scale=with_scale,
                                              conv_cfg=conv_cfg,
                                              norm_cfg=norm_cfg)
    # 定义一个前向传播函数，接受输入张量 x 和引导张量 guide，返回处理后的张量
    def forward(self, x: Tensor, guide: Tensor) -> Tensor:
        """Forward process."""
        # 使用主要卷积层处理输入张量 x
        x_main = self.main_conv(x)
        # 将处理后的张量按照通道数分割成两部分
        x_main = list(x_main.split((self.mid_channels, self.mid_channels), 1))
        # 对每个分割后的部分依次应用不同的块
        x_main.extend(blocks(x_main[-1]) for blocks in self.blocks)
        # 将最后一个处理后的部分与引导张量 guide 一起传入注意力块
        x_main.append(self.attn_block(x_main[-1], guide))
        # 将所有处理后的部分拼接在一起，然后通过最终卷积层处理得到最终输出
        return self.final_conv(torch.cat(x_main, 1))
# 注册 ImagePoolingAttentionModule 类到 MODELS 模块
@MODELS.register_module()
class ImagePoolingAttentionModule(nn.Module):
    # 初始化函数，接受多个参数
    def __init__(self,
                 image_channels: List[int],
                 text_channels: int,
                 embed_channels: int,
                 with_scale: bool = False,
                 num_feats: int = 3,
                 num_heads: int = 8,
                 pool_size: int = 3):
        # 调用父类的初始化函数
        super().__init__()

        # 初始化各个属性
        self.text_channels = text_channels
        self.embed_channels = embed_channels
        self.num_heads = num_heads
        self.num_feats = num_feats
        self.head_channels = embed_channels // num_heads
        self.pool_size = pool_size

        # 根据 with_scale 参数决定是否添加可学习的缩放参数
        if with_scale:
            self.scale = nn.Parameter(torch.tensor([0.]), requires_grad=True)
        else:
            self.scale = 1.0
        # 创建投影层，将输入的图像通道数映射到嵌入通道数
        self.projections = nn.ModuleList([
            ConvModule(in_channels, embed_channels, 1, act_cfg=None)
            for in_channels in image_channels
        ])
        # 创建查询、键、值和投影层
        self.query = nn.Sequential(nn.LayerNorm(text_channels),
                                   Linear(text_channels, embed_channels))
        self.key = nn.Sequential(nn.LayerNorm(embed_channels),
                                 Linear(embed_channels, embed_channels))
        self.value = nn.Sequential(nn.LayerNorm(embed_channels),
                                   Linear(embed_channels, embed_channels))
        self.proj = Linear(embed_channels, text_channels)

        # 创建图像池化层，用于对图像特征进行池化
        self.image_pools = nn.ModuleList([
            nn.AdaptiveMaxPool2d((pool_size, pool_size))
            for _ in range(num_feats)
        ])
    # 前向传播函数，接收文本特征和图像特征作为输入
    def forward(self, text_features, image_features):
        # 获取 batch size
        B = image_features[0].shape[0]
        # 断言图像特征列表长度等于预定义的特征数量
        assert len(image_features) == self.num_feats
        # 计算每个图像特征的像素块数量
        num_patches = self.pool_size**2
        # 对每个图像特征进行投影和池化操作，然后将结果拼接在一起
        mlvl_image_features = [
            pool(proj(x)).view(B, -1, num_patches)
            for (x, proj, pool
                 ) in zip(image_features, self.projections, self.image_pools)
        ]
        # 将拼接后的图像特征进行维度转置
        mlvl_image_features = torch.cat(mlvl_image_features,
                                        dim=-1).transpose(1, 2)
        # 对文本特征进行查询操作
        q = self.query(text_features)
        # 对图像特征进行键值对操作
        k = self.key(mlvl_image_features)
        v = self.value(mlvl_image_features)

        # 将查询、键、值进行维度重塑
        q = q.reshape(B, -1, self.num_heads, self.head_channels)
        k = k.reshape(B, -1, self.num_heads, self.head_channels)
        v = v.reshape(B, -1, self.num_heads, self.head_channels)

        # 计算注意力权重
        attn_weight = torch.einsum('bnmc,bkmc->bmnk', q, k)
        # 缩放注意力权重
        attn_weight = attn_weight / (self.head_channels**0.5)
        # 对注意力权重进行 softmax 操作
        attn_weight = F.softmax(attn_weight, dim=-1)

        # 根据注意力权重计算加权值
        x = torch.einsum('bmnk,bkmc->bnmc', attn_weight, v)
        # 将加权值进行投影操作
        x = self.proj(x.reshape(B, -1, self.embed_channels))
        # 返回最终结果，加上文本特征并乘以缩放因子
        return x * self.scale + text_features
# 注册模块为VanillaSigmoidBlock，表示使用Sigmoid激活函数的注意力块
@MODELS.register_module()
class VanillaSigmoidBlock(BaseModule):
    """Sigmoid attention block."""
    def __init__(self,
                 in_channels: int,
                 out_channels: int,
                 guide_channels: int,
                 embed_channels: int,
                 kernel_size: int = 3,
                 padding: int = 1,
                 num_heads: int = 1,
                 use_depthwise: bool = False,
                 with_scale: bool = False,
                 conv_cfg: OptConfigType = None,
                 norm_cfg: ConfigType = dict(type='BN',
                                             momentum=0.03,
                                             eps=0.001),
                 init_cfg: OptMultiConfig = None) -> None:
        super().__init__(init_cfg=init_cfg)
        # 根据是否使用深度可分离卷积选择不同的卷积模块
        conv = DepthwiseSeparableConvModule if use_depthwise else ConvModule

        # 确保输出通道数和嵌入通道数能够被头数整除
        assert (out_channels % num_heads == 0 and
                embed_channels % num_heads == 0), \
            'out_channels and embed_channels should be divisible by num_heads.'
        self.num_heads = num_heads
        self.head_channels = out_channels // num_heads

        # 定义投影卷积层
        self.project_conv = conv(in_channels,
                                 out_channels,
                                 kernel_size,
                                 stride=1,
                                 padding=padding,
                                 conv_cfg=conv_cfg,
                                 norm_cfg=norm_cfg,
                                 act_cfg=None)

    def forward(self, x: Tensor, guide: Tensor) -> Tensor:
        """Forward process."""
        # 进行投影卷积
        x = self.project_conv(x)
        # 使用Sigmoid激活函数进行注意力加权
        x = x * x.sigmoid()
        return x


# 注册模块为EfficientCSPLayerWithTwoConv，表示使用两个卷积层的CSP层，基于Sigmoid注意力机制
@MODELS.register_module()
class EfficientCSPLayerWithTwoConv(CSPLayerWithTwoConv):
    """Sigmoid-attention based CSP layer with two convolution layers."""
    # 初始化函数，定义了一个自定义的神经网络模块
    def __init__(
            self,
            in_channels: int,  # 输入通道数
            out_channels: int,  # 输出通道数
            guide_channels: int,  # 引导通道数
            embed_channels: int,  # 嵌入通道数
            num_heads: int = 1,  # 多头注意力机制的头数，默认为1
            expand_ratio: float = 0.5,  # 扩展比例，默认为0.5
            num_blocks: int = 1,  # 块的数量，默认为1
            with_scale: bool = False,  # 是否使用缩放，默认为False
            add_identity: bool = True,  # 是否添加身份映射，默认为True
            conv_cfg: OptConfigType = None,  # 卷积配置，默认为None
            norm_cfg: ConfigType = dict(type='BN', momentum=0.03, eps=0.001),  # 归一化配置，默认为BatchNorm
            act_cfg: ConfigType = dict(type='SiLU', inplace=True),  # 激活函数配置，默认为SiLU
            init_cfg: OptMultiConfig = None) -> None:  # 初始化配置，默认为None
        # 调用父类的初始化函数
        super().__init__(in_channels=in_channels,
                         out_channels=out_channels,
                         expand_ratio=expand_ratio,
                         num_blocks=num_blocks,
                         add_identity=add_identity,
                         conv_cfg=conv_cfg,
                         norm_cfg=norm_cfg,
                         act_cfg=act_cfg,
                         init_cfg=init_cfg)

        # 定义最终的卷积层
        self.final_conv = ConvModule((3 + num_blocks) * self.mid_channels,
                                     out_channels,
                                     1,
                                     conv_cfg=conv_cfg,
                                     norm_cfg=norm_cfg,
                                     act_cfg=act_cfg)

        # 定义注意力块
        self.attn_block = VanillaSigmoidBlock(self.mid_channels,
                                              self.mid_channels,
                                              guide_channels=guide_channels,
                                              embed_channels=embed_channels,
                                              num_heads=num_heads,
                                              with_scale=with_scale,
                                              conv_cfg=conv_cfg,
                                              norm_cfg=norm_cfg)
    # 定义一个前向传播函数，接受输入张量 x 和引导张量 guide，返回处理后的张量
    def forward(self, x: Tensor, guide: Tensor) -> Tensor:
        """Forward process."""
        # 使用主要卷积层处理输入张量 x
        x_main = self.main_conv(x)
        # 将处理后的张量按照通道数分割成两部分
        x_main = list(x_main.split((self.mid_channels, self.mid_channels), 1))
        # 对每个块中的处理函数对最后一个处理后的张量进行处理，并将结果添加到 x_main 中
        x_main.extend(blocks(x_main[-1]) for blocks in self.blocks)
        # 将最后一个处理后的张量和引导张量传入注意力块中进行处理，并将结果添加到 x_main 中
        x_main.append(self.attn_block(x_main[-1], guide))
        # 将所有处理后的张量拼接在一起，并传入最终卷积层进行处理，返回结果
        return self.final_conv(torch.cat(x_main, 1))

.\YOLO-World\yolo_world\models\layers\__init__.py

# 版权声明，版权归腾讯公司所有
# 基于 CSPLayers 的 PAFPN 的基本模块

# 导入 yolo_bricks 模块中的相关类
from .yolo_bricks import (
    CSPLayerWithTwoConv,
    MaxSigmoidAttnBlock,
    MaxSigmoidCSPLayerWithTwoConv,
    ImagePoolingAttentionModule,
    )

# 导出给外部使用的类列表
__all__ = ['CSPLayerWithTwoConv',
           'MaxSigmoidAttnBlock',
           'MaxSigmoidCSPLayerWithTwoConv',
           'ImagePoolingAttentionModule']

.\YOLO-World\yolo_world\models\losses\dynamic_loss.py

# 导入必要的库
from typing import Optional
import torch
import torch.nn as nn
from torch import Tensor
from mmdet.models.losses.mse_loss import mse_loss
from mmyolo.registry import MODELS

# 注册模型类为CoVMSELoss
@MODELS.register_module()
class CoVMSELoss(nn.Module):

    def __init__(self,
                 dim: int = 0,
                 reduction: str = 'mean',
                 loss_weight: float = 1.0,
                 eps: float = 1e-6) -> None:
        super().__init__()
        # 初始化参数
        self.dim = dim
        self.reduction = reduction
        self.loss_weight = loss_weight
        self.eps = eps

    def forward(self,
                pred: Tensor,
                weight: Optional[Tensor] = None,
                avg_factor: Optional[int] = None,
                reduction_override: Optional[str] = None) -> Tensor:
        """Forward function of loss."""
        # 确保重写的减少参数在合法范围内
        assert reduction_override in (None, 'none', 'mean', 'sum')
        # 根据重写的减少参数或者默认减少参数来确定减少方式
        reduction = (
            reduction_override if reduction_override else self.reduction)
        # 计算协方差
        cov = pred.std(self.dim) / pred.mean(self.dim).clamp(min=self.eps)
        # 创建目标张量
        target = torch.zeros_like(cov)
        # 计算损失
        loss = self.loss_weight * mse_loss(
            cov, target, weight, reduction=reduction, avg_factor=avg_factor)
        return loss

.\YOLO-World\yolo_world\models\losses\__init__.py

# 版权声明，版权归腾讯公司所有
# 导入动态损失模块中的CoVMSELoss类
from .dynamic_loss import CoVMSELoss

# 导出CoVMSELoss类，供外部使用
__all__ = ['CoVMSELoss']

.\YOLO-World\yolo_world\models\necks\yolo_world_pafpn.py

# 导入必要的库
import copy
from typing import List, Union
import torch
import torch.nn as nn
from torch import Tensor
from mmdet.utils import ConfigType, OptMultiConfig

# 导入自定义的模型注册器和工具函数
from mmyolo.registry import MODELS
from mmyolo.models.utils import make_divisible, make_round
from mmyolo.models.necks.yolov8_pafpn import YOLOv8PAFPN

# 注册YOLOWorldPAFPN类为模型
@MODELS.register_module()
class YOLOWorldPAFPN(YOLOv8PAFPN):
    """Path Aggregation Network used in YOLO World
    Following YOLOv8 PAFPN, including text to image fusion
    """
    # 初始化函数，定义模型结构和参数
    def __init__(self,
                 in_channels: List[int],
                 out_channels: Union[List[int], int],
                 guide_channels: int,
                 embed_channels: List[int],
                 num_heads: List[int],
                 deepen_factor: float = 1.0,
                 widen_factor: float = 1.0,
                 num_csp_blocks: int = 3,
                 freeze_all: bool = False,
                 block_cfg: ConfigType = dict(type='CSPLayerWithTwoConv'),
                 norm_cfg: ConfigType = dict(type='BN',
                                             momentum=0.03,
                                             eps=0.001),
                 act_cfg: ConfigType = dict(type='SiLU', inplace=True),
                 init_cfg: OptMultiConfig = None) -> None:
        # 设置引导通道数、嵌入通道数和头数
        self.guide_channels = guide_channels
        self.embed_channels = embed_channels
        self.num_heads = num_heads
        self.block_cfg = block_cfg
        # 调用父类的初始化函数，传入参数
        super().__init__(in_channels=in_channels,
                         out_channels=out_channels,
                         deepen_factor=deepen_factor,
                         widen_factor=widen_factor,
                         num_csp_blocks=num_csp_blocks,
                         freeze_all=freeze_all,
                         norm_cfg=norm_cfg,
                         act_cfg=act_cfg,
                         init_cfg=init_cfg)
    # 构建自顶向下的层
    def build_top_down_layer(self, idx: int) -> nn.Module:
        """build top down layer.

        Args:
            idx (int): layer idx.

        Returns:
            nn.Module: The top down layer.
        """
        # 深拷贝块配置
        block_cfg = copy.deepcopy(self.block_cfg)
        # 更新块配置参数
        block_cfg.update(
            dict(in_channels=make_divisible(
                (self.in_channels[idx - 1] + self.in_channels[idx]),
                self.widen_factor),
                 out_channels=make_divisible(self.out_channels[idx - 1],
                                             self.widen_factor),
                 guide_channels=self.guide_channels,
                 embed_channels=make_round(self.embed_channels[idx - 1],
                                           self.widen_factor),
                 num_heads=make_round(self.num_heads[idx - 1],
                                      self.widen_factor),
                 num_blocks=make_round(self.num_csp_blocks,
                                       self.deepen_factor),
                 add_identity=False,
                 norm_cfg=self.norm_cfg,
                 act_cfg=self.act_cfg))
        # 构建模型
        return MODELS.build(block_cfg)
    # 构建底部向上的层
    def build_bottom_up_layer(self, idx: int) -> nn.Module:
        """build bottom up layer.

        Args:
            idx (int): layer idx.

        Returns:
            nn.Module: The bottom up layer.
        """
        # 深拷贝块配置
        block_cfg = copy.deepcopy(self.block_cfg)
        # 更新块配置
        block_cfg.update(
            dict(in_channels=make_divisible(
                (self.out_channels[idx] + self.out_channels[idx + 1]),
                self.widen_factor),
                 out_channels=make_divisible(self.out_channels[idx + 1],
                                             self.widen_factor),
                 guide_channels=self.guide_channels,
                 embed_channels=make_round(self.embed_channels[idx + 1],
                                           self.widen_factor),
                 num_heads=make_round(self.num_heads[idx + 1],
                                      self.widen_factor),
                 num_blocks=make_round(self.num_csp_blocks,
                                       self.deepen_factor),
                 add_identity=False,
                 norm_cfg=self.norm_cfg,
                 act_cfg=self.act_cfg))
        # 构建模型
        return MODELS.build(block_cfg)
    # 定义前向传播函数，接受多层级的图像特征和文本特征作为输入，返回元组
    def forward(self, img_feats: List[Tensor], txt_feats: Tensor) -> tuple:
        """Forward function.
        including multi-level image features, text features: BxLxD
        """
        # 断言图像特征的数量与输入通道数相同
        assert len(img_feats) == len(self.in_channels)
        
        # 减少层级
        reduce_outs = []
        for idx in range(len(self.in_channels)):
            reduce_outs.append(self.reduce_layers[idx](img_feats[idx]))

        # 自顶向下路径
        inner_outs = [reduce_outs[-1]]
        for idx in range(len(self.in_channels) - 1, 0, -1):
            feat_high = inner_outs[0]
            feat_low = reduce_outs[idx - 1]
            upsample_feat = self.upsample_layers[len(self.in_channels) - 1 - idx](feat_high)
            if self.upsample_feats_cat_first:
                top_down_layer_inputs = torch.cat([upsample_feat, feat_low], 1)
            else:
                top_down_layer_inputs = torch.cat([feat_low, upsample_feat], 1)
            inner_out = self.top_down_layers[len(self.in_channels) - 1 - idx](top_down_layer_inputs, txt_feats)
            inner_outs.insert(0, inner_out)

        # 自底向上路径
        outs = [inner_outs[0]]
        for idx in range(len(self.in_channels) - 1):
            feat_low = outs[-1]
            feat_high = inner_outs[idx + 1]
            downsample_feat = self.downsample_layers[idx](feat_low)
            out = self.bottom_up_layers[idx](torch.cat([downsample_feat, feat_high], 1), txt_feats)
            outs.append(out)

        # 输出层
        results = []
        for idx in range(len(self.in_channels)):
            results.append(self.out_layers[idx](outs[idx]))

        return tuple(results)
# 使用 @MODELS 注册 YOLOWorldDualPAFPN 类
@MODELS.register_module()
# 定义 YOLOWorldDualPAFPN 类，继承自 YOLOWorldPAFPN 类
class YOLOWorldDualPAFPN(YOLOWorldPAFPN):
    """Path Aggregation Network used in YOLO World v8."""
    # 初始化函数，接受多个参数
    def __init__(self,
                 in_channels: List[int],  # 输入通道列表
                 out_channels: Union[List[int], int],  # 输出通道列表或整数
                 guide_channels: int,  # 引导通道数
                 embed_channels: List[int],  # 嵌入通道列表
                 num_heads: List[int],  # 多头注意力机制的头数列表
                 deepen_factor: float = 1.0,  # 加深因子，默认为1.0
                 widen_factor: float = 1.0,  # 扩宽因子，默认为1.0
                 num_csp_blocks: int = 3,  # CSP块的数量，默认为3
                 freeze_all: bool = False,  # 是否冻结所有层，默认为False
                 text_enhancder: ConfigType = dict(  # 文本增强器配置
                     type='ImagePoolingAttentionModule',  # 类型为图像池化注意力模块
                     embed_channels=256,  # 嵌入通道数为256
                     num_heads=8,  # 多头注意力机制的头数为8
                     pool_size=3),  # 池化大小为3
                 block_cfg: ConfigType = dict(type='CSPLayerWithTwoConv'),  # 块配置，默认为CSPLayerWithTwoConv
                 norm_cfg: ConfigType = dict(type='BN',  # 归一化配置，默认为BN
                                             momentum=0.03,  # 动量为0.03
                                             eps=0.001),  # epsilon为0.001
                 act_cfg: ConfigType = dict(type='SiLU', inplace=True),  # 激活函数配置，默认为SiLU
                 init_cfg: OptMultiConfig = None) -> None:  # 初始化配置，默认为None，返回None
        # 调用父类的初始化函数
        super().__init__(in_channels=in_channels,
                         out_channels=out_channels,
                         guide_channels=guide_channels,
                         embed_channels=embed_channels,
                         num_heads=num_heads,
                         deepen_factor=deepen_factor,
                         widen_factor=widen_factor,
                         num_csp_blocks=num_csp_blocks,
                         freeze_all=freeze_all,
                         block_cfg=block_cfg,
                         norm_cfg=norm_cfg,
                         act_cfg=act_cfg,
                         init_cfg=init_cfg)

        # 更新文本增强器配置
        text_enhancder.update(
            dict(
                image_channels=[int(x * widen_factor) for x in out_channels],  # 图像通道数根据输出通道和扩宽因子计算
                text_channels=guide_channels,  # 文本通道数为引导通道数
                num_feats=len(out_channels),  # 特征数量为输出通道数的长度
            ))
        # 打印文本增强器配置
        print(text_enhancder)
        # 构建文本增强器模型
        self.text_enhancer = MODELS.build(text_enhancder)
    # 定义前向传播函数，接受图像特征列表和文本特征作为输入，返回元组
    def forward(self, img_feats: List[Tensor], txt_feats: Tensor) -> tuple:
        """Forward function."""
        # 断言图像特征列表的长度与输入通道数相同
        assert len(img_feats) == len(self.in_channels)
        
        # 减少层
        reduce_outs = []
        for idx in range(len(self.in_channels)):
            reduce_outs.append(self.reduce_layers[idx](img_feats[idx]))

        # 自顶向下路径
        inner_outs = [reduce_outs[-1]]
        for idx in range(len(self.in_channels) - 1, 0, -1):
            feat_high = inner_outs[0]
            feat_low = reduce_outs[idx - 1]
            upsample_feat = self.upsample_layers[len(self.in_channels) - 1 - idx](feat_high)
            if self.upsample_feats_cat_first:
                top_down_layer_inputs = torch.cat([upsample_feat, feat_low], 1)
            else:
                top_down_layer_inputs = torch.cat([feat_low, upsample_feat], 1)
            inner_out = self.top_down_layers[len(self.in_channels) - 1 - idx](top_down_layer_inputs, txt_feats)
            inner_outs.insert(0, inner_out)

        # 对文本特征进行增强
        txt_feats = self.text_enhancer(txt_feats, inner_outs)
        
        # 自底向上路径
        outs = [inner_outs[0]]
        for idx in range(len(self.in_channels) - 1):
            feat_low = outs[-1]
            feat_high = inner_outs[idx + 1]
            downsample_feat = self.downsample_layers[idx](feat_low)
            out = self.bottom_up_layers[idx](torch.cat([downsample_feat, feat_high], 1), txt_feats)
            outs.append(out)

        # 输出层
        results = []
        for idx in range(len(self.in_channels)):
            results.append(self.out_layers[idx](outs[idx]))

        return tuple(results)

.\YOLO-World\yolo_world\models\necks\__init__.py

# 版权声明，版权归腾讯公司所有
# 导入yolo_world_pafpn模块中的YOLOWorldPAFPN和YOLOWorldDualPAFPN类
from .yolo_world_pafpn import YOLOWorldPAFPN, YOLOWorldDualPAFPN
# 定义__all__列表，包含YOLOWorldPAFPN和YOLOWorldDualPAFPN类，用于模块导入时指定可导入的内容
__all__ = ['YOLOWorldPAFPN', 'YOLOWorldDualPAFPN']

.\YOLO-World\yolo_world\models\__init__.py

# 导入 Tencent 公司所有权的代码库
# 从 backbones 模块中导入所有内容
from .backbones import *  # noqa
# 从 layers 模块中导入所有内容
from .layers import *  # noqa
# 从 detectors 模块中导入所有内容
from .detectors import *  # noqa
# 从 losses 模块中导入所有内容
from .losses import *  # noqa
# 从 data_preprocessors 模块中导入所有内容
from .data_preprocessors import *  # noqa
# 从 dense_heads 模块中导入所有内容
from .dense_heads import *  # noqa
# 从 necks 模块中导入所有内容
from .necks import *  # noqa

.\YOLO-World\yolo_world\version.py

# 版权声明
# 版权所有 © 腾讯公司

# 定义版本号
__version__ = '0.1.0'

# 解析版本信息的函数
def parse_version_info(version_str):
    """Parse a version string into a tuple.

    Args:
        version_str (str): The version string.
    Returns:
        tuple[int | str]: The version info, e.g., "1.3.0" is parsed into
            (1, 3, 0), and "2.0.0rc1" is parsed into (2, 0, 0, 'rc1').
    """
    # 初始化版本信息列表
    version_info = []
    # 根据 '.' 分割版本号字符串
    for x in version_str.split('.'):
        # 如果是数字，则转换为整数
        if x.isdigit():
            version_info.append(int(x))
        # 如果包含 'rc'，则分割出补丁版本号
        elif x.find('rc') != -1:
            patch_version = x.split('rc')
            version_info.append(int(patch_version[0]))
            version_info.append(f'rc{patch_version[1]}')
    # 返回版本信息元组
    return tuple(version_info)

# 调用解析版本信息函数，得到版本信息元组
version_info = parse_version_info(__version__)

# 导出的变量列表
__all__ = ['__version__', 'version_info', 'parse_version_info']

.\YOLO-World\yolo_world\__init__.py

# 导入当前目录下的 models 模块中的所有内容
from .models import *  # noqa
# 导入当前目录下的 datasets 模块中的所有内容
from .datasets import *  # noqa
# 导入当前目录下的 engine 模块中的所有内容
from .engine import *  # noqa