如果你正在处理高分辨率图像比如医学影像分析或卫星图像识别是否经常遇到显存不足的问题传统的注意力机制在处理长序列时计算复杂度呈平方级增长这让很多CV任务在现实场景中难以落地。而滑动窗口注意力通过局部注意力机制将复杂度降低到线性级别但现有的实现往往忽略了像素级别的精细依赖关系。最近CCF-A类期刊TIFS 2026年发表的重构滑动窗口注意力RSWAtt正是针对这一痛点提出的创新解决方案。它不仅继承了滑动窗口注意力的高效特性更重要的是通过重构机制实现了真正的像素级依赖建模在保持计算效率的同时大幅提升了视觉任务的精度。本文将从实际应用角度深入解析RSWAtt的核心原理提供完整的代码实现并展示如何在常见的CV任务中即插即用这一先进技术。无论你是从事目标检测、图像分割还是其他视觉任务都能从中获得实用的技术方案。1. RSWAtt要解决的核心问题1.1 传统注意力机制在CV任务中的瓶颈在计算机视觉领域随着图像分辨率的不断提升传统的自注意力机制面临着严峻的计算挑战。一个1080p的高清图像包含超过200万个像素点如果使用标准的Transformer架构注意力矩阵的大小将达到(200万×200万)这在实际应用中是完全不可行的。滑动窗口注意力通过将全局注意力限制在局部窗口内确实解决了计算复杂度的问题。但这种方法引入了新的问题窗口之间的信息交互不足难以捕捉长距离的像素依赖关系。特别是在边缘检测、细粒度分割等任务中这种局限性尤为明显。1.2 RSWAtt的创新突破点RSWAtt的核心创新在于引入了重构机制它不仅仅是在局部窗口内计算注意力更重要的是通过多层次的特征重构来实现窗口间的信息融合。这种设计使得模型能够在保持线性计算复杂度的同时建立更加丰富的像素级依赖关系。具体来说RSWAtt通过三个关键改进实现了这一目标动态窗口重构根据图像内容自适应调整窗口大小和形状跨窗口信息传递设计专门的机制促进不同窗口间的特征交互多尺度特征融合在不同尺度上建立像素依赖关系2. 滑动窗口注意力的基础概念2.1 标准自注意力机制回顾在深入理解RSWAtt之前我们需要回顾一下标准的自注意力机制。自注意力的计算公式为import torch import torch.nn as nn import torch.nn.functional as F class StandardSelfAttention(nn.Module): def __init__(self, dim, heads8, dim_headNone): super().__init__() self.heads heads self.dim_head dim_head if dim_head else dim // heads self.scale self.dim_head ** -0.5 self.to_qkv nn.Linear(dim, self.dim_head * heads * 3, biasFalse) self.to_out nn.Linear(self.dim_head * heads, dim) def forward(self, x): b, n, d x.shape qkv self.to_qkv(x).chunk(3, dim-1) q, k, v map(lambda t: t.reshape(b, n, self.heads, self.dim_head).transpose(1, 2), qkv) dots torch.matmul(q, k.transpose(-1, -2)) * self.scale attn dots.softmax(dim-1) out torch.matmul(attn, v) out out.transpose(1, 2).reshape(b, n, -1) return self.to_out(out)这种标准实现的计算复杂度为O(n²)其中n是序列长度。对于高分辨率图像这种复杂度是无法接受的。2.2 滑动窗口注意力的基本原理滑动窗口注意力通过限制每个位置只能关注其邻近的位置来降低计算复杂度。具体实现如下class SlidingWindowAttention(nn.Module): def __init__(self, dim, window_size, heads8, dim_headNone): super().__init__() self.heads heads self.dim_head dim_head if dim_head else dim // heads self.scale self.dim_head ** -0.5 self.window_size window_size self.to_qkv nn.Linear(dim, self.dim_head * heads * 3, biasFalse) self.to_out nn.Linear(self.dim_head * heads, dim) def create_window_masks(self, h, w, device): 创建滑动窗口掩码 masks [] for i in range(0, h, self.window_size): for j in range(0, w, self.window_size): mask torch.zeros(h, w, devicedevice) i_end min(i self.window_size, h) j_end min(j self.window_size, w) mask[i:i_end, j:j_end] 1 masks.append(mask) return torch.stack(masks) def forward(self, x, h, w): b, n, d x.shape assert n h * w, 序列长度必须等于高度×宽度 # 将序列重塑为空间格式 x_spatial x.reshape(b, h, w, d) # 应用滑动窗口注意力 output_windows [] window_masks self.create_window_masks(h, w, x.device) for mask in window_masks: window_indices mask.nonzero(as_tupleFalse) if len(window_indices) 0: continue # 提取窗口内的特征 window_features x_spatial[:, window_indices[:, 0], window_indices[:, 1], :] window_features window_features.reshape(b, -1, d) # 在窗口内应用标准注意力 qkv self.to_qkv(window_features).chunk(3, dim-1) q, k, v map(lambda t: t.reshape(b, -1, self.heads, self.dim_head).transpose(1, 2), qkv) dots torch.matmul(q, k.transpose(-1, -2)) * self.scale attn dots.softmax(dim-1) window_out torch.matmul(attn, v) window_out window_out.transpose(1, 2).reshape(b, -1, self.dim_head * self.heads) window_out self.to_out(window_out) # 将结果放回原位置 output_window torch.zeros(b, n, d, devicex.device) flat_indices window_indices[:, 0] * w window_indices[:, 1] output_window[:, flat_indices] window_out output_windows.append(output_window) # 合并所有窗口的结果 output sum(output_windows) / len(output_windows) return output这种实现的复杂度降低到了O(n×w²)其中w是窗口大小通常远小于序列长度n。3. RSWAtt的核心原理与架构设计3.1 重构机制的核心思想RSWAtt的核心创新在于引入了特征重构机制。传统的滑动窗口注意力在独立的窗口内计算注意力窗口之间缺乏有效的信息交互。RSWAtt通过多层次的重构操作实现了窗口间的特征融合。重构机制包含三个关键步骤局部特征提取在每个窗口内进行精细的特征学习跨窗口信息传递通过重叠窗口或特征重采样实现信息交换全局特征重构整合所有窗口的信息重建全局特征表示3.2 RSWAtt的完整架构实现下面是RSWAtt的完整PyTorch实现import torch import torch.nn as nn import torch.nn.functional as F class RSWAtt(nn.Module): def __init__(self, dim, window_size, shift_size0, num_heads8, qkv_biasTrue, attn_drop0., proj_drop0.): super().__init__() self.dim dim self.window_size window_size self.shift_size shift_size self.num_heads num_heads head_dim dim // num_heads self.scale head_dim ** -0.5 self.qkv nn.Linear(dim, dim * 3, biasqkv_bias) self.attn_drop nn.Dropout(attn_drop) self.proj nn.Linear(dim, dim) self.proj_drop nn.Dropout(proj_drop) # 重构相关的参数 self.reconstruction_mlp nn.Sequential( nn.Linear(dim, dim * 2), nn.GELU(), nn.Linear(dim * 2, dim) ) self.softmax nn.Softmax(dim-1) def forward(self, x, maskNone): 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] # 形状: (B_, num_heads, N, head_dim) q q * self.scale attn (q k.transpose(-2, -1)) # (B_, num_heads, N, N) 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) attn self.attn_drop(attn) x (attn v).transpose(1, 2).reshape(B_, N, C) # 特征重构步骤 reconstructed_x self.reconstruction_mlp(x) x x reconstructed_x # 残差连接 x self.proj(x) x self.proj_drop(x) return x class WindowPartition(nn.Module): 将特征图划分为不重叠的窗口 def __init__(self, window_size): super().__init__() self.window_size window_size def forward(self, x): Args: x: (B, H, W, C) Returns: windows: (num_windows*B, window_size, window_size, C) B, H, W, C x.shape x x.view(B, H // self.window_size, self.window_size, W // self.window_size, self.window_size, C) windows x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, self.window_size, self.window_size, C) return windows class WindowReverse(nn.Module): 将窗口反转回特征图 def __init__(self, window_size, H, W): super().__init__() self.window_size window_size self.H H self.W W def forward(self, windows): Args: windows: (num_windows*B, window_size, window_size, C) Returns: x: (B, H, W, C) B int(windows.shape[0] / (self.H * self.W / self.window_size / self.window_size)) x windows.view(B, self.H // self.window_size, self.W // self.window_size, self.window_size, self.window_size, -1) x x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, self.H, self.W, -1) return x3.3 像素级依赖建模的实现细节RSWAtt通过特殊的注意力权重计算和特征重构机制实现像素级依赖建模class PixelLevelDependency(nn.Module): 像素级依赖建模模块 def __init__(self, dim, kernel_size3): super().__init__() self.dim dim self.kernel_size kernel_size self.padding kernel_size // 2 # 局部卷积用于捕捉像素间关系 self.local_conv nn.Conv2d(dim, dim, kernel_size, paddingself.padding, groupsdim) self.norm nn.LayerNorm(dim) self.activation nn.GELU() def forward(self, x, H, W): Args: x: (B, H*W, C) Returns: x: (B, H*W, C) 增强像素依赖的特征 B, N, C x.shape assert N H * W # 转换为空间格式 x_spatial x.transpose(1, 2).reshape(B, C, H, W) # 应用局部卷积捕捉像素关系 local_features self.local_conv(x_spatial) local_features local_features.reshape(B, C, -1).transpose(1, 2) # 与原始特征融合 enhanced_x x local_features enhanced_x self.norm(enhanced_x) enhanced_x self.activation(enhanced_x) return enhanced_x class EnhancedRSWAtt(nn.Module): 增强版的RSWAtt包含像素级依赖建模 def __init__(self, dim, window_size, num_heads, mlp_ratio4., drop0., attn_drop0.): super().__init__() self.window_size window_size self.dim dim # 注意力模块 self.attn RSWAtt(dim, window_size, num_headsnum_heads, attn_dropattn_drop, proj_dropdrop) # 像素级依赖建模 self.pixel_dependency PixelLevelDependency(dim) # MLP层 self.mlp nn.Sequential( nn.Linear(dim, int(dim * mlp_ratio)), nn.GELU(), nn.Dropout(drop), nn.Linear(int(dim * mlp_ratio), dim), nn.Dropout(drop) ) self.norm1 nn.LayerNorm(dim) self.norm2 nn.LayerNorm(dim) self.norm3 nn.LayerNorm(dim) def forward(self, x, H, W): # 第一部分窗口注意力 shortcut x x self.norm1(x) # 窗口划分 x_windows window_partition(x, self.window_size, H, W) x_windows x_windows.view(-1, self.window_size * self.window_size, self.dim) # 应用RSWAtt attn_windows self.attn(x_windows) attn_windows attn_windows.view(-1, self.window_size, self.window_size, self.dim) # 窗口反转 x window_reverse(attn_windows, self.window_size, H, W) x shortcut x # 第二部分像素级依赖建模 x x self.pixel_dependency(self.norm2(x), H, W) # 第三部分MLP x x self.mlp(self.norm3(x)) return x4. 环境准备与依赖安装4.1 基础环境配置在开始使用RSWAtt之前需要确保环境满足以下要求# 创建conda环境 conda create -n rswatt python3.8 conda activate rswatt # 安装PyTorch根据CUDA版本选择 pip install torch1.12.1cu113 torchvision0.13.1cu113 torchaudio0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113 # 安装其他依赖 pip install numpy matplotlib opencv-python pillow pip install timm # 图像模型库 pip install einops # 张量操作工具4.2 项目结构规划建议的项目结构如下rswatt_project/ ├── models/ │ ├── __init__.py │ ├── rswatt.py # RSWAtt核心实现 │ └── backbones.py # 骨干网络集成 ├── datasets/ │ ├── __init__.py │ └── vision_datasets.py # 数据加载器 ├── configs/ │ └── default.yaml # 配置文件 ├── utils/ │ ├── logger.py # 日志工具 │ └── metrics.py # 评估指标 ├── train.py # 训练脚本 ├── eval.py # 评估脚本 └── requirements.txt # 依赖列表4.3 依赖检查脚本创建环境检查脚本确保所有依赖正确安装# check_environment.py import sys import pkg_resources def check_environment(): 检查环境依赖是否满足要求 required_packages { torch: 1.12.0, torchvision: 0.13.0, numpy: 1.21.0, opencv-python: 4.5.0, timm: 0.6.0, einops: 0.6.0 } missing_packages [] version_mismatch [] for package, required_version in required_packages.items(): try: installed_version pkg_resources.get_distribution(package).version if pkg_resources.parse_version(installed_version) pkg_resources.parse_version(required_version): version_mismatch.append(f{package} (需要: {required_version}, 已安装: {installed_version})) except pkg_resources.DistributionNotFound: missing_packages.append(package) if missing_packages or version_mismatch: print(环境检查失败:) if missing_packages: print(f缺少包: {, .join(missing_packages)}) if version_mismatch: print(f版本不匹配: {, .join(version_mismatch)}) return False else: print(环境检查通过!) return True if __name__ __main__: check_environment()5. RSWAtt在CV任务中的即插即用实现5.1 图像分类任务集成将RSWAtt集成到标准的Vision Transformer中import torch.nn as nn from einops import rearrange class RSWAttVisionTransformer(nn.Module): 基于RSWAtt的Vision Transformer def __init__(self, img_size224, patch_size16, in_chans3, num_classes1000, embed_dim768, depth12, num_heads12, mlp_ratio4., window_size7, drop_rate0., attn_drop_rate0.): super().__init__() self.num_classes num_classes self.num_features self.embed_dim embed_dim self.patch_size patch_size # 图像分块嵌入 self.patch_embed PatchEmbed( img_sizeimg_size, patch_sizepatch_size, in_chansin_chans, embed_dimembed_dim) num_patches self.patch_embed.num_patches # 位置编码 self.pos_embed nn.Parameter(torch.zeros(1, num_patches, embed_dim)) self.pos_drop nn.Dropout(pdrop_rate) # RSWAtt blocks self.blocks nn.ModuleList([ EnhancedRSWAtt( dimembed_dim, window_sizewindow_size, num_headsnum_heads, mlp_ratiomlp_ratio, dropdrop_rate, attn_dropattn_drop_rate ) for i in range(depth)]) self.norm nn.LayerNorm(embed_dim) self.head nn.Linear(embed_dim, num_classes) if num_classes 0 else nn.Identity() # 初始化权重 self.apply(self._init_weights) def _init_weights(self, m): if isinstance(m, nn.Linear): nn.init.trunc_normal_(m.weight, std.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) elif isinstance(m, nn.Parameter): nn.init.trunc_normal_(m, std.02) def forward_features(self, x): B x.shape[0] x self.patch_embed(x) # (B, L, C) x x self.pos_embed x self.pos_drop(x) H, W self.patch_embed.grid_size for blk in self.blocks: x blk(x, H, W) x self.norm(x) return x.mean(dim1) # 全局平均池化 def forward(self, x): x self.forward_features(x) x self.head(x) return x class PatchEmbed(nn.Module): 图像分块嵌入 def __init__(self, img_size224, patch_size16, in_chans3, embed_dim768): super().__init__() self.img_size (img_size, img_size) self.patch_size (patch_size, patch_size) self.grid_size (img_size // patch_size, img_size // patch_size) self.num_patches self.grid_size[0] * self.grid_size[1] self.proj nn.Conv2d(in_chans, embed_dim, kernel_sizepatch_size, stridepatch_size) def forward(self, x): B, C, H, W x.shape assert H self.img_size[0] and W self.img_size[1], \ f输入图像大小({H}*{W})不匹配模型期望的{self.img_size[0]}*{self.img_size[1]} x self.proj(x).flatten(2).transpose(1, 2) # (B, L, C) return x5.2 目标检测任务适配将RSWAtt集成到目标检测框架中class RSWAttFPN(nn.Module): 基于RSWAtt的特征金字塔网络 def __init__(self, in_channels_list, out_channels, window_sizes[7, 7, 7]): super().__init__() self.lateral_convs nn.ModuleList() self.fpn_convs nn.ModuleList() self.rswatt_blocks nn.ModuleList() for i, in_channels in enumerate(in_channels_list): lateral_conv nn.Conv2d(in_channels, out_channels, 1) fpn_conv nn.Conv2d(out_channels, out_channels, 3, padding1) rswatt_block EnhancedRSWAtt( dimout_channels, window_sizewindow_sizes[i], num_headsout_channels // 32 ) self.lateral_convs.append(lateral_conv) self.fpn_convs.append(fpn_conv) self.rswatt_blocks.append(rswatt_block) # 初始化权重 for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_uniform_(m.weight, a1) if m.bias is not None: nn.init.constant_(m.bias, 0) def forward(self, inputs): Args: inputs: 不同尺度的特征图列表 Returns: outputs: 增强后的多尺度特征 # lateral连接 laterals [ lateral_conv(inputs[i]) for i, lateral_conv in enumerate(self.lateral_convs) ] # 自上而下的路径 for i in range(len(laterals) - 1, 0, -1): laterals[i - 1] F.interpolate( laterals[i], scale_factor2, modenearest ) # FPN卷积和RSWAtt增强 outputs [] for i, (lateral, fpn_conv, rswatt_block) in enumerate( zip(laterals, self.fpn_convs, self.rswatt_blocks) ): # 应用卷积 x fpn_conv(lateral) # 应用RSWAtt B, C, H, W x.shape x x.flatten(2).transpose(1, 2) # (B, H*W, C) x rswatt_block(x, H, W) x x.transpose(1, 2).reshape(B, C, H, W) outputs.append(x) return outputs5.3 语义分割任务应用在语义分割任务中集成RSWAttclass RSWAttUNet(nn.Module): 基于RSWAtt的UNet分割网络 def __init__(self, in_channels3, num_classes21, base_dim64, window_size7, depths[2, 2, 2, 2]): super().__init__() self.depths depths # 编码器 self.encoders nn.ModuleList() self.downsamplers nn.ModuleList() dim base_dim for i, depth in enumerate(depths): encoder_blocks nn.ModuleList([ RSWAttEncoderBlock(dim, window_size) for _ in range(depth) ]) self.encoders.append(encoder_blocks) if i len(depths) - 1: downsampler nn.Sequential( nn.Conv2d(dim, dim * 2, kernel_size3, stride2, padding1), nn.BatchNorm2d(dim * 2), nn.ReLU(inplaceTrue) ) self.downsamplers.append(downsampler) dim * 2 # 解码器 self.decoders nn.ModuleList() self.upsamplers nn.ModuleList() for i, depth in enumerate(reversed(depths[:-1])): upsampler nn.Sequential( nn.ConvTranspose2d(dim, dim // 2, kernel_size2, stride2), nn.BatchNorm2d(dim // 2), nn.ReLU(inplaceTrue) ) self.upsamplers.append(upsampler) dim // 2 decoder_blocks nn.ModuleList([ RSWAttDecoderBlock(dim, window_size) for _ in range(depth) ]) self.decoders.append(decoder_blocks) # 输出层 self.output_conv nn.Conv2d(base_dim, num_classes, kernel_size1) def forward(self, x): # 编码路径 skips [] for i, (encoder_blocks, downsampler) in enumerate( zip(self.encoders, self.downsamplers [None]) ): for block in encoder_blocks: x block(x) skips.append(x) if downsampler is not None: x downsampler(x) # 解码路径 for i, (upsampler, decoder_blocks, skip) in enumerate( zip(self.upsamplers, self.decoders, reversed(skips[:-1])) ): x upsampler(x) # 跳跃连接 x torch.cat([x, skip], dim1) for block in decoder_blocks: x block(x) return self.output_conv(x) class RSWAttEncoderBlock(nn.Module): RSWAtt编码器块 def __init__(self, dim, window_size): super().__init__() self.conv nn.Sequential( nn.Conv2d(dim, dim, 3, padding1), nn.BatchNorm2d(dim), nn.ReLU(inplaceTrue) ) self.rswatt EnhancedRSWAtt(dim, window_size, num_headsdim//32) def forward(self, x): # 卷积路径 conv_out self.conv(x) # RSWAtt路径 B, C, H, W x.shape x_flat x.flatten(2).transpose(1, 2) attn_out self.rswatt(x_flat, H, W) attn_out attn_out.transpose(1, 2).reshape(B, C, H, W) return conv_out attn_out6. 完整训练流程与代码实现6.1 训练配置与参数设置创建完整的训练配置类import yaml from dataclasses import dataclass from typing import Dict, Any dataclass class TrainingConfig: 训练配置参数 # 数据参数 data_path: str ./data batch_size: int 32 num_workers: int 4 image_size: int 224 # 模型参数 model_name: str rswatt_vit_small num_classes: int 1000 patch_size: int 16 embed_dim: int 384 depth: int 12 num_heads: int 6 window_size: int 7 mlp_ratio: float 4.0 # 优化器参数 learning_rate: float 1e-3 weight_decay: float 0.05 momentum: float 0.9 optimizer: str adamw # 学习率调度 scheduler: str cosine warmup_epochs: int 5 min_lr: float 1e-6 # 训练参数 epochs: int 300 eval_interval: int 1 save_interval: int 10 resume: str # 恢复训练的检查点路径 # 设备参数 device: str cuda if torch.cuda.is_available() else cpu amp: bool True # 自动混合精度 def to_dict(self) - Dict[str, Any]: return {k: v for k, v in self.__dict__.items() if not k.startswith(_)} classmethod def from_yaml(cls, yaml_path: str): with open(yaml_path, r) as f: config_dict yaml.safe_load(f) return cls(**config_dict) def save_to_yaml(self, yaml_path: str): with open(yaml_path, w) as f: yaml.dump(self.to_dict(), f, default_flow_styleFalse)6.2 完整的训练脚本import torch import torch.nn as nn import torch.optim as optim from torch.utils.data import DataLoader from torch.cuda.amp import GradScaler, autocast import time import os from pathlib import Path class RSWAttTrainer: RSWAtt模型训练器 def __init__(self, config: TrainingConfig): self.config config self.device torch.device(config.device) self.scaler GradScaler() if config.amp else None # 创建模型 self.model self._create_model() self.model.to(self.device) # 创建数据加载器 self.train_loader, self.val_loader self._create_dataloaders() # 创建优化器和调度器 self.optimizer self._create_optimizer() self.scheduler self._create_scheduler() # 损失函数 self.criterion nn.CrossEntropyLoss() # 训练状态 self.epoch 0 self.best_acc 0.0 self.train_losses [] self.val_accuracies [] # 创建输出目录 self.output_dir Path(./outputs) self.output_dir.mkdir(exist_okTrue) def _create_model(self): 创建RSWAtt模型 if self.config.model_name rswatt_vit_small: return RSWAttVisionTransformer( img_sizeself.config.image_size, patch_sizeself.config.patch_size, num_classesself.config.num_classes, embed_dimself.config.embed_dim, depthself.config.depth, num_headsself.config.num_heads, window_sizeself.config.window_size, mlp_ratioself.config.mlp_ratio ) else: raise ValueError(f未知模型: {self.config.model_name}) def _create_dataloaders(self): 创建数据加载器 # 这里使用伪数据集实际项目中替换为真实数据集 from torch.utils.data import TensorDataset # 创建伪训练数据 x_train torch.randn(1000, 3, self.config.image_size, self.config.image_size) y_train torch.randint(0, self.config.num_classes, (1000,)) train_dataset TensorDataset(x_train, y_train) # 创建伪验证数据 x_val torch.randn(200, 3, self.config.image_size, self.config.image_size) y_val torch.randint(0, self.config.num_classes, (200,)) val_dataset TensorDataset(x_val, y_val) train_loader DataLoader( train_dataset, batch_sizeself.config.batch_size, shuffleTrue, num_workersself.config.num_workers ) val_loader DataLoader( val_dataset, batch_sizeself.config.batch_size, shuffleFalse, num_workersself.config.num_workers ) return train_loader, val_loader def _create_optimizer(self): 创建优化器
RSWAtt重构滑动窗口注意力：像素级依赖建模与CV任务实战
发布时间：2026/7/12 2:44:35
如果你正在处理高分辨率图像比如医学影像分析或卫星图像识别是否经常遇到显存不足的问题传统的注意力机制在处理长序列时计算复杂度呈平方级增长这让很多CV任务在现实场景中难以落地。而滑动窗口注意力通过局部注意力机制将复杂度降低到线性级别但现有的实现往往忽略了像素级别的精细依赖关系。最近CCF-A类期刊TIFS 2026年发表的重构滑动窗口注意力RSWAtt正是针对这一痛点提出的创新解决方案。它不仅继承了滑动窗口注意力的高效特性更重要的是通过重构机制实现了真正的像素级依赖建模在保持计算效率的同时大幅提升了视觉任务的精度。本文将从实际应用角度深入解析RSWAtt的核心原理提供完整的代码实现并展示如何在常见的CV任务中即插即用这一先进技术。无论你是从事目标检测、图像分割还是其他视觉任务都能从中获得实用的技术方案。1. RSWAtt要解决的核心问题1.1 传统注意力机制在CV任务中的瓶颈在计算机视觉领域随着图像分辨率的不断提升传统的自注意力机制面临着严峻的计算挑战。一个1080p的高清图像包含超过200万个像素点如果使用标准的Transformer架构注意力矩阵的大小将达到(200万×200万)这在实际应用中是完全不可行的。滑动窗口注意力通过将全局注意力限制在局部窗口内确实解决了计算复杂度的问题。但这种方法引入了新的问题窗口之间的信息交互不足难以捕捉长距离的像素依赖关系。特别是在边缘检测、细粒度分割等任务中这种局限性尤为明显。1.2 RSWAtt的创新突破点RSWAtt的核心创新在于引入了重构机制它不仅仅是在局部窗口内计算注意力更重要的是通过多层次的特征重构来实现窗口间的信息融合。这种设计使得模型能够在保持线性计算复杂度的同时建立更加丰富的像素级依赖关系。具体来说RSWAtt通过三个关键改进实现了这一目标动态窗口重构根据图像内容自适应调整窗口大小和形状跨窗口信息传递设计专门的机制促进不同窗口间的特征交互多尺度特征融合在不同尺度上建立像素依赖关系2. 滑动窗口注意力的基础概念2.1 标准自注意力机制回顾在深入理解RSWAtt之前我们需要回顾一下标准的自注意力机制。自注意力的计算公式为import torch import torch.nn as nn import torch.nn.functional as F class StandardSelfAttention(nn.Module): def __init__(self, dim, heads8, dim_headNone): super().__init__() self.heads heads self.dim_head dim_head if dim_head else dim // heads self.scale self.dim_head ** -0.5 self.to_qkv nn.Linear(dim, self.dim_head * heads * 3, biasFalse) self.to_out nn.Linear(self.dim_head * heads, dim) def forward(self, x): b, n, d x.shape qkv self.to_qkv(x).chunk(3, dim-1) q, k, v map(lambda t: t.reshape(b, n, self.heads, self.dim_head).transpose(1, 2), qkv) dots torch.matmul(q, k.transpose(-1, -2)) * self.scale attn dots.softmax(dim-1) out torch.matmul(attn, v) out out.transpose(1, 2).reshape(b, n, -1) return self.to_out(out)这种标准实现的计算复杂度为O(n²)其中n是序列长度。对于高分辨率图像这种复杂度是无法接受的。2.2 滑动窗口注意力的基本原理滑动窗口注意力通过限制每个位置只能关注其邻近的位置来降低计算复杂度。具体实现如下class SlidingWindowAttention(nn.Module): def __init__(self, dim, window_size, heads8, dim_headNone): super().__init__() self.heads heads self.dim_head dim_head if dim_head else dim // heads self.scale self.dim_head ** -0.5 self.window_size window_size self.to_qkv nn.Linear(dim, self.dim_head * heads * 3, biasFalse) self.to_out nn.Linear(self.dim_head * heads, dim) def create_window_masks(self, h, w, device): 创建滑动窗口掩码 masks [] for i in range(0, h, self.window_size): for j in range(0, w, self.window_size): mask torch.zeros(h, w, devicedevice) i_end min(i self.window_size, h) j_end min(j self.window_size, w) mask[i:i_end, j:j_end] 1 masks.append(mask) return torch.stack(masks) def forward(self, x, h, w): b, n, d x.shape assert n h * w, 序列长度必须等于高度×宽度 # 将序列重塑为空间格式 x_spatial x.reshape(b, h, w, d) # 应用滑动窗口注意力 output_windows [] window_masks self.create_window_masks(h, w, x.device) for mask in window_masks: window_indices mask.nonzero(as_tupleFalse) if len(window_indices) 0: continue # 提取窗口内的特征 window_features x_spatial[:, window_indices[:, 0], window_indices[:, 1], :] window_features window_features.reshape(b, -1, d) # 在窗口内应用标准注意力 qkv self.to_qkv(window_features).chunk(3, dim-1) q, k, v map(lambda t: t.reshape(b, -1, self.heads, self.dim_head).transpose(1, 2), qkv) dots torch.matmul(q, k.transpose(-1, -2)) * self.scale attn dots.softmax(dim-1) window_out torch.matmul(attn, v) window_out window_out.transpose(1, 2).reshape(b, -1, self.dim_head * self.heads) window_out self.to_out(window_out) # 将结果放回原位置 output_window torch.zeros(b, n, d, devicex.device) flat_indices window_indices[:, 0] * w window_indices[:, 1] output_window[:, flat_indices] window_out output_windows.append(output_window) # 合并所有窗口的结果 output sum(output_windows) / len(output_windows) return output这种实现的复杂度降低到了O(n×w²)其中w是窗口大小通常远小于序列长度n。3. RSWAtt的核心原理与架构设计3.1 重构机制的核心思想RSWAtt的核心创新在于引入了特征重构机制。传统的滑动窗口注意力在独立的窗口内计算注意力窗口之间缺乏有效的信息交互。RSWAtt通过多层次的重构操作实现了窗口间的特征融合。重构机制包含三个关键步骤局部特征提取在每个窗口内进行精细的特征学习跨窗口信息传递通过重叠窗口或特征重采样实现信息交换全局特征重构整合所有窗口的信息重建全局特征表示3.2 RSWAtt的完整架构实现下面是RSWAtt的完整PyTorch实现import torch import torch.nn as nn import torch.nn.functional as F class RSWAtt(nn.Module): def __init__(self, dim, window_size, shift_size0, num_heads8, qkv_biasTrue, attn_drop0., proj_drop0.): super().__init__() self.dim dim self.window_size window_size self.shift_size shift_size self.num_heads num_heads head_dim dim // num_heads self.scale head_dim ** -0.5 self.qkv nn.Linear(dim, dim * 3, biasqkv_bias) self.attn_drop nn.Dropout(attn_drop) self.proj nn.Linear(dim, dim) self.proj_drop nn.Dropout(proj_drop) # 重构相关的参数 self.reconstruction_mlp nn.Sequential( nn.Linear(dim, dim * 2), nn.GELU(), nn.Linear(dim * 2, dim) ) self.softmax nn.Softmax(dim-1) def forward(self, x, maskNone): 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] # 形状: (B_, num_heads, N, head_dim) q q * self.scale attn (q k.transpose(-2, -1)) # (B_, num_heads, N, N) 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) attn self.attn_drop(attn) x (attn v).transpose(1, 2).reshape(B_, N, C) # 特征重构步骤 reconstructed_x self.reconstruction_mlp(x) x x reconstructed_x # 残差连接 x self.proj(x) x self.proj_drop(x) return x class WindowPartition(nn.Module): 将特征图划分为不重叠的窗口 def __init__(self, window_size): super().__init__() self.window_size window_size def forward(self, x): Args: x: (B, H, W, C) Returns: windows: (num_windows*B, window_size, window_size, C) B, H, W, C x.shape x x.view(B, H // self.window_size, self.window_size, W // self.window_size, self.window_size, C) windows x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, self.window_size, self.window_size, C) return windows class WindowReverse(nn.Module): 将窗口反转回特征图 def __init__(self, window_size, H, W): super().__init__() self.window_size window_size self.H H self.W W def forward(self, windows): Args: windows: (num_windows*B, window_size, window_size, C) Returns: x: (B, H, W, C) B int(windows.shape[0] / (self.H * self.W / self.window_size / self.window_size)) x windows.view(B, self.H // self.window_size, self.W // self.window_size, self.window_size, self.window_size, -1) x x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, self.H, self.W, -1) return x3.3 像素级依赖建模的实现细节RSWAtt通过特殊的注意力权重计算和特征重构机制实现像素级依赖建模class PixelLevelDependency(nn.Module): 像素级依赖建模模块 def __init__(self, dim, kernel_size3): super().__init__() self.dim dim self.kernel_size kernel_size self.padding kernel_size // 2 # 局部卷积用于捕捉像素间关系 self.local_conv nn.Conv2d(dim, dim, kernel_size, paddingself.padding, groupsdim) self.norm nn.LayerNorm(dim) self.activation nn.GELU() def forward(self, x, H, W): Args: x: (B, H*W, C) Returns: x: (B, H*W, C) 增强像素依赖的特征 B, N, C x.shape assert N H * W # 转换为空间格式 x_spatial x.transpose(1, 2).reshape(B, C, H, W) # 应用局部卷积捕捉像素关系 local_features self.local_conv(x_spatial) local_features local_features.reshape(B, C, -1).transpose(1, 2) # 与原始特征融合 enhanced_x x local_features enhanced_x self.norm(enhanced_x) enhanced_x self.activation(enhanced_x) return enhanced_x class EnhancedRSWAtt(nn.Module): 增强版的RSWAtt包含像素级依赖建模 def __init__(self, dim, window_size, num_heads, mlp_ratio4., drop0., attn_drop0.): super().__init__() self.window_size window_size self.dim dim # 注意力模块 self.attn RSWAtt(dim, window_size, num_headsnum_heads, attn_dropattn_drop, proj_dropdrop) # 像素级依赖建模 self.pixel_dependency PixelLevelDependency(dim) # MLP层 self.mlp nn.Sequential( nn.Linear(dim, int(dim * mlp_ratio)), nn.GELU(), nn.Dropout(drop), nn.Linear(int(dim * mlp_ratio), dim), nn.Dropout(drop) ) self.norm1 nn.LayerNorm(dim) self.norm2 nn.LayerNorm(dim) self.norm3 nn.LayerNorm(dim) def forward(self, x, H, W): # 第一部分窗口注意力 shortcut x x self.norm1(x) # 窗口划分 x_windows window_partition(x, self.window_size, H, W) x_windows x_windows.view(-1, self.window_size * self.window_size, self.dim) # 应用RSWAtt attn_windows self.attn(x_windows) attn_windows attn_windows.view(-1, self.window_size, self.window_size, self.dim) # 窗口反转 x window_reverse(attn_windows, self.window_size, H, W) x shortcut x # 第二部分像素级依赖建模 x x self.pixel_dependency(self.norm2(x), H, W) # 第三部分MLP x x self.mlp(self.norm3(x)) return x4. 环境准备与依赖安装4.1 基础环境配置在开始使用RSWAtt之前需要确保环境满足以下要求# 创建conda环境 conda create -n rswatt python3.8 conda activate rswatt # 安装PyTorch根据CUDA版本选择 pip install torch1.12.1cu113 torchvision0.13.1cu113 torchaudio0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113 # 安装其他依赖 pip install numpy matplotlib opencv-python pillow pip install timm # 图像模型库 pip install einops # 张量操作工具4.2 项目结构规划建议的项目结构如下rswatt_project/ ├── models/ │ ├── __init__.py │ ├── rswatt.py # RSWAtt核心实现 │ └── backbones.py # 骨干网络集成 ├── datasets/ │ ├── __init__.py │ └── vision_datasets.py # 数据加载器 ├── configs/ │ └── default.yaml # 配置文件 ├── utils/ │ ├── logger.py # 日志工具 │ └── metrics.py # 评估指标 ├── train.py # 训练脚本 ├── eval.py # 评估脚本 └── requirements.txt # 依赖列表4.3 依赖检查脚本创建环境检查脚本确保所有依赖正确安装# check_environment.py import sys import pkg_resources def check_environment(): 检查环境依赖是否满足要求 required_packages { torch: 1.12.0, torchvision: 0.13.0, numpy: 1.21.0, opencv-python: 4.5.0, timm: 0.6.0, einops: 0.6.0 } missing_packages [] version_mismatch [] for package, required_version in required_packages.items(): try: installed_version pkg_resources.get_distribution(package).version if pkg_resources.parse_version(installed_version) pkg_resources.parse_version(required_version): version_mismatch.append(f{package} (需要: {required_version}, 已安装: {installed_version})) except pkg_resources.DistributionNotFound: missing_packages.append(package) if missing_packages or version_mismatch: print(环境检查失败:) if missing_packages: print(f缺少包: {, .join(missing_packages)}) if version_mismatch: print(f版本不匹配: {, .join(version_mismatch)}) return False else: print(环境检查通过!) return True if __name__ __main__: check_environment()5. RSWAtt在CV任务中的即插即用实现5.1 图像分类任务集成将RSWAtt集成到标准的Vision Transformer中import torch.nn as nn from einops import rearrange class RSWAttVisionTransformer(nn.Module): 基于RSWAtt的Vision Transformer def __init__(self, img_size224, patch_size16, in_chans3, num_classes1000, embed_dim768, depth12, num_heads12, mlp_ratio4., window_size7, drop_rate0., attn_drop_rate0.): super().__init__() self.num_classes num_classes self.num_features self.embed_dim embed_dim self.patch_size patch_size # 图像分块嵌入 self.patch_embed PatchEmbed( img_sizeimg_size, patch_sizepatch_size, in_chansin_chans, embed_dimembed_dim) num_patches self.patch_embed.num_patches # 位置编码 self.pos_embed nn.Parameter(torch.zeros(1, num_patches, embed_dim)) self.pos_drop nn.Dropout(pdrop_rate) # RSWAtt blocks self.blocks nn.ModuleList([ EnhancedRSWAtt( dimembed_dim, window_sizewindow_size, num_headsnum_heads, mlp_ratiomlp_ratio, dropdrop_rate, attn_dropattn_drop_rate ) for i in range(depth)]) self.norm nn.LayerNorm(embed_dim) self.head nn.Linear(embed_dim, num_classes) if num_classes 0 else nn.Identity() # 初始化权重 self.apply(self._init_weights) def _init_weights(self, m): if isinstance(m, nn.Linear): nn.init.trunc_normal_(m.weight, std.02) if isinstance(m, nn.Linear) and m.bias is not None: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.LayerNorm): nn.init.constant_(m.bias, 0) nn.init.constant_(m.weight, 1.0) elif isinstance(m, nn.Parameter): nn.init.trunc_normal_(m, std.02) def forward_features(self, x): B x.shape[0] x self.patch_embed(x) # (B, L, C) x x self.pos_embed x self.pos_drop(x) H, W self.patch_embed.grid_size for blk in self.blocks: x blk(x, H, W) x self.norm(x) return x.mean(dim1) # 全局平均池化 def forward(self, x): x self.forward_features(x) x self.head(x) return x class PatchEmbed(nn.Module): 图像分块嵌入 def __init__(self, img_size224, patch_size16, in_chans3, embed_dim768): super().__init__() self.img_size (img_size, img_size) self.patch_size (patch_size, patch_size) self.grid_size (img_size // patch_size, img_size // patch_size) self.num_patches self.grid_size[0] * self.grid_size[1] self.proj nn.Conv2d(in_chans, embed_dim, kernel_sizepatch_size, stridepatch_size) def forward(self, x): B, C, H, W x.shape assert H self.img_size[0] and W self.img_size[1], \ f输入图像大小({H}*{W})不匹配模型期望的{self.img_size[0]}*{self.img_size[1]} x self.proj(x).flatten(2).transpose(1, 2) # (B, L, C) return x5.2 目标检测任务适配将RSWAtt集成到目标检测框架中class RSWAttFPN(nn.Module): 基于RSWAtt的特征金字塔网络 def __init__(self, in_channels_list, out_channels, window_sizes[7, 7, 7]): super().__init__() self.lateral_convs nn.ModuleList() self.fpn_convs nn.ModuleList() self.rswatt_blocks nn.ModuleList() for i, in_channels in enumerate(in_channels_list): lateral_conv nn.Conv2d(in_channels, out_channels, 1) fpn_conv nn.Conv2d(out_channels, out_channels, 3, padding1) rswatt_block EnhancedRSWAtt( dimout_channels, window_sizewindow_sizes[i], num_headsout_channels // 32 ) self.lateral_convs.append(lateral_conv) self.fpn_convs.append(fpn_conv) self.rswatt_blocks.append(rswatt_block) # 初始化权重 for m in self.modules(): if isinstance(m, nn.Conv2d): nn.init.kaiming_uniform_(m.weight, a1) if m.bias is not None: nn.init.constant_(m.bias, 0) def forward(self, inputs): Args: inputs: 不同尺度的特征图列表 Returns: outputs: 增强后的多尺度特征 # lateral连接 laterals [ lateral_conv(inputs[i]) for i, lateral_conv in enumerate(self.lateral_convs) ] # 自上而下的路径 for i in range(len(laterals) - 1, 0, -1): laterals[i - 1] F.interpolate( laterals[i], scale_factor2, modenearest ) # FPN卷积和RSWAtt增强 outputs [] for i, (lateral, fpn_conv, rswatt_block) in enumerate( zip(laterals, self.fpn_convs, self.rswatt_blocks) ): # 应用卷积 x fpn_conv(lateral) # 应用RSWAtt B, C, H, W x.shape x x.flatten(2).transpose(1, 2) # (B, H*W, C) x rswatt_block(x, H, W) x x.transpose(1, 2).reshape(B, C, H, W) outputs.append(x) return outputs5.3 语义分割任务应用在语义分割任务中集成RSWAttclass RSWAttUNet(nn.Module): 基于RSWAtt的UNet分割网络 def __init__(self, in_channels3, num_classes21, base_dim64, window_size7, depths[2, 2, 2, 2]): super().__init__() self.depths depths # 编码器 self.encoders nn.ModuleList() self.downsamplers nn.ModuleList() dim base_dim for i, depth in enumerate(depths): encoder_blocks nn.ModuleList([ RSWAttEncoderBlock(dim, window_size) for _ in range(depth) ]) self.encoders.append(encoder_blocks) if i len(depths) - 1: downsampler nn.Sequential( nn.Conv2d(dim, dim * 2, kernel_size3, stride2, padding1), nn.BatchNorm2d(dim * 2), nn.ReLU(inplaceTrue) ) self.downsamplers.append(downsampler) dim * 2 # 解码器 self.decoders nn.ModuleList() self.upsamplers nn.ModuleList() for i, depth in enumerate(reversed(depths[:-1])): upsampler nn.Sequential( nn.ConvTranspose2d(dim, dim // 2, kernel_size2, stride2), nn.BatchNorm2d(dim // 2), nn.ReLU(inplaceTrue) ) self.upsamplers.append(upsampler) dim // 2 decoder_blocks nn.ModuleList([ RSWAttDecoderBlock(dim, window_size) for _ in range(depth) ]) self.decoders.append(decoder_blocks) # 输出层 self.output_conv nn.Conv2d(base_dim, num_classes, kernel_size1) def forward(self, x): # 编码路径 skips [] for i, (encoder_blocks, downsampler) in enumerate( zip(self.encoders, self.downsamplers [None]) ): for block in encoder_blocks: x block(x) skips.append(x) if downsampler is not None: x downsampler(x) # 解码路径 for i, (upsampler, decoder_blocks, skip) in enumerate( zip(self.upsamplers, self.decoders, reversed(skips[:-1])) ): x upsampler(x) # 跳跃连接 x torch.cat([x, skip], dim1) for block in decoder_blocks: x block(x) return self.output_conv(x) class RSWAttEncoderBlock(nn.Module): RSWAtt编码器块 def __init__(self, dim, window_size): super().__init__() self.conv nn.Sequential( nn.Conv2d(dim, dim, 3, padding1), nn.BatchNorm2d(dim), nn.ReLU(inplaceTrue) ) self.rswatt EnhancedRSWAtt(dim, window_size, num_headsdim//32) def forward(self, x): # 卷积路径 conv_out self.conv(x) # RSWAtt路径 B, C, H, W x.shape x_flat x.flatten(2).transpose(1, 2) attn_out self.rswatt(x_flat, H, W) attn_out attn_out.transpose(1, 2).reshape(B, C, H, W) return conv_out attn_out6. 完整训练流程与代码实现6.1 训练配置与参数设置创建完整的训练配置类import yaml from dataclasses import dataclass from typing import Dict, Any dataclass class TrainingConfig: 训练配置参数 # 数据参数 data_path: str ./data batch_size: int 32 num_workers: int 4 image_size: int 224 # 模型参数 model_name: str rswatt_vit_small num_classes: int 1000 patch_size: int 16 embed_dim: int 384 depth: int 12 num_heads: int 6 window_size: int 7 mlp_ratio: float 4.0 # 优化器参数 learning_rate: float 1e-3 weight_decay: float 0.05 momentum: float 0.9 optimizer: str adamw # 学习率调度 scheduler: str cosine warmup_epochs: int 5 min_lr: float 1e-6 # 训练参数 epochs: int 300 eval_interval: int 1 save_interval: int 10 resume: str # 恢复训练的检查点路径 # 设备参数 device: str cuda if torch.cuda.is_available() else cpu amp: bool True # 自动混合精度 def to_dict(self) - Dict[str, Any]: return {k: v for k, v in self.__dict__.items() if not k.startswith(_)} classmethod def from_yaml(cls, yaml_path: str): with open(yaml_path, r) as f: config_dict yaml.safe_load(f) return cls(**config_dict) def save_to_yaml(self, yaml_path: str): with open(yaml_path, w) as f: yaml.dump(self.to_dict(), f, default_flow_styleFalse)6.2 完整的训练脚本import torch import torch.nn as nn import torch.optim as optim from torch.utils.data import DataLoader from torch.cuda.amp import GradScaler, autocast import time import os from pathlib import Path class RSWAttTrainer: RSWAtt模型训练器 def __init__(self, config: TrainingConfig): self.config config self.device torch.device(config.device) self.scaler GradScaler() if config.amp else None # 创建模型 self.model self._create_model() self.model.to(self.device) # 创建数据加载器 self.train_loader, self.val_loader self._create_dataloaders() # 创建优化器和调度器 self.optimizer self._create_optimizer() self.scheduler self._create_scheduler() # 损失函数 self.criterion nn.CrossEntropyLoss() # 训练状态 self.epoch 0 self.best_acc 0.0 self.train_losses [] self.val_accuracies [] # 创建输出目录 self.output_dir Path(./outputs) self.output_dir.mkdir(exist_okTrue) def _create_model(self): 创建RSWAtt模型 if self.config.model_name rswatt_vit_small: return RSWAttVisionTransformer( img_sizeself.config.image_size, patch_sizeself.config.patch_size, num_classesself.config.num_classes, embed_dimself.config.embed_dim, depthself.config.depth, num_headsself.config.num_heads, window_sizeself.config.window_size, mlp_ratioself.config.mlp_ratio ) else: raise ValueError(f未知模型: {self.config.model_name}) def _create_dataloaders(self): 创建数据加载器 # 这里使用伪数据集实际项目中替换为真实数据集 from torch.utils.data import TensorDataset # 创建伪训练数据 x_train torch.randn(1000, 3, self.config.image_size, self.config.image_size) y_train torch.randint(0, self.config.num_classes, (1000,)) train_dataset TensorDataset(x_train, y_train) # 创建伪验证数据 x_val torch.randn(200, 3, self.config.image_size, self.config.image_size) y_val torch.randint(0, self.config.num_classes, (200,)) val_dataset TensorDataset(x_val, y_val) train_loader DataLoader( train_dataset, batch_sizeself.config.batch_size, shuffleTrue, num_workersself.config.num_workers ) val_loader DataLoader( val_dataset, batch_sizeself.config.batch_size, shuffleFalse, num_workersself.config.num_workers ) return train_loader, val_loader def _create_optimizer(self): 创建优化器
