Swin Transformer: Hierarchical Vision Transformer using Shifted Windows复现

### Swin Transformer 的复现教程 #### 1. 模型概述 Swin Transformer 是一种分层视觉变换器 (Hierarchical Vision Transformer)，它通过滑动窗口机制构建局部表示并支持跨窗口连接[^1]。该模型的核心组件包括分层设计、移位窗口 (Shifted Window) 和自注意力机制。 #### 2. 数据预处理 数据预处理阶段涉及将输入图像划分为多个 patch，并将其映射到 token 序列中。具体过程如下： - 将输入图像切分成大小为 \(P \times P\) 的 patches。 - 对每个 patch 进行线性嵌入操作，得到初始的 token 表示。 - 使用卷积下采样层进一步减少空间分辨率，形成多尺度特征图。 此阶段通常称为“阶段 1”，其中 transformer 块的数量为 \(H/4 \times W/4\)，即每张图片被分解成若干 tokens[^2]。 #### 3. 移位窗口机制 为了提高效率和建模能力，Swin Transformer 引入了移位窗口策略。在标准窗口划分的基础上，每隔一层会调整窗口的位置以引入交叉窗口的信息交互。这种方法显著提升了性能，在 ImageNet-1K 图像分类任务上 top-1 准确率提高了 +1.1%，而在 COCO 目标检测任务中则分别提升 +2.8 box AP 和 +2.2 mask AP[^4]。 #### 4. PyTorch 实现代码 以下是基于 PyTorch 的 Swin Transformer 核心模块实现： ```python import torch from torch import nn class PatchEmbed(nn.Module): """Patch Embedding Layer""" def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96): super().__init__() self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) def forward(self, x): x = self.proj(x).flatten(2).transpose(1, 2) return x class Mlp(nn.Module): """Multilayer Perceptron""" def __init__(self, in_features, hidden_features=None, out_features=None): 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 = nn.GELU() self.fc2 = nn.Linear(hidden_features, out_features) def forward(self, x): x = self.fc1(x) x = self.act(x) x = self.fc2(x) return x class WindowAttention(nn.Module): """Window-based Multi-head Self Attention (MSA) module with relative position bias.""" def __init__(self, dim, window_size, num_heads): super().__init__() self.dim = dim self.window_size = window_size self.num_heads = num_heads head_dim = dim // num_heads self.scale = head_dim ** -0.5 self.qkv = nn.Linear(dim, dim * 3, bias=True) self.attn_drop = nn.Dropout(0.) self.proj = nn.Linear(dim, dim) def forward(self, x): 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.unbind(0) attn = (q @ k.transpose(-2, -1)) * self.scale attn = attn.softmax(dim=-1) attn = self.attn_drop(attn) x = (attn @ v).transpose(1, 2).reshape(B_, N, C) x = self.proj(x) return x class SwinTransformerBlock(nn.Module): """Swin Transformer Block""" def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0): super().__init__() self.input_resolution = input_resolution self.window_size = window_size self.shift_size = shift_size if min(self.input_resolution) <= self.window_size: self.shift_size = 0 self.window_size = min(self.input_resolution) self.norm1 = nn.LayerNorm(dim) self.attn = WindowAttention( dim, window_size=(self.window_size, self.window_size), num_heads=num_heads ) self.norm2 = nn.LayerNorm(dim) mlp_hidden_dim = int(dim * 4) self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim) def forward(self, x): H, W = self.input_resolution B, L, C = x.shape assert L == H * W, "input feature has wrong size" shortcut = x x = self.norm1(x) x = x.view(B, H, W, C) # cyclic shift if self.shift_size > 0: shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2)) else: shifted_x = x # partition windows x_windows = shifted_x.unfold(1, self.window_size, self.window_size)\ .unfold(2, self.window_size, self.window_size) x_windows = x_windows.contiguous().view(-1, self.window_size*self.window_size, C) # attention and projection attn_windows = self.attn(x_windows) attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C) # reverse windows shifted_x = attn_windows.permute(0, 1, 2, 3).contiguous().view(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 x = x.view(B, H*W, C) # FFN x = shortcut + x x = x + self.mlp(self.norm2(x)) return x class BasicLayer(nn.Module): """A basic Swin Transformer layer for one stage.""" def __init__(self, dim, depth, num_heads, window_size=7): super().__init__() self.blocks = nn.ModuleList([ SwinTransformerBlock( dim=dim, input_resolution=(window_size, window_size), num_heads=num_heads, window_size=window_size, shift_size=0 if i % 2 == 0 else window_size // 2 ) for i in range(depth)]) def forward(self, x): for blk in self.blocks: x = blk(x) return x class SwinTransformer(nn.Module): """Overall architecture of the Swin Transformer model.""" def __init__(self, img_size=224, patch_size=4, in_chans=3, num_classes=1000, embed_dim=96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24], window_size=7): super().__init__() self.patch_embed = PatchEmbed(img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim) dpr = [x.item() for x in torch.linspace(0, 0.1, sum(depths))] self.layers = nn.ModuleList() for i_layer in range(len(depths)): layer = BasicLayer( dim=int(embed_dim * 2 ** i_layer), depth=depths[i_layer], num_heads=num_heads[i_layer], window_size=window_size ) self.layers.append(layer) self.norm = nn.LayerNorm(int(embed_dim * 2 ** (len(depths)-1))) self.avgpool = nn.AdaptiveAvgPool1d(1) self.head = nn.Linear(int(embed_dim * 2 ** (len(depths)-1)), num_classes) if num_classes > 0 else nn.Identity() def forward(self, x): x = self.patch_embed(x) for layer in self.layers: x = layer(x) x = self.norm(x.mean(1)) x = self.head(x) return x ``` #### 5. 训练与验证流程 训练过程中可以使用常见的优化算法（如 Adam 或 SGD），并通过学习率调度器动态调整超参数。对于下游任务（如目标检测或语义分割），可以通过微调预训练权重来加速收敛。 ---