深度学习之Swin Transformer学习篇(详细 - 附代码）

一、摘要

        提出了一种新的vision Transformer，称为Swin Transformer，它可以作为计算机视觉的通用骨干。将Transformer从语言应用到视觉的挑战来自于这两个领域之间的差异，例如视觉实体规模的巨大差异以及与文本中的单词相比，图像中像素的高分辨率。为了解决这些差异，我们提出了一个分层的Transformer，它的表示是用移位窗口计算的。分层设计和移位窗口方法也被证明对所有mlp体系结构都是有益的。

---

 

二、Swin Transformer所解决的问题

• 超高分辨率的图像所带来的计算量：参考卷积网络的工作方式，获得全局注意力能力的同时，又将计算量从图像大小的平方关系降为线性关系，大大地减少了运算量，串联窗口自注意力运算(W-MSA)以及滑动窗口自注意力运算(SW-MSA)。
• 最初的Vision Transformer不具备多尺度预测：通过特征融合的方式PatchMerging（可参考卷积网络里的池化操作)，每次特征抽取之后都进行一次下采样，增加了下一次窗口注意力运算在原始图像上的感受野，从而对输入图像进行了多尺度的特征提取。
• 核心技术：SwinTransformer 针对ViT使用了“窗口”和“分层”的方式来替代长序列进行改进。

---

 

三、模型概述

 

3.1 模型结构

        首先将图片输入到Patch Partition模块中进行分块，即每4x4相邻的像素为一个Patch，然后在channel方向展平（flatten）。假设输入的是RGB三通道图片，那么每个patch就有4x4=16个像素，然后每个像素有R、G、B三个值所以展平后是16x3=48。通过Patch Partition后图像shape由 [H, W, 3]变成了 [H/4, W/4, 48]。然后在通过Linear Embeding层对每个像素的channel数据做线性变换，由48变成C，即图像shape再由 [H/4, W/4, 48]变成了 [H/4, W/4, C]。

        接着通过四个Stage构建不同大小的特征图，除了Stage1中先通过一个Linear Embeding层外，剩下三个stage都是先通过一个Patch Merging层进行下采样。重复堆叠Swin Transformer Block，注意这里的Block其实有两种结构，如图(b)中所示，这两种结构的不同之处仅在于一个使用了W-MSA结构，一个使用了SW-MSA结构。而且这两个结构是成对使用的，先使用一个W-MSA结构再使用一个SW-MSA结构，所以堆叠Swin Transformer Block的次数都是偶数。

 Patch Merging

        由图可见在除Stage1外的每个Stage中首先要通过一个Patch Merging层进行下采样。假设输入Patch Merging的是一个4x4大小的单通道特征图（feature map），Patch Merging会将每个2x2的相邻像素划分为一个patch，然后将每个patch中相同位置（同一颜色）像素给拼在一起就得到了4个feature map。接着将这四个feature map在深度方向进行concat拼接，接着通过一个LayerNorm层。最后通过一个全连接层在feature map的深度方向做线性变化，将feature map的深度由C变成C/2。可以看出，通过Patch Merging层后，feature map的高和宽减半，深度翻倍。

 

---

3.2 W-MSA

        Swin Transformer使用了类似卷积神经网络中的层次化构建方法（Hierarchical feature maps），比如特征图尺寸中有对图像下采样4倍的，8倍的以及16倍的，这样的backbone有助于在此基础上构建目标检测，实例分割等任务。而在之前的Vision Transformer中是一开始就直接下采样16倍，后面的特征图也是维持这个下采样率不变。在Swin Transformer中使用了Windows Multi-Head Self-Attention(W-MSA)的概念，比如在下图的4倍下采样和8倍下采样中，将特征图划分成了多个不相交的区域（Window），并且Multi-Head Self-Attention只在每个窗口内进行，这样做的目的是能够减少计算量的，尤其是在浅层特征图很大的时候。

 

3.3 SW-MSA

        采用W-MSA模块时，虽然减少了计算量但也会隔绝不同窗口之间的信息传递，所以在论文中作者又提出了 Shifted Windows Multi-Head Self-Attention(SW-MSA)的概念，通过此方法能够让信息在相邻的窗口中进行传递。如图所示，左侧使用的是刚刚讲的W-MSA（假设是第L层），那么根据之前介绍的W-MSA和SW-MSA是成对使用的，那么第L+1层使用的就是SW-MSA。根据左右两幅图对比能够发现窗口发生了偏移（可以理解成窗口从左上角分别向右侧和下方各偏移了 M/2 个像素）。看下偏移后的窗口，比如，第二行第二列的4x4的窗口，能够使第L层的四个窗口信息进行交流，其他的同理，那么这就解决了不同窗口之间无法进行信息交流的问题。

 

 

        通过将窗口进行偏移后，由原来的4个窗口变成9个窗口了。后续对每个窗口内部进行MSA的话又过于繁琐。为了解决这个问题，作者又提出而了Efficient batch computation for shifted configuration ，如下图所示，先将AC移至底部，再将BA移至最右侧。移动完后，A、B、C与初始模块重新组合，又重新变为4块4x4的窗口，所以能够保证计算量是相同的。

 

        但是由于不同的区域合并在一起进行MSA会导致信息错乱，为了防止这个问题，在实际计算中使用的是masked MSA，这样就能够通过设置蒙板来隔绝不同区域的信息了。关于mask如何使用，以的区域5和区域3为例。

        对于该窗口内的每一个像素（或称token，patch）在进行MSA计算时，都要先生成对应的query(q)，key(k)，value(v)。假设对于上图的像素0而言，得到q0后要与每一个像素的k进行匹配（match），假设α 0,0代表q 0 与像素0对应的k 0进行匹配的结果，那么同理可以得到α 0,0至α 0,15 。按照普通的MSA计算，接下来就是SoftMax操作了。但对于这里的masked MSA，像素0是属于区域5的，我们只想让它和区域5内的像素进行匹配。那么我们可以将像素0与区域3中的所有像素匹配结果都减去100（例如α 0 , 2 , α 0 , 3 , α 0 , 6 , α 0 , 7 等等），由于α的值都很小，一般都是零点几的数，将其中一些数减去100后再通过SoftMax得到对应的权重都等于0了。所以对于像素0而言实际上还是只和区域5内的像素进行了MSA。在计算完后还要把A，B，C给挪回到原来的位置上。

 

3.4 相对位置偏置

        

        论文中提到使用了相对位置偏置后能够带来明显的提升，下图为公式，与自注意力机制公式相比加了B（偏置）。

        如下图，假设输入的feature map高宽都为2，那么首先我们可以构建出每个像素的绝对位置（左下方的矩阵），对于每个像素的绝对位置是使用行号和列号表示的。比如蓝色的像素对应的是第0行第0列所以绝对位置索引是(0,0)，接下来再看看相对位置索引。首先看下蓝色的像素，在蓝色像素使用q与所有像素k进行匹配过程中，是以蓝色像素为参考点。然后用蓝色像素的绝对位置索引与其他位置索引进行相减，就得到其他位置相对蓝色像素的相对位置索引。例如黄色像素的绝对位置索引是(0,1)，则它相对蓝色像素的相对位置索引为(0,0)−(0,1)=(0,−1)，。那么同理可以得到其他位置相对蓝色像素的相对位置索引矩阵。同样，也能得到相对黄色，红色以及绿色像素的相对位置索引矩阵。接下来将每个相对位置索引矩阵按行展平，并拼接在一起可以得到下面的4x4矩阵 。

        这里描述的一直是相对位置索引，并不是相对位置偏执参数。因为后面我们会根据相对位置索引去取对应的参数。比如说黄色像素是在蓝色像素的右边，所以相对蓝色像素的相对位置索引为(0,−1)。绿色像素是在红色像素的右边，所以相对红色像素的相对位置索引为(0,−1)。可以发现这两者的相对位置索引都是(0,−1)，所以他们使用的相对位置偏执参数都是一样的。

        在源码中作者为了方便把二维索引给转成了一维索引。如果直接把行、列索引相加，例如上面的相对位置索引中有(0,−1)和(−1,0)，在二维的相对位置索引中明显是代表不同的位置，但如果简单相加都等于-1就出问题了。源码中首先在原始的相对位置索引上加上M-1(M为窗口的大小，在本示例中M=2)，加上之后索引中就不会有负数了。

接着将所有行标都乘上2M-1，最后将行标和列表进行相加，就不会出现上述问题了。 

        由于之前计算的是相对位置索引，并不是相对位置偏执参数。真正使用到的可训练参数B是保存在relative position bias table表里的，这个表的长度是等于(2M−1)×(2M−1)的。那么公式中的相对位置偏置参数B是根据上面的相对位置索引表根据查relative position bias table表得到的，如下图所示。

 

---

四、代码

        下图是原论文中给出的关于不同Swin Transformer的配置，T(Tiny)，S(Small)，B(Base)，L(Large)，其中：

• win. sz. 7x7表示使用的窗口（Windows）的大小
• dim表示feature map的channel深度（或者说token的向量长度）
• head表示多头注意力模块中head的个数

4.1 Swin Transformer

         展示 Swin Transformer 的整体架构，参数与上图对应。

class SwinTransformer(nn.Module):
    r""" Swin Transformer
        A PyTorch impl of : `Swin Transformer: Hierarchical Vision Transformer using Shifted Windows`  -
          https://arxiv.org/pdf/2103.14030

    Args:
        patch_size (int | tuple(int)): Patch size. Default: 4
        in_chans (int): Number of input image channels. Default: 3
        num_classes (int): Number of classes for classification head. Default: 1000
        embed_dim (int): Patch embedding dimension. Default: 96
        depths (tuple(int)): Depth of each Swin Transformer layer.
        num_heads (tuple(int)): Number of attention heads in different layers.
        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
        drop_rate (float): Dropout rate. Default: 0
        attn_drop_rate (float): Attention dropout rate. Default: 0
        drop_path_rate (float): Stochastic depth rate. Default: 0.1
        norm_layer (nn.Module): Normalization layer. Default: nn.LayerNorm.
        patch_norm (bool): If True, add normalization after patch embedding. Default: True
        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
    """
    # patch_size=4对应模型下采样4倍，即高宽/4
    def __init__(self, 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, mlp_ratio=4., qkv_bias=True,
                 drop_rate=0., attn_drop_rate=0., drop_path_rate=0.1,
                 norm_layer=nn.LayerNorm, patch_norm=True,
                 use_checkpoint=False, **kwargs):
        super().__init__()

        self.num_classes = num_classes
        self.num_layers = len(depths)
        self.embed_dim = embed_dim
        self.patch_norm = patch_norm
        # stage4输出特征矩阵的channels
        self.num_features = int(embed_dim * 2 ** (self.num_layers - 1))  #对应模型8C
        self.mlp_ratio = mlp_ratio

        # split image into non-overlapping patches将图片划分为一个个没有重叠的patches
        self.patch_embed = PatchEmbed(
            patch_size=patch_size, in_c=in_chans, embed_dim=embed_dim,
            norm_layer=norm_layer if self.patch_norm else None)
        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):
            # 注意这里构建的stage和论文图中有些差异
            # 这里的stage不包含该stage的patch_merging层，包含的是下个stage的
            layers = BasicLayer(dim=int(embed_dim * 2 ** i_layer),
                                depth=depths[i_layer],
                                num_heads=num_heads[i_layer],
                                window_size=window_size,
                                mlp_ratio=self.mlp_ratio,
                                qkv_bias=qkv_bias,
                                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(layers)

        self.norm = norm_layer(self.num_features)
        self.avgpool = nn.AdaptiveAvgPool1d(1)
        self.head = nn.Linear(self.num_features, num_classes) if num_classes > 0 else nn.Identity()

        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)

    def forward(self, x):
        # x: [B, L, C]
        x, H, W = self.patch_embed(x)
        x = self.pos_drop(x)

        for layer in self.layers:
            x, H, W = layer(x, H, W)

        x = self.norm(x)  # [B, L, C]
        x = self.avgpool(x.transpose(1, 2))  # [B, C, 1]
        x = torch.flatten(x, 1)
        x = self.head(x)
        return x

4.1.1 PatchEmbed

        将图片输入 Swin Transformer Block 前，需将图片划分成若干 patch tokens 并投影为嵌入向量。更具体地，将输入原始图片划分成一个个 patch_size * patch_size 大小的 patch token，然后投影嵌入。可通过将 2D 卷积层的 stride 和 kernel_size 的大小设为 patch_size，并将输出通道数设为 embed_dim 来实现投影嵌入。最后，展平并置换维度。

class PatchEmbed(nn.Module):
    """
    2D Image to Patch Embedding
    """
    def __init__(self, patch_size=4, in_c=3, embed_dim=96, norm_layer=None):
        super().__init__()
        patch_size = (patch_size, patch_size)
        self.patch_size = patch_size
        self.in_chans = in_c
        self.embed_dim = embed_dim
        self.proj = nn.Conv2d(in_c, embed_dim, kernel_size=patch_size, stride=patch_size)
        self.norm = norm_layer(embed_dim) if norm_layer else nn.Identity()

    def forward(self, x):
        _, _, H, W = x.shape

        # padding
        # 如果输入图片的H，W不是patch_size的整数倍，需要进行padding
        pad_input = (H % self.patch_size[0] != 0) or (W % self.patch_size[1] != 0)
        if pad_input:
            # to pad the last 3 dimensions,
            # (W_left, W_right, H_top,H_bottom, C_front, C_back)
            x = F.pad(x, (0, self.patch_size[1] - W % self.patch_size[1],
                          0, self.patch_size[0] - H % self.patch_size[0],
                          0, 0))

        # 下采样patch_size倍
        x = self.proj(x)
        _, _, H, W = x.shape
        # flatten: [B, C, H, W] -> [B, C, HW]，将hw展平
        # transpose: [B, C, HW] -> [B, HW, C]
        x = x.flatten(2).transpose(1, 2)
        x = self.norm(x)
        return x, H, W

4.1.2 PatchMerging

        在每个 Stage 前下采样缩小分辨率并减半通道数，从而形成层次化设计并降低运算量。

class PatchMerging(nn.Module):
    r""" 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):
        """
        x: B, H*W, C
        """
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size"

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

        # padding
        # 如果输入feature map的H，W不是2的整数倍，需要进行padding
        pad_input = (H % 2 == 1) or (W % 2 == 1)
        if pad_input:
            # to pad the last 3 dimensions, starting from the last dimension and moving forward.
            # (C_front, C_back, W_left, W_right, H_top, H_bottom)
            # 注意这里的Tensor通道是[B, H, W, C]，所以会和官方文档有些不同
            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)  # [B, H/2*W/2, 2*C]

        return x

4.1.3 BasicLayer

        Basic Layer 即 Swin Transformer 的各 Stage，包含了若干 Swin Transformer Blocks 及 其他层。一个 Stage 包含的 Swin Transformer Blocks 的个数必须是 偶数，因为需交替包含一个含有 Window Attention (W-MSA) 的 Layer 和含有 Shifted Window Attention (SW-MSA) 的 Layer。 

class BasicLayer(nn.Module):
    """
    A basic Swin Transformer layer for one stage.

    Args:
        dim (int): Number of input channels.
        depth (int): Number of blocks.
        num_heads (int): Number of attention heads.
        window_size (int): Local window size.
        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
        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,
                 mlp_ratio=4., qkv_bias=True, drop=0., attn_drop=0.,
                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False):
        super().__init__()
        self.dim = dim
        self.depth = depth
        self.window_size = window_size
        self.use_checkpoint = use_checkpoint
        self.shift_size = window_size // 2

        # 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 self.shift_size,
                mlp_ratio=mlp_ratio,
                qkv_bias=qkv_bias,
                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 create_mask(self, x, H, W):
        # calculate attention mask for SW-MSA
        # 保证Hp和Wp是window_size的整数倍
        Hp = int(np.ceil(H / self.window_size)) * self.window_size
        Wp = int(np.ceil(W / self.window_size)) * self.window_size
        # 拥有和feature map一样的通道排列顺序，方便后续window_partition
        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, Mh, Mw, 1]
        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)  # [nW, Mh*Mw]
        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)  # [nW, 1, Mh*Mw] - [nW, Mh*Mw, 1]
        # [nW, Mh*Mw, Mh*Mw]
        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
        return attn_mask

    def forward(self, x, H, W):
        attn_mask = self.create_mask(x, H, W)  # [nW, Mh*Mw, Mh*Mw]
        for blk in self.blocks:
            blk.H, blk.W = H, W
            if not torch.jit.is_scripting() and self.use_checkpoint:
                x = checkpoint.checkpoint(blk, x, attn_mask)
            else:
                x = blk(x, attn_mask)
        if self.downsample is not None:
            x = self.downsample(x, H, W)
            H, W = (H + 1) // 2, (W + 1) // 2

        return x, H, W

4.1.3.1 creat_mask

    def create_mask(self, x, H, W):
        # calculate attention mask for SW-MSA
        # 保证Hp和Wp是window_size的整数倍
        Hp = int(np.ceil(H / self.window_size)) * self.window_size
        Wp = int(np.ceil(W / self.window_size)) * self.window_size
        # 拥有和feature map一样的通道排列顺序，方便后续window_partition
        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, Mh, Mw, 1]
        mask_windows = mask_windows.view(-1, self.window_size * self.window_size)  # [nW, Mh*Mw]
        attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)  # [nW, 1, Mh*Mw] - [nW, Mh*Mw, 1]
        # [nW, Mh*Mw, Mh*Mw]
        attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
        return attn_mask

window_partition

def window_partition(x, window_size: int):
    """
    将feature map按照window_size划分成一个个没有重叠的window
    Args:
        x: (B, H, W, C)
        window_size (int): window size(M)

    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)
    # permute: [B, H//Mh, Mh, W//Mw, Mw, C] -> [B, H//Mh, W//Mh, Mw, Mw, C]
    # view: [B, H//Mh, W//Mw, Mh, Mw, C] -> [B*num_windows, Mh, Mw, C]
    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
    return windows

window_reverse

def window_reverse(windows, window_size: int, H: int, W: int):
    """
    将一个个window还原成一个feature map
    Args:
        windows: (num_windows*B, window_size, window_size, C)
        window_size (int): Window size(M)
        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))
    # view: [B*num_windows, Mh, Mw, C] -> [B, H//Mh, W//Mw, Mh, Mw, C]
    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
    # permute: [B, H//Mh, W//Mw, Mh, Mw, C] -> [B, H//Mh, Mh, W//Mw, Mw, C]
    # view: [B, H//Mh, Mh, W//Mw, Mw, C] -> [B, H, W, C]
    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, H, W, -1)
    return x

4.1.3.2 SwinTransformerBlock

class SwinTransformerBlock(nn.Module):
    r""" 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
        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, 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=(self.window_size, self.window_size), num_heads=num_heads, qkv_bias=qkv_bias,
            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)

    def forward(self, x, attn_mask):
        H, W = self.H, self.W
        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)

        # pad feature maps to multiples of window size
        # 把feature map给pad到window size的整数倍
        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))
        else:
            shifted_x = x
            attn_mask = None

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

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

        # merge windows
        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)  # [nW*B, Mh, Mw, 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:
            # 把前面pad的数据移除掉
            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

Mlp

使用 GELU 激活函数 + Dropout 的两层 FCs。 

class Mlp(nn.Module):
    """ MLP as used in Vision Transformer, MLP-Mixer and related networks
    """
    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.drop1 = nn.Dropout(drop)
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop2 = nn.Dropout(drop)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop1(x)
        x = self.fc2(x)
        x = self.drop2(x)
        return x

WindowAttention

比较复杂，具体可参考

【12.2 使用Pytorch搭建Swin-Transformer网络】https://www.bilibili.com/video/BV1yg411K7Yc?vd_source=5feb724f579d85f69271c141d448d95a 视频54：43处

或

【深度学习】详解 Swin Transformer (SwinT)-CSDN博客  中window attention代码解释（非常详细）

class WindowAttention(nn.Module):
    r""" 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
        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, attn_drop=0., proj_drop=0.):

        super().__init__()
        self.dim = dim
        self.window_size = window_size  # [Mh, Mw]
        self.num_heads = num_heads
        head_dim = dim // num_heads
        self.scale = 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*Mh-1 * 2*Mw-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], indexing="ij"))  # [2, Mh, Mw]
        coords_flatten = torch.flatten(coords, 1)  # [2, Mh*Mw]
        # [2, Mh*Mw, 1] - [2, 1, Mh*Mw]
        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # [2, Mh*Mw, Mh*Mw]
        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # [Mh*Mw, Mh*Mw, 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)  # [Mh*Mw, Mh*Mw]
        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)

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

    def forward(self, x, mask: Optional[torch.Tensor] = None):
        """
        Args:
            x: input features with shape of (num_windows*B, Mh*Mw, C)
            mask: (0/-inf) mask with shape of (num_windows, Wh*Ww, Wh*Ww) or None
        """
        # [batch_size*num_windows, Mh*Mw, total_embed_dim]
        B_, N, C = x.shape
        # qkv(): -> [batch_size*num_windows, Mh*Mw, 3 * total_embed_dim]
        # reshape: -> [batch_size*num_windows, Mh*Mw, 3, num_heads, embed_dim_per_head]
        # permute: -> [3, batch_size*num_windows, num_heads, Mh*Mw, embed_dim_per_head]
        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4)
        # [batch_size*num_windows, num_heads, Mh*Mw, embed_dim_per_head]
        q, k, v = qkv.unbind(0)  # make torchscript happy (cannot use tensor as tuple)

        # transpose: -> [batch_size*num_windows, num_heads, embed_dim_per_head, Mh*Mw]
        # @: multiply -> [batch_size*num_windows, num_heads, Mh*Mw, Mh*Mw]
        q = q * self.scale
        attn = (q @ k.transpose(-2, -1))

        # relative_position_bias_table.view: [Mh*Mw*Mh*Mw,nH] -> [Mh*Mw,Mh*Mw,nH]
        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)
        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # [nH, Mh*Mw, Mh*Mw]
        attn = attn + relative_position_bias.unsqueeze(0)

        if mask is not None:
            # mask: [nW, Mh*Mw, Mh*Mw]
            nW = mask.shape[0]  # num_windows
            # attn.view: [batch_size, num_windows, num_heads, Mh*Mw, Mh*Mw]
            # mask.unsqueeze: [1, nW, 1, Mh*Mw, Mh*Mw]
            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)

        # @: multiply -> [batch_size*num_windows, num_heads, Mh*Mw, embed_dim_per_head]
        # transpose: -> [batch_size*num_windows, Mh*Mw, num_heads, embed_dim_per_head]
        # reshape: -> [batch_size*num_windows, Mh*Mw, total_embed_dim]
        x = (attn @ v).transpose(1, 2).reshape(B_, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x

参考链接：

AI大模型系列之三：Swin Transformer 最强CV图解（深度好文）_cv大模型-CSDN博客 

Swin-Transformer网络结构详解_swin transformer-CSDN博客

【12.1 Swin-Transformer网络结构详解】https://www.bilibili.com/video/BV1pL4y1v7jC?vd_source=5feb724f579d85f69271c141d448d95a

Swin Transformer代码详解（必懂）-CSDN博客