SwinTransformer整体网络结构代码详解

SwinTransfomer在发布后，赢得了广泛关注，本人也曾用该模型进行实验，发现该模型的确有较好的结果。这里主要结合论文和代码中的网络结构进行详细解析。

论文中的网络结构

其中SwinTransformer Block的结构为，即为一个标准W-MSA和SW-MSA的级联：

网络结构的细节

关于网络整体的细节，在文中的第三部分有非常详细的说明，这里先把文章中的相关内容截下，然后再结合代码详细展开叙述。

1 关于网络中的输入图像预处理部分，该部分对应的网络整体架构图如下面的截图。

所对应的代码如下：

class PatchEmbed(nn.Module):
    r""" Image to Patch Embedding

    Args:
        img_size (int): Image size.  Default: 224.
        patch_size (int): Patch token size. Default: 4.
        in_chans (int): Number of input image channels. Default: 3.
        embed_dim (int): Number of linear projection output channels. Default: 96.
        norm_layer (nn.Module, optional): Normalization layer. Default: None
    """

    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
        super().__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
        self.img_size = img_size
        self.patch_size = patch_size
        self.patches_resolution = patches_resolution
        self.num_patches = patches_resolution[0] * patches_resolution[1]
        self.in_chans = in_chans
        self.embed_dim = embed_dim

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

    def forward(self, x):
        B, C, H, W = x.shape
        # FIXME look at relaxing size constraints
        assert H == self.img_size[0] and W == self.img_size[1], \
            f"Input image size ({H}*{W}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
        x = self.proj(x).flatten(2).transpose(1, 2)  # B Ph*Pw C
        if self.norm is not None:
            x = self.norm(x)
        return x

即代码中在输入端使用了 4 x 4 且步长为4的卷积，则假设原输入为224 x 224，经过此过程后，图像变为56 x 56的大小，此过程没有用非线性层，所以文中说的是将图像分为 4 x 4 的小图像块，然后进行线性变换，得到图像的嵌入特征向量。

2 SwinTransformerBlock

论文中给出的 SwinTransformer模块，在图的标注部分，说的非常明确，这个是两个SwinTransformer Block的连接，其中不同点在于前一个阶段是 W-MSA模块,而后一个是 SW-MSA模块，则可以看出，SwinTransfomer Block包含了这两种模块，关于SwinBlock的说明，在后续的文章内容中也有说明。但是如果不细读文章或者细看代码，很可能会认为在单个SwinBlock中包含了这两个部分，其实不是的，这是两个级联的SwinBlock。

接下来结合代码来看一下：(BasicLayer方法中的构建SwinBlock的部分，大家可以看到，其实W-MSA和SW-MSA是交替出现的，具体通过shift_size这个变量进行选择)

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

SwinTransformerBlock方法(这个应该是本论文中的重点部分之一了)：

class SwinTransformerBlock(nn.Module):
    r""" Swin Transformer Block.

    Args:
        dim (int): Number of input channels.
        input_resolution (tuple[int]): Input resulotion.
        num_heads (int): Number of attention heads.
        window_size (int): Window size.
        shift_size (int): Shift size for SW-MSA.
        mlp_ratio (float): Ratio of mlp hidden dim to embedding dim.
        qkv_bias (bool, optional): If True, add a learnable bias to query, key, value. Default: True
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set.
        drop (float, optional): Dropout rate. Default: 0.0
        attn_drop (float, optional): Attention dropout rate. Default: 0.0
        drop_path (float, optional): Stochastic depth rate. Default: 0.0
        act_layer (nn.Module, optional): Activation layer. Default: nn.GELU
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
        fused_window_process (bool, optional): If True, use one kernel to fused window shift & window partition for acceleration, similar for the reversed part. Default: False
    """

    def __init__(self, dim, input_resolution, num_heads, window_size=7, shift_size=0,
                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0., drop_path=0.,
                 act_layer=nn.GELU, norm_layer=nn.LayerNorm,
                 fused_window_process=False):
        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.num_heads = num_heads
        self.window_size = window_size
        self.shift_size = shift_size
        self.mlp_ratio = mlp_ratio
        if min(self.input_resolution) <= self.window_size:
            # if window size is larger than input resolution, we don't partition windows
            self.shift_size = 0
            self.window_size = min(self.input_resolution)
        assert 0 <= self.shift_size < self.window_size, "shift_size must in 0-window_size"

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

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

        if self.shift_size > 0:
            # calculate attention mask for SW-MSA
            H, W = self.input_resolution
            img_mask = torch.zeros((1, H, W, 1))  # 1 H W 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:
                    #相当于将 img_mask分了9块
                    img_mask[:, h, w, :] = cnt
                    cnt += 1
            #num_windows*1, window_size, window_size, 1
            mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
            mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
            #(num_windows,window_size*window_size)
            # ->(num_windows,1,window_size*window_size)-(num_windows,window_size*window_size,1)
            #->(num_windows,window_size*window_size,window_size*window_size)
            attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
            attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))
        else:
            attn_mask = None

        self.register_buffer("attn_mask", attn_mask)
        self.fused_window_process = fused_window_process

    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:
            if not self.fused_window_process:
                shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
                # partition windows
                x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
            else:
                x_windows = WindowProcess.apply(x, B, H, W, C, -self.shift_size, self.window_size)
        else:
            shifted_x = x
            # partition windows
            x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C

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

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

        # merge windows
        attn_windows = attn_windows.view(-1, self.window_size, self.window_size, C)

        # reverse cyclic shift
        if self.shift_size > 0:
            if not self.fused_window_process:
                shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
                x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
            else:
                x = WindowProcessReverse.apply(attn_windows, B, H, W, C, self.shift_size, self.window_size)
        else:
            shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
            x = shifted_x
        x = x.view(B, H * W, C)
        x = shortcut + self.drop_path(x)

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

        return x

其实大家先看该方法中的 forward 函数，可以大致推导出SwinTranformer的计算流程，其实是和论文描述一致的：

我们需要关注的是，在该部分到底是如何进行 W-MSA和SW-MSA过程的。
首先看该方法中的 WindowAttention 方法,当然这个是多头注意力机制模块(有点类似于分组卷积)，此外这段代码中包含了论文中提到 relative position bias

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
        qk_scale (float | None, optional): Override default qk scale of head_dim ** -0.5 if set
        attn_drop (float, optional): Dropout ratio of attention weight. Default: 0.0
        proj_drop (float, optional): Dropout ratio of output. Default: 0.0
    """

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

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

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

        # get pair-wise relative position index for each token inside the window
        coords_h = torch.arange(self.window_size[0])
        coords_w = torch.arange(self.window_size[1])
        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
        #49个位置的值，分别于自身的49个位置值做差，得到相对位置坐标
        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
        relative_coords[:, :, 1] += self.window_size[1] - 1
        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
        #值从 0 - 168(156=(6+6)*13+12(6+6)) 主对角线上的值为 84
        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
        self.register_buffer("relative_position_index", relative_position_index)

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
        self.attn_drop = nn.Dropout(attn_drop)
        self.proj = nn.Linear(dim, dim)
        self.proj_drop = nn.Dropout(proj_drop)

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

    def forward(self, x, mask=None):
        """
        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)
        #(B,num_heads,N,C // self.num_heads)
        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)

        q = q * self.scale
        attn = (q @ k.transpose(-2, -1))#(B,num_heads,N,N) N = 49
        # 得到Wh*Ww,Wh*Ww,nH每个坐标位置的相对位置偏置量(越靠近中心值越大)
        relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
            self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)  # Wh*Ww,Wh*Ww,nH
        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
        attn = attn + relative_position_bias.unsqueeze(0)#(B,num_heads,N,N)

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

        attn = self.attn_drop(attn)

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

如果理解相对位置编码的是实现，可以看该部分。其中 relative_position_bias_table 存储的是可学习[(2 * window_size[0] - 1) * (2 * window_size[1] - 1)]*num_heads 的位置编码值，需要注意的是该编码值为截断正态随机采样，整体的值分布是随机的，没有像正态分布那样在中间大两边小的情况，因为该张量为可学习的，relative_position_index 是给定窗口大小的相对位置索引，其索引长度为（window_size[0] * window_size[1]）x（window_size[0] * window_size[1），该大小对应窗口自注意力模块形成的尺寸大小，每个索引位的取值为范围为 [0,(2 * window_size[0] - 1) x (2 * window_size[1] - 1)-1]。

 self.relative_position_bias_table = nn.Parameter(
            torch.zeros((2 * window_size[0] - 1) * (2 * window_size[1] - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH

        # get pair-wise relative position index for each token inside the window
        coords_h = torch.arange(self.window_size[0])
        coords_w = torch.arange(self.window_size[1])
        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
        #49个位置的值，分别于自身的49个位置值做差，得到相对位置坐标
        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
        relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
        relative_coords[:, :, 0] += self.window_size[0] - 1  # shift to start from 0
        relative_coords[:, :, 1] += self.window_size[1] - 1
        relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1
        #值从 0 - 168(156=(6+6)*13+12(6+6)) 主对角线上的值为 84
        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
        self.register_buffer("relative_position_index", relative_position_index)
        trunc_normal_(self.relative_position_bias_table, std=.02)

其实在窗口注意力的代码中，还有和 mask 相关的代码，这个mask就是为了配合 shift_size 生成的，目的是为了实现在同一窗口下，实现不同偏移量子窗口的注意力机制，这个在文中有相关的说明。

if mask is not None:
            nW = mask.shape[0]#n_windows
            #mask：num_windows, Wh*Ww, Wh*Ww-> 1, num_windows, 1,Wh*Ww, Wh*Ww
            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)

接下来看一下这个 mask 的实现，可以通过代码发现，mask是在 img_mask这个母版上移动得到的，我们可以根据论文代码中给定的关系，假设 H=W=14 window_size=7 shift_size=3，则会划分出 2*2=4个子窗口。

if self.shift_size > 0:
            # calculate attention mask for SW-MSA
            H, W = self.input_resolution
            img_mask = torch.zeros((1, H, W, 1))  # 1 H W 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:
                    #相当于将 img_mask分了9块
                    img_mask[:, h, w, :] = cnt
                    cnt += 1
            #num_windows*1, window_size, window_size, 1
            mask_windows = window_partition(img_mask, self.window_size)  # nW, window_size, window_size, 1
            mask_windows = mask_windows.view(-1, self.window_size * self.window_size)
            #(num_windows,window_size*window_size)
            # ->(num_windows,1,window_size*window_size)-(num_windows,window_size*window_size,1)
            #->(num_windows,window_size*window_size,window_size*window_size)
            attn_mask = mask_windows.unsqueeze(1) - mask_windows.unsqueeze(2)
            attn_mask = attn_mask.masked_fill(attn_mask != 0, float(-100.0)).masked_fill(attn_mask == 0, float(0.0))

则此时 img_mask的输出为

上图的 img_mask 经过 window_partition 之后，变为

这个基本与文中对应，为什么要形成这样的mask模板，这个可以继续往下看。

然后对经过partition之后的窗口进行最终输出4个窗口注意力的掩码图，这4个掩码图对应输入张量(B,C,H,W)切分窗口后对应的子窗口，实现子窗口自注意力机制(变换后的特征图对应位置之间的注意力计算)的计算，从而实现整个输入张量自注意力机制的计算。以下展示的是最终的 mask 二值图的可视化。

与mask相关的是 SW-MSA模块，这一部分代码是关键代码，即在已经计算过(q x k)且加入相对位置偏移的算子中加上 mask, 即实现了注意力机制是在自身对应位置计算的，不对应的位置加了一个比较大负值(-100)，则经过softmax后，这部分的分值就变低了，将最终的分支与 v 变量进行相乘，基本可以实现遮挡与想要计算位置不同位置特征的目的。

if mask is not None:
            nW = mask.shape[0]#n_windows
            #mask：num_windows, Wh*Ww, Wh*Ww-> 1, num_windows, 1,Wh*Ww, Wh*Ww
            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)
x = self.proj(x)
x = self.proj_drop(x)

以上基本是本文中子注意力模块计算的过程，当然我们主要讲的是 W-MSA，由于文中还有个模块为 SW-MSA,在进行SW-MSA之前，其实还需要注意的一点是，将原始输入的特征图进行适当的循环移位，但值得注意的是，在代码中 shift_size在任何一个阶段均为 (window_size//2)，由于window_size为定值，所以 shift_size也为定值，论文内容和代码如下：


下面代码的效果对应上面的图。

# cyclic shift
        if self.shift_size > 0:
            if not self.fused_window_process:
                shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
                # partition windows
                x_windows = window_partition(shifted_x, self.window_size)  # nW*B, window_size, window_size, C
            else:
                x_windows = WindowProcess.apply(x, B, H, W, C, -self.shift_size, self.window_size)
        else:
            shifted_x = x
            # partition windows
            x_windows = window_partition(shifted_x, self.window_size)

计算完 SW-MSA后，会将该窗口重新复原

# reverse cyclic shift
        if self.shift_size > 0:
            if not self.fused_window_process:
                shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # B H' W' C
                x = torch.roll(shifted_x, shifts=(self.shift_size, self.shift_size), dims=(1, 2))
            else:
                x = WindowProcessReverse.apply(attn_windows, B, H, W, C, self.shift_size, self.window_size)

以上就是SwinTransformerBlock模块做的事情了。下面再说下本文的一种池化机制：PatchMerging。

池化机制 PatchMerging

这个模块主要是进行特征图的池化操作，与卷积神经网路中比较直接的池化策略不同，这里采用的 横纵间隔抽取特征图中的像素点，物理上将特征图由(H,W)变为 (H/2,W/2)，再经过一个全连接进行降维(4C->2C)，最终将(H,W,C)的输入变为了(H/2,W/2,2C)的输出，实现2倍缩放的池化操作。

class PatchMerging(nn.Module):
    r""" Patch Merging Layer.

    Args:
        input_resolution (tuple[int]): Resolution of input feature.
        dim (int): Number of input channels.
        norm_layer (nn.Module, optional): Normalization layer.  Default: nn.LayerNorm
    """

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

    def forward(self, x):
        """
        x: B, H*W, C
        """
        H, W = self.input_resolution
        B, L, C = x.shape
        assert L == H * W, "input feature has wrong size"
        assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."

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

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

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

        return x

值得注意的是，文中图示部分是将Patch Merging模块与SwinBlock模块组合为stage 2的，其实在代码中，Patch Merging 为 stage1的最后一步，但这个其实与文中图示不冲突。

总结

SwinTransformer的代码确实值得好好看一下，这样文中的技术才会较全面的理解。码字不易，如有问题，欢迎各位留言。