对Swin-Transformer的理解

        Swin-Transformer是微软近期提出将Transformer用到密集图像预测任务中所提出的一种可作为骨干的Transformer骨干，目前正在各大CV领域疯狂屠榜！

Introduction

        目前Transformer应用到图像领域主要有两大挑战：

• 视觉目标变化大，在不同的场景下视觉Transfrmer（如VIT）的性能未必好
• 图像分辨率高，像素点多，Transformer基于全局自注意力的计算导致计算量比较大
  针对上述问题 Swin-Transformer架构被提出，这种架构包含滑窗操作，具有层级设计，其中滑窗操作包括不重叠的local window，和重叠的cross-window。在窗口中计算各自的注意力，这样做的好处是既能引入CNN卷积操作的局限性，另一方面能节省计算量。
  Swin-Transformer

整体结构

        整个模型采取层次化的设计（主流做法），一共有4个Stage，每个Stage通过Patch Merging来缩小输入特征图的分辨率（这点和CNN一样通过逐层来扩大感受野）。

• 在输入开始，做一个Patch Embedding/Patch Partition，将图片切成一个个小块，并嵌入到Embedding。
• 在每个Stage中，由Patch Merging和多个Block（上图右侧） 组成。
• 其中Patch Merging模块的作用是在每个Stage开头来降低图片分辨率。
• Block 块如图右所示，主要由LN(LayerNnorm)，W MSA(Window Attention)，MLP,SW MSA(Shifted Window Attention)

class SwinTransformer(nn.Module):
    def __init__(self, img_size=224, patch_size=4, in_chans=3, num_class=1000,
                 embed_dim = 96, depths=[2, 2, 6, 2], num_heads=[3, 6, 12, 24],
                 window_size=7, mlp_ratio = 4, qkv_bias = True, qk_scale=None,
                 drop_rate = 0., attn_drop_rate=0., drop_path_rate=0.1,
                 norm_layer=nn.LayerNorm, ape = False, patch_norm=True,
                 use_checkpoint=False, **kwargs):
        super(SwinTransformer, self).__init__()
        self.num_class = num_class
        self.num_layers = len(depths)
        self.embed_dim = embed_dim
        self.ape = ape
        self.patch_norm = patch_norm
        self.num_features = int(embed_dim * 2 ** (self.num_layers - 1))
        self.mlp_ratio = mlp_ratio

        #split image into non-overlapping patches
        self.patch_embed = PatchEmbed(
            img_size=img_size, patch_size=patch_size, in_chans=in_chans, embed_dim=embed_dim,
            norm_layer=norm_layer if self.patch_norm else None)
        num_patches = self.patch_embed.num_patches
        patches_resolution = self.patch_embed.patchs_resolution
        self.patches_resloution = patches_resolution

        # absolute position embedding
        if self.ape:
            self.absolute_pos_embed = nn.Parameter(torch.zeros(1, num_patches, embed_dim))
            trunc_normal_(self.absolute_pos_embed, std=.02)
        self.pos_drop = nn.Dropout(p=drop_rate)

        # stochastic depth 随机深度
        dpr = [x.item() for x in torch.linspace(0, drop_path_rate, sum(depths))]

        # build layer
        self.layers = nn.ModuleList()
        for i_layer in range(self.num_layers):
            layer = BasicLayer(dim = int(embed_dim * 2 ** i_layer),
                               input_resolution = (patches_resolution[0] // (2 ** i_layer),
                                                 patches_resolution[1] // (2 ** i_layer)),
                               depth = depths[i_layer],
                               num_heads = num_heads[i_layer],
                               window_size = window_size,
                               mlp_ratio = mlp_ratio,
                               qkv_bias = qkv_bias,qk_scale=qk_scale,
                               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.appens(layer)
        self.norm = norm_layer(self.num_features)
        self.avgpool = nn.AdaptiveAvgPool1d(1)
        self.apply(self._init_weights)

    def _int_weights(self,m):
        if isinstance(m, nn.Linear):
            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_features(self, x):
        x = self.patch_embed(x)
        if self.ape:
            x += self.absolute_pos_embed
        x = self.pos_drop(x)

        for layer in self.layers:
            x = layer(x)
        x = self.norm(x)
        x = self.avgpool(x.transpose(1, 2))
        x = torch.flatten(x, 1)
        return x

    def forward(self, x):
        x = self.forward_features(x)
        return x

        与VIT的差别之处：

1. VIT在输入会给Embedding（Patch Embeding前）进行位置编码。而Swin-Transformer中的此处位置编码是个可选项（self.ape=True/False）,Swin-Transformer的位置编码是在计算Attention的时候做了个相对位置编码

Patch Embedding(Partition)

        在将图片输入进Block之前，需要将图片切成一个个Patch，然后嵌入向量。
具体做法：将原始图片裁成一个个window_size * window_size的窗口大小，然后进行嵌入向量，这里的做法可以使用二维卷积层，将stride，kernel-size，设置为window_size大小。设定输出通道来确定嵌入向量的大小，然后将H,W维度展开，并移到第一维度

class PatchEmbed(nn.Module):
    def __init__(self, img_size=224, patch_size=4, in_chans=3, embed_dim=96, norm_layer=None):
        super(PatchEmbed, self).__init__()
        img_size = to_2tuple(img_size)
        patch_size = to_2tuple(patch_size)
        patchs_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]]
        self.img_size = img_size
        self.patch_size = patch_size
        self.patchs_resolution = patchs_resolution
        self.num_patches = patchs_resolution[0] * patchs_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
        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)
        # x的维度变化
        # (B,96,224/4,224/4)
        # (B,96,56*56)
        # (B,56*56,96)
        if self.norm is not None:
            x = self.norm(x)
        return x

Patch Merging

        此模块的作用是在每个Stage之前进行下采样，用于缩小分辨率，调整通道数，进而形成层次化的设计，同时也能节省一定的计算量。

CNN中，使用Stride>1的卷积来下采样

        每次降采样都是2倍，因此在行和列方向上间隔2选取元素。
        然后拼接到一起作为一整个张量，最后展开。此时通道位数会变成原来的4倍(保持总的不变)，此时在通过一个全连接层再调整通道维度为原来的2倍。

class PatchMerging(nn.Module):
    def __init__(self, input_resolution, dim, norm_layer=nn.LayerNorm):
        super(self).__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 nor even(偶数)"

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

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

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

        return x

Window Partition/Reverse

        Window partition函数用于对张量划分窗口，指定窗口大小。将原本的张量变成num_windows * B，Window_size,Window_size,C,其中 num_windows= (H//Window_size) * (W/Window_size)，也就是窗口的个数，而Window_reverse则是对应的逆过程。

def window_partition(x, window_size):
    B, H, W, C = x.shape
    x = x.view(B, H//window_size, window_size, W//window_size, window_size, C)
    windows = x.permute(0,1,3,2,4,5).contiguous().view(-1, window_size, window_size, C)
    return windows

def window_reverse(windows, window_size, H, W):
    B = int(windows.shape[0] / (H * W / window_size / window_size))
    x = windows.view(B, H // window_size, W // window_size, window_size, window_size, -1)
    x = x.permute(0,1,3,2,4,5).contiguous().view*(B, H, W, -1)
    return x

Window Attention

        本文关键，传统的Transformer都是基于全局来计算注意力的，因此计算量高。而Swin_Transformer则将注意力的计算限制在每个窗口内，从而减少了计算量。

        主要区别是在原始计算Attention的公式中的Q，K时加入了相对位置编码，后续试验证明相对位置编码的加入提升了模型性能。

class WindowAttention(nn.Module):
    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
        self.num_heads = num_heads
        head_dim = dim // self.num_heads
        self.scale = qk_scale or head_dim ** -0.5

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

        coords_h = torch.arrange(self.window_size[0])
        coords_w = torch.arrange(self.window_size[1])

        coords = torch.stack(torch.meshgrid([coords_h, coords_w]))
        coords_flatten = torch.flatten(coords,1)
        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]
        relative_coords = relative_coords.permute(1,2,0).contiguous()
        relative_coords[:, :, 0] += self.window_size[0] - 1
        relative_coords[:, :, 1] += self.window_size[1] - 1
        relative_coords[:, :, 0] *= 2 * self.window_size[1] -1
        self.relative_position_index = relative_coords.sum(-1)
        self.register_buffer("relative_position_index", self.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):
        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]

        q = q * self.scale
        attn = (q @ k.transpose(-2, -1))

        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)
        attn = attn + relative_position_bias.unsqueeze(0)

        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)

        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

相对位置编码

        首先QK计算出来的Attention张量形状为(numB，num_heads,window_sizewindow_size,window_size*window_size)。对于Attention张量来说，以不同的元素为源点，其他坐标也是不同的，以window_size=2为例，其相对位置编码如下图:

        首先，我们使用torch.arrange和torch.meshgrid函数生成对应的坐标：

torch.meshgrid（）的功能是生成网格，可以用于生成坐标。函数输入两个数据类型相同的一维张量，两个输出张量的行数为第一个输入张量的元素个数，列数为第二个输入张量的元素个数，当两个输入张量数据类型不同或维度不是一维时会报错。
其中第一个输出张量填充第一个输入张量中的元素，各行元素相同；第二个输出张量填充第二个输入张量中的元素各列元素相同

        以window_size=2为例：

coords_h = torch.arange(self.window_size[0])
coords_w = torch.arange(self.window_size[1])
coords = torch.meshgrid([coords_h, coords_w]) # -> 2*(wh, ww)
"""
  (tensor([[0, 0],
           [1, 1]]), 
   tensor([[0, 1],
           [0, 1]]))
"""

        然后堆叠起来，形成一个二维向量

coords = torch.stack(coords)  # 2, Wh, Ww
coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
"""
tensor([[0, 0, 1, 1],
        [0, 1, 0, 1]])
"""

        利用广播机制，分别在第一维，第二维插入一个维度，进行广播相减，得到2，whww，whww的张量

relative_coords_first = coords_flatten[:, :, None]  # 2, wh*ww, 1
relative_coords_second = coords_flatten[:, None, :] # 2, 1, wh*ww
relative_coords = relative_coords_first - relative_coords_second # 最终得到 2, wh*ww, wh*ww 形状的张量

        因为采取的是相减，所以得到的索引是从负数开始的，我们加上偏移量，让其从0 开始

relative_coords = relative_coords.permute(1, 2, 0).contiguous() # Wh*Ww, Wh*Ww, 2
relative_coords[:, :, 0] += self.window_size[0] - 1
relative_coords[:, :, 1] += self.window_size[1] - 1

        后面我们需要将其展开成一维偏移量。而对于(1,2)和(2,1)这两个坐标，在二维上是不同的，但是通过将x，y坐标相加转换为一维偏移的时候，他的偏移量是相等的。

        所以最后对其中做了个乘法操作，以进行区分

relative_coords[:, :, 0] *= 2 * self.window_size[1] - 1

        再看 forward 代码

    def forward(self, x, mask=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]

        q = q * self.scale
        attn = (q @ k.transpose(-2, -1))

        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)
        attn = attn + relative_position_bias.unsqueeze(0)

        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)

        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

• 首先输入张量形状为num_windowsB，window_sizewindow_size，C
• 然后经过self.qkv这个全连接层后进行reshape调增轴的顺序，得到形状为3，num_widowsB，num_heads，window_sizewindow_size，C//num_heads，并分别分配各Q,K,V。
• 根据公式，我们将Q乘以一个scale缩放系数，然后与K（为了满足矩阵相乘的要求，需要将最后两个维度进行交换）进行相乘。得到形状为(num_WindowB，num_heads，window_sizewindow_size,window_size*window_size)的attn张量
• 之前我们针对位置编码设置了个形状为((2window_size-1)(2window_size，num_heads))的可学习变量。我们用计算得到的相对位置编码位置索引 self.relative_position_index选取，得到形状为(window_sizewindow_size，window_size*window_size，num_heads)的编码，加到attn张量上
• 暂不考虑mask的情况，剩下的就是和transformer一样的softmax，dropout，与V矩阵进行相乘，在经过一层全连接层和dropout

Shifted Window Attention

        前面的Window Attention是在每个窗口下计算注意力的，为了更好的和其他window进行信息交互，Swin_Transformer还引入了shifted window操作。

        左边是没有重叠的Window Attention，而右边则是将窗口进行移位的Shifted Window Attentin。可以看到的思移位后的窗口包含了原本相邻窗口的元素。但这也引入了一个系问题，即windows的个数翻倍了，由原本的四个窗口变成了9个窗口。
在实际代码里，我们是通过对特征图移位，并给Attenton设置mask来间接实现的。能在保持原有的windows个数下，最后的计算结果等价

特征图移位操作

        代码里对特征图移位是通过torch.roll来实现的，下面是示意图;

Attention Mask

        此处通过设置合理的mask，让Shifted Windowu Attention在与Window Attention相同的窗口个数下，达到等价的计算结果。
首先，对Shift Window后的每个窗口都给上index，并且做一个roll操作(window_size=2，shift_size=1)

        我们希望在计算Attention时，让具有相同index 的 QK进行计算，而忽略不同index QK计算结果。

        而想在原始4个窗口下得到正确结果，我们必须给Attention的结果加入一个mask(如上图右边所示)

代码如下：

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:
                    img_mask[:, h, w, :] = cnt
                    cnt += 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)
            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))

        得到的mask

tensor([[[[[   0.,    0.,    0.,    0.],
           [   0.,    0.,    0.,    0.],
           [   0.,    0.,    0.,    0.],
           [   0.,    0.,    0.,    0.]]],


         [[[   0., -100.,    0., -100.],
           [-100.,    0., -100.,    0.],
           [   0., -100.,    0., -100.],
           [-100.,    0., -100.,    0.]]],


         [[[   0.,    0., -100., -100.],
           [   0.,    0., -100., -100.],
           [-100., -100.,    0.,    0.],
           [-100., -100.,    0.,    0.]]],


         [[[   0., -100., -100., -100.],
           [-100.,    0., -100., -100.],
           [-100., -100.,    0., -100.],
           [-100., -100., -100.,    0.]]]]])

        在之前的window attention模块的前向代码里，包含这么一段

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)

        将mask加到attention的计算结果，并进行softmax。mask的值设置为-100，softmax后就会忽略掉对应的值。

Transformer Block整体架构

        两个连续的Block架构如上图所示，需要注意的是一个Stage包含的Block个数必须是偶数，因为需要交替包含一个含有Window Attention的Layer和含有Shifted Window Attention的Layer。
看下Block的forward代码

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 = 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)
        shifted_x = window_reverse(attn_windows, self.window_size, H, W)  # 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 + self.drop_path(x)
        x = x + self.drop_path(self.mlp(self.norm2(x)))

        return x

        整体流程如下：

• 先对特征图进行LayerNorm
• 通过 self.shifted_size 决定是否需要对特征图进行shift
• 然后将特征图切成一个个窗口
• 计算Attention，通过self.attn_mask来区分Window Attention还是shift-Window Attention
• 将各个窗口合并回来
• 如果之间有shift操作，此时进行reverse shift，把之前的shift操作恢复
• 做dropout和残差连接
• 再通过一层LayerNorm+全连接层，以及dropout和残差连

参考

• 知乎:zzzk
• Swin Transformer: Hierarchical Vision Transformer using Shifted Windows
• 知乎专栏