Swin-Transformer（2021-08）

简介

一直到写下这篇笔记开始，基于Swin的模型仍然霸榜Object Detection等多个榜单。
很多博客都已经介绍的非常详细了，这里只记录一下自己学习过程中遇到的困惑。

Swin与ViT的对比，ViT将image划分为固定大小的patch，以patch为单位进行attention计算，计算过程中的feature map 分辨率是保持不变的，并且ViT为了保持与NLP的一致性，添加了一个额外的class token，最后用于分类。因此ViT不太适用于检测等下游任务，无法提取多尺度特征。

想要transformer用于检测等视觉任务，一种就是模仿CNN，将transformer改造为层级机构，一种就是使用纯transformer结构进行探索。

Swin显然属于前者，采用局部窗口自注意力和shift window的方式巧妙构造出了层级结构，可以作为视觉领域的通用backbone使用。

上图是Swin-T的结构图，输入图像首先通过Patch Partitiion 和Linear Embedding 转化为token形式的向量序列，然后输入Swin Transformer Block中，每个Block都是由一个window–Multi-Head self Attention 和Shift-Window Multi-head self-attention 组成，因此总是偶数。

window partition 窗口划分

将(B, H, W, C) 划分为(num_windows*B, window_size, window_size, C)的windows

def window_partition(x, window_size):
    """
    Args:
        x: (B, H, W, C)
        window_size (int): window size

    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)
    #view()  必须针对contiguous数据存放格式
    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
    return windows
#恢复原始输入 x
def window_reverse(windows, window_size, H, W):
    """
    Args:
        windows: (num_windows*B, window_size, window_size, C)
        window_size (int): Window size
        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))
    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

测试运行如下

x = torch.randn(1,224,224,3)
p =window_partition(x,7)
print(p.size())
o = window_reverse(p,7,224,224)
print(o.size())

***output

 torch.Size([1024, 7, 7, 3])
torch.Size([1, 224, 224, 3])***

partition就是将输入image转换为指定window size的patch向量，这里就是将（1 ，224，224，3）的batch转换为1024个窗口尺寸为(7,7)的patch。

reverse函数是partition的反函数

PatchEmbedding

[B C H W]->[B ,Ph*Pw,96]

import torch 
from torch import nn
from timm.models.layers import to_2tuple

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) #(224,224)
        patch_size = to_2tuple(patch_size)  #patch size (4,4)
        patches_resolution = [img_size[0] // patch_size[0], img_size[1] // patch_size[1]] #(56,56)
        self.img_size = img_size 
        self.patch_size = patch_size
        self.patches_resolution = patches_resolution #分辨率指patch数
        self.num_patches = patches_resolution[0] * patches_resolution[1] #56*56=3136

        self.in_chans = in_chans
        self.embed_dim = embed_dim
        #使用2d卷积进行patch 划分，输入channe默认为3
        self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=patch_size) #output:(batch,96,56,56)
        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
        #限制图像尺寸必须为224*224
        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]})." 
                   
        #B C h W  -> b embed h*w ->b h*W embed 得到 Patch Embeding 形状
        x = self.proj(x) #Patch 划分 [1,96,56,56]
        x=x.flatten(2).transpose(1, 2)  # B Ph*Pw C
        if self.norm is not None:
            x = self.norm(x)
        return x

    def flops(self):
        Ho, Wo = self.patches_resolution
        flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1])
        if self.norm is not None:
            flops += Ho * Wo * self.embed_dim
        return flops
    
x =  torch.randn(1,3,224,224)
PM  = PatchEmbed()
out = PM(x)
print(out.shape)

**output

torch.Size([1, 3136, 96])**

PatchMerging

[B,H*W,C]->[B,H/2,W/2,2C]
PatchMerging 相当于卷积中的下采样，降低feature map的分辨率，同时增加channel的维度，这里分辨率H，W都降为原来的一半(整体减少4倍)，channel数变为原来的两倍。

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" #限制输入x的第二个维度与H,W匹配
        assert H % 2 == 0 and W % 2 == 0, f"x size ({H}*{W}) are not even."

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

        #按照 H W 间隔为2 ，两两组合 共4组
        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
        
        print("\n",x0,"\n",x1,"\n",x2,"\n",x3)
        
        x = torch.cat([x0, x1, x2, x3], -1)  # B H/2 W/2 4*C  #特征图分辨率降低一半
        x = x.view(B, -1, 4 * C)  # B H/2*W/2 4*C  此时通道维度变为原来的4倍

        x = self.norm(x)
        x = self.reduction(x)  #4C -> 2C  通过linear 通道维度由4倍变为原来的2倍

        return x

    def extra_repr(self) -> str:
        return f"input_resolution={self.input_resolution}, dim={self.dim}"

    def flops(self):
        H, W = self.input_resolution
        flops = H * W * self.dim
        flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
        return flops
    
PM = PatchMerging(input_resolution=(4,4),dim=3)
x = torch.arange(48,dtype=torch.float).view(1,16,3)
o=PM(x)
print(o.shape)

output：

***tensor([[[[ 0.,  1.,  2.],
          [ 6.,  7.,  8.]],
         [[24., 25., 26.],
          [30., 31., 32.]]]])*** 
          
 ***tensor([[[[12., 13., 14.],
          [18., 19., 20.]],
         [[36., 37., 38.],
          [42., 43., 44.]]]])*** 
          
 ***tensor([[[[ 3.,  4.,  5.],
          [ 9., 10., 11.]],
         [[27., 28., 29.],
          [33., 34., 35.]]]])*** 
          
 ***tensor([[[[15., 16., 17.],
          [21., 22., 23.]],
         [[39., 40., 41.],
          [45., 46., 47.]]]])***
          
**torch.Size([1, 4, 6])**

Window Attention

窗口自注意力计算，限制在一个window中，公式相比与之前的attention添加了一个额外的B-相对位置偏移(relative postional bias)

从论文给出的结果来看，采用相对位置偏置后的结果要比其它方法更好。

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  #每个注意力头对应的维数
        self.scale = qk_scale or head_dim ** -0.5  #QK^T 放缩倍数 根号D

        # define a parameter table of relative position bias
        #构造可学习的相对位置偏置table，长度为 （2H-1)*(2W-1)*(num_head)  
        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
        #获取window中每一个token的相对位置索引对
        coords_h = torch.arange(self.window_size[0]) #[0,1,2...H]
        coords_w = torch.arange(self.window_size[1]) #[0,1,2...W]
        
        # mehsgrid相当于生成一个二维坐标，利用stack将两个w*h shape的张量按dim0拼接  
        coords = torch.stack(torch.meshgrid([coords_h, coords_w])) #（2, Wh, Ww） 
        
        coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww  拉直为2个一维序列，方便索引
        
        
        #计算相对位置索引坐标
         #[2,Wh*Ww,1]-[2,1,Wh*Ww]=[2, Wh*Ww, Wh*Ww]  广播机制
        relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :] 
        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
        relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
        
        #注册buffer保存到state_dict中作为永久变量relative_position_index保存，
        # 不需要参与梯度计算，只作为索引到bias table中取值
        self.register_buffer("relative_position_index", relative_position_index)

        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)#得到Q，K，V
        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) #初始化
        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
        '''
        x.shape = (num_windows*B, N, C) N=Wh*Ww 
        self.qkv(x).shape = (num_windows*B, N, 3C) 经过全连接层后，调整维度分配给Q，K，V
        self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).shape = (num_windows*B, N, 3, num_heads, C//num_heads)
        self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4).shape = (3, num_windows*B, num_heads, N, C//num_heads)
        '''
        qkv = self.qkv(x).reshape(B_, N, 3, self.num_heads, C // self.num_heads).permute(2, 0, 3, 1, 4) #Q,K,V
        q, k, v = qkv[0], qkv[1], qkv[2]  # make torchscript happy (cannot use tensor as tuple)

        '''
        q.shape = k.shape = v.shape = (num_windows*B, num_heads, N, C//num_heads)
        N = 代表patches的数量
        C//num_heads代表Q,K,V的维数
        '''
        q = q * self.scale #Q/sqrt(dim)
        attn = (q @ k.transpose(-2, -1))  #Q*K^T
        """
        attn.shape = (num_windows*B, num_heads, N, N)
        """

        #相对位置偏置B,将table中的值按照 (W*H,W*H,nH)的shape排列
        """ 
        self.relative_position_index矩阵中的所有值都是从self.relative_position_bias_table中取的
        self.relative_position_index是计算出来不可学习的量
        self.relative_position_index.view(-1).shape=Wh*Ww*Wh*Ww  把索引矩阵展开成一维，再到Bias Table中根据索引取值
        self.relative_position_bias_table.shape = (2*Wh-1 * 2*Ww-1, nH)
        
        xxxx.view(self.window_size[0] * self.window_size[1], self.window_size[0] * self.window_size[1], -1)
        得到的值为：每个attn头对应的Wh*Ww*Wh*Ww【索引self.relative_position_index.view(-1)】长度的值，每一列对应一个attn头
        """
        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
        
        #转换为与attention shape一致
        relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
        attn = attn + relative_position_bias.unsqueeze(0) #
        """ 
                                                                     第一个维度  广播机制 相加 
        attn.shape = [num_windows*B,nH, Wh*Ww, Wh*Ww]==(num_windows*B, num_heads, N, N)+[1,nH, Wh*Ww, Wh*Ww]
        所有attn头共用一个索引矩阵和bias table
        """
        #mask windows 部分
        if mask is not None:
            nW = mask.shape[0]
            # attn.view(B_ // nW, nW, self.num_heads, N, N).shape = (B, num_windows, num_heads, N, N) 第一个N代表有M2个token，第二个N代表每个token要计算M2次QKT的值
            # mask.unsqueeze(1).unsqueeze(0).shape =                (1, num_windows, 1, N, N) 第一个N代表有M2个token，第二个N代表每个token要计算M2次QKT的值
            # broadcast相加
            attn = attn.view(B_ // nW, nW, self.num_heads, N, N) + mask.unsqueeze(1).unsqueeze(0)
            """ 
            attn.view().shape = [num_windows*B//nW,nH, N, N]    N=Wh*Ww即window中的patch数
            attn.shape = (B, num_windows, num_heads, N, N)
            """
            attn = attn.view(-1, self.num_heads, N, N) 
            """ 
            attn.shape = (num_windows*B, 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) #全连接
        x = self.proj_drop(x) #dropout
        return x  # x.shape = (num_windows*B, N, C)  N:窗口中所有patches的数量

    def extra_repr(self) -> str:
        return f'dim={self.dim}, window_size={self.window_size}, num_heads={self.num_heads}'

    def flops(self, N):
        # calculate flops for 1 window with token length of N
        flops = 0
        # qkv = self.qkv(x)
        flops += N * self.dim * 3 * self.dim
        # attn = (q @ k.transpose(-2, -1))
        flops += self.num_heads * N * (self.dim // self.num_heads) * N
        #  x = (attn @ v)
        flops += self.num_heads * N * N * (self.dim // self.num_heads)
        # x = self.proj(x)
        flops += N * self.dim * self.dim
        return flops

mask shift-window 可视化

更直观的方式，有颜色的结果是需要自注意力计算保留的，其他白色部分是需要忽略(mask)的地方，可以参考从后边代码生成的结果与这个图是一致的。

 #attn mask 部分测试代码

#假设输入image 为4X4大小，划分为4个window，每个window为2X2大小
input_resolution =(4,4)
window_size =2
shift_size=1  #shift_size 也就是[M/2]下取整

def window_partition(x, window_size):#[B H W C]->[BHW/(window size)^2 , window size,winsow size,C ]
    """
    Args:
        x: (B, H, W, C)
        window_size (int): window size

    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)
    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
    return windows


if shift_size > 0:  
     
    # calculate attention mask for SW-MSA
    H, W = input_resolution
    img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
    
    # 提前规划好的分片位置索引-slice(start,stop,step)
    h_slices = (slice(0, -window_size),
                slice(-window_size, -shift_size),
                slice(-shift_size, None))
    w_slices = (slice(0, -window_size),
                slice(-window_size, -shift_size),
                slice(-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, window_size)  # nW, window_size, window_size, 1
    mask_windows = mask_windows.view(-1, 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


print(attn_mask)

output

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.]]])

#mask shift window 可视化代码

import torch

import matplotlib.pyplot as plt


def window_partition(x, window_size):
    """
    Args:
        x: (B, H, W, C)
        window_size (int): window size

    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)
    windows = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
    return windows


window_size = 7
shift_size = 3
H, W = 14, 14
img_mask = torch.zeros((1, H, W, 1))  # 1 H W 1
h_slices = (slice(0, -window_size),
            slice(-window_size, -shift_size),
            slice(-shift_size, None))
w_slices = (slice(0, -window_size),
            slice(-window_size, -shift_size),
            slice(-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, window_size)  # nW, window_size, window_size, 1
mask_windows = mask_windows.view(-1, 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))

plt.matshow(img_mask[0, :, :, 0].numpy())
plt.matshow(attn_mask[0].numpy())
plt.matshow(attn_mask[1].numpy())
plt.matshow(attn_mask[2].numpy())
plt.matshow(attn_mask[3].numpy())

plt.show()

Swin 模块实现

#SwinTransformer(BasicLayer(SwinBlock))
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
    """

    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):
        super().__init__()
        self.dim = dim  #输入通道
        self.input_resolution = input_resolution  #输入image 分辨率
        self.num_heads = num_heads #attn头
        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)#(num_windows*B, N, C)

        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
            
            # 提前规划好的分片索引-slice(start,stop,step)
            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))
        else:
            attn_mask = None

        self.register_buffer("attn_mask", attn_mask)

    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:
            #图像重新组合window计算attention
            shifted_x = torch.roll(x, shifts=(-self.shift_size, -self.shift_size), dims=(1, 2))
        else:
            shifted_x = x

        # partition windows
        #移动后的图像划分窗口
        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 计算完后还原回去
        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)
        x = shortcut + self.drop_path(x)

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

        return x

    def extra_repr(self) -> str:
        return f"dim={self.dim}, input_resolution={self.input_resolution}, num_heads={self.num_heads}, " \
               f"window_size={self.window_size}, shift_size={self.shift_size}, mlp_ratio={self.mlp_ratio}"

    def flops(self):
        flops = 0
        H, W = self.input_resolution
        # norm1
        flops += self.dim * H * W
        # W-MSA/SW-MSA
        nW = H * W / self.window_size / self.window_size
        flops += nW * self.attn.flops(self.window_size * self.window_size)
        # mlp
        flops += 2 * H * W * self.dim * self.dim * self.mlp_ratio
        # norm2
        flops += self.dim * H * W
        return flops


class PatchMerging(nn.Module):  #[B H W C]->[B H/4 W/4 2C]
    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)
        #H，W 以2为间隔取数
        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

    def extra_repr(self) -> str:
        return f"input_resolution={self.input_resolution}, dim={self.dim}"

    def flops(self):
        H, W = self.input_resolution
        flops = H * W * self.dim
        flops += (H // 2) * (W // 2) * 4 * self.dim * 2 * self.dim
        return flops

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

    Args:
        dim (int): Number of input channels.
        input_resolution (tuple[int]): Input resolution.
        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
        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 | 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, input_resolution, depth, num_heads, window_size,
                 mlp_ratio=4., qkv_bias=True, qk_scale=None, drop=0., attn_drop=0.,
                 drop_path=0., norm_layer=nn.LayerNorm, downsample=None, use_checkpoint=False):

        super().__init__()
        self.dim = dim
        self.input_resolution = input_resolution
        self.depth = depth
        self.use_checkpoint = use_checkpoint

        # 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, # 整除2为0表示W-MSA，否则shift-size决定是否SW-MSA 
                                 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)
            for i in range(depth)]) #每个transformer block是偶数，depth对应stage遍历W-MSA/SW-MSA,如[2,2,6,2]

        # patch merging layer
        if downsample is not None:
            self.downsample = downsample(input_resolution, dim=dim, norm_layer=norm_layer)
        else:
            self.downsample = None

    def forward(self, x):
        for blk in self.blocks:
            if self.use_checkpoint:
                x = checkpoint.checkpoint(blk, x)
            else:
                x = blk(x)
        if self.downsample is not None:
            x = self.downsample(x)
        return x

    def extra_repr(self) -> str:
        return f"dim={self.dim}, input_resolution={self.input_resolution}, depth={self.depth}"

    def flops(self):
        flops = 0
        for blk in self.blocks:
            flops += blk.flops()
        if self.downsample is not None:
            flops += self.downsample.flops()
        return flops


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
        #使用卷积进行patch 划分
        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]})." 
                   
        #B C (h W ) -> b c h*w ->b h*W c 得到 Patch Embeding 形状
        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

    def flops(self):
        Ho, Wo = self.patches_resolution
        flops = Ho * Wo * self.embed_dim * self.in_chans * (self.patch_size[0] * self.patch_size[1])
        if self.norm is not None:
            flops += Ho * Wo * self.embed_dim
        return flops

#SwinTransformer(BasicLayer(SwinBlock))
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:
        img_size (int | tuple(int)): Input image size. Default 224
        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
        qk_scale (float): Override default qk scale of head_dim ** -0.5 if set. Default: None
        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.
        ape (bool): If True, add absolute position embedding to the patch embedding. Default: False
        patch_norm (bool): If True, add normalization after patch embedding. Default: True
        use_checkpoint (bool): Whether to use checkpointing to save memory. Default: False
    """

    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, 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().__init__()

        self.num_classes = num_classes
        self.num_layers = len(depths)
        self.embed_dim = embed_dim
        self.ape = ape
        self.patch_norm = patch_norm
        #satge4 输出的channel 8C
        self.num_features = int(embed_dim * 2 ** (self.num_layers - 1)) # embed_dim*8
        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.patches_resolution
        self.patches_resolution = patches_resolution

        # absolute position embedding--ape
        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 droprate：0->drop_path_rate
        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): #4
            #layer 整个网络处理过程
            layer = BasicLayer(dim=int(embed_dim * 2 ** i_layer), #stage输出维度：c 2c 4c 8c
                               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=self.mlp_ratio,
                               qkv_bias=qkv_bias, qk_scale=qk_scale,
                               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(layer)

        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)
    #参数初始化函数 
    def _init_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)

    @torch.jit.ignore
    def no_weight_decay(self):
        return {'absolute_pos_embed'}

    @torch.jit.ignore
    def no_weight_decay_keywords(self):
        return {'relative_position_bias_table'}

    def forward_features(self, x):
        x = self.patch_embed(x)
        if self.ape:
            x = x + self.absolute_pos_embed
        x = self.pos_drop(x)

        for layer in self.layers:
            x = layer(x)

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

    def forward(self, x):
        x = self.forward_features(x)
        x = self.head(x)  #应该是用来分类
        return x

    def flops(self):
        flops = 0
        flops += self.patch_embed.flops()
        for i, layer in enumerate(self.layers):
            flops += layer.flops()
        flops += self.num_features * self.patches_resolution[0] * self.patches_resolution[1] // (2 ** self.num_layers)
        flops += self.num_features * self.num_classes
        return flops

Swin-X 参数配置列表

包含关系：Swin_transformer（Basic_layer(Swin_Block)）

复杂度分析

reference(推荐阅读)：

1.https://zhuanlan.zhihu.com/p/367111046
2. https://hub.fastgit.xyz/microsoft/Swin-Transformer/issues/38
3. https://zhuanlan.zhihu.com/p/430047908 -很清楚
4. https://blog.csdn.net/qq_37541097/article/details/121119988
5. https://blog.csdn.net/qq_39478403/article/details/120042232